Add files using upload-large-folder tool

jlens_materials/prompts/README.md

@@ -0,0 +1,53 @@
+# Prompt and experiment data
+
+`run_lens.py --prompts PATH` accepts either this toolkit's `{"prompts": [...]}`
+schema or the released Jacobian Lens `{"items": [...]}` evaluation/probe-swap
+schema. A directory loads every JSON file in lexical order.
+
+For a paper-protocol lens-quality item, provide:
+
+```json
+{
+  "slug": "stable-id",
+  "shape": "MULTIHOP",
+  "protocol": "lens_eval",
+  "domain": "mechanics",
+  "text": "The question ending immediately before its answer is",
+  "readout_selector": "before_answer",
+  "tracked": ["latent-intermediate"],
+  "synonyms": {"latent-intermediate": ["alternative"]},
+  "answer": "answer",
+  "must_be_absent_from_input": true,
+  "must_be_absent_from_output": true
+}
+```
+
+The available fixed selectors are `before_answer`, `final_prompt_token`,
+`last_newline`, `assistant_response`, and `explicit`. `all_prompt` exists only
+for exploratory visualization and is rejected by strict protocol mode.
+
+For `probe_swap`, also provide `swap_from`, `swap_to`, and `swap_answer`.
+For `verbal_report`, omit `swap_from`: the clean greedy answer is selected
+dynamically. Directed-modulation trials use the identical teacher-forced
+carrier/targets across `focus`, `suppress`, and `control`, connected by a shared
+`control_group`.
+For paper-scale modulation, also provide stable `phrasing_id`, `carrier_id`,
+and `target_family` values; strict paper runs require 24 distinct phrasings.
+
+[`mechanics-paper-example.json`](mechanics-paper-example.json) demonstrates all
+three families. It is deliberately too small for quantitative reporting;
+create at least 50 independent items per evaluation distribution and multiple
+instruction phrasings/trials for modulation.
+
+[`materials-qualitative-pack.json`](materials-qualitative-pack.json) provides
+14 materials-science case studies covering Hall--Petch strengthening, Orowan
+bypass, martensitic transformation, Schmid slip, diffusional creep, dislocation
+character, diffraction recognition, unnamed associations, matched modulation,
+and causal swaps. Run it with `--recipe demo --min-items-per-shape 50` for close
+inspection without accidentally labeling the small sample quantitative. Use
+`--generation-max-new-tokens 1` when auditing the input/output-absence controls.
+
+`fracture.json` and `protein.json` are retained as legacy visualization sets.
+The authoritative built-ins in `domain_prompts.py` include stricter scoring and
+matched modulation controls; use the mechanics example as the starting point
+for new paper-protocol datasets.

jlens_materials/prompts/fracture.json

@@ -0,0 +1,146 @@
+{
+  "prompts": [
+    {
+      "slug": "fracture-griffith",
+      "shape": "MULTIHOP",
+      "domain": "fracture",
+      "title": "Griffith criterion (multi-hop)",
+      "description": "Two-hop: the failure law for a brittle solid with a sharp flaw. Watch 'energy' / 'surface' / 'Griffith' surface before 'crack'.",
+      "text": "In linear elastic fracture mechanics, the critical stress for a brittle solid to fail is set by balancing the released strain energy against the energy needed to create new surface. The critical stress therefore scales inversely with the square root of the",
+      "readout_at": -1,
+      "tracked": [
+        "energy",
+        "surface",
+        "Griffith",
+        "crack",
+        "length"
+      ],
+      "note": "Griffith's result sigma_c ~ sqrt(2 E gamma / (pi a)). The answer token is 'crack (length)'; 'energy', 'surface' and the eponym 'Griffith' are unspoken intermediates the derivation passes through -- the fracture analogue of Mars->color->red."
+    },
+    {
+      "slug": "fracture-fatigue",
+      "shape": "MULTIHOP",
+      "domain": "fracture",
+      "title": "Cyclic loading -> failure mode (multi-hop)",
+      "description": "Cyclic sub-yield loading. 'fatigue' should lead 'crack', with 'Paris'/'striations' as deeper-layer associates.",
+      "text": "A steel component is loaded repeatedly far below its yield stress for millions of cycles. Tiny cracks nucleate at the surface and grow a little on every cycle until the part suddenly breaks. This progressive mode of failure is called",
+      "readout_at": -1,
+      "tracked": [
+        "fatigue",
+        "crack",
+        "cyclic",
+        "Paris",
+        "striations"
+      ],
+      "note": "The answer is 'fatigue'. Whether 'Paris' (the Paris crack-growth law) or 'striations' (the fractographic fingerprint) appear in deeper layers probes how much associated fracture knowledge the workspace pulls in beyond the one-word answer."
+    },
+    {
+      "slug": "fracture-toughness-chain",
+      "shape": "MULTIHOP",
+      "domain": "fracture",
+      "title": "Stress-intensity factor chain (multi-hop)",
+      "description": "Chain: geometry+load -> stress intensity K -> compare to K_Ic -> propagate. Track K / toughness / propagate in order.",
+      "text": "An engineer has a plate with an edge crack under tension. She first computes the stress intensity factor at the crack tip, then compares it against the material's fracture toughness. If the first exceeds the second, the crack will",
+      "readout_at": -1,
+      "tracked": [
+        "stress",
+        "intensity",
+        "toughness",
+        "propagate",
+        "grow",
+        "unstable"
+      ],
+      "note": "Multi-step like (4+17)*2+7: compute K, then compare to K_Ic, then conclude 'propagate'. The paper's claim is that these steps surface in the lens in computed order and at successively later layers."
+    },
+    {
+      "slug": "fracture-brittle-assoc",
+      "shape": "ASSOCIATION",
+      "domain": "fracture",
+      "title": "Brittle fracture (association, unnamed)",
+      "description": "A vignette of glass shattering. Does 'brittle' / 'fracture' rise in the lens at the final token though never named?",
+      "text": "The technician tapped the cold pane once with the hammer. Without bending or warning it flew apart into a hundred glittering shards that skated across the floor, each edge sharp as the moment before.",
+      "readout_at": -1,
+      "tracked": [
+        "brittle",
+        "fracture",
+        "glass",
+        "shatter",
+        "crack"
+      ],
+      "note": "Association eval (paper: grief/Einstein/noir). The passage evokes brittle fracture without the words 'brittle' or 'fracture'; a workspace representation of the concept should be readable at the closing token."
+    },
+    {
+      "slug": "fracture-ductile-assoc",
+      "shape": "ASSOCIATION",
+      "domain": "fracture",
+      "title": "Ductile failure (association, unnamed)",
+      "description": "A copper rod necking and drawing out. Contrast partner to the brittle vignette: expect 'ductile'/'plastic'/'neck'.",
+      "text": "As the load climbed the copper rod stopped springing back. It grew visibly thinner at one waist, stretching and drawing out like warm taffy, holding on far past the point anyone expected before it finally parted.",
+      "readout_at": -1,
+      "tracked": [
+        "ductile",
+        "plastic",
+        "necking",
+        "yield",
+        "deformation"
+      ],
+      "note": "The matched control for fracture-brittle-assoc. Comparing the two tests whether the lens cleanly separates the two canonical failure regimes from prose alone."
+    },
+    {
+      "slug": "fracture-stress-recognition",
+      "shape": "RECOGNITION",
+      "domain": "fracture",
+      "title": "Stress state from a tensor (recognition)",
+      "description": "A raw Cauchy stress tensor with one large tensile eigenvalue. Does the lens read out 'tension'/'crack'/'mode'?",
+      "text": "Consider the Cauchy stress tensor (MPa):\n[[420,   0,   0],\n [  0,  15,   0],\n [  0,   0,  10]]\nThe dominant loading state this represents is",
+      "readout_at": -1,
+      "tracked": [
+        "tension",
+        "tensile",
+        "uniaxial",
+        "crack",
+        "opening",
+        "mode"
+      ],
+      "note": "Analogue of the paper reading 'protein' out of a raw sequence: here the model must parse a numeric tensor, recognise near-uniaxial tension, and (in fracture terms) associate it with Mode I opening."
+    },
+    {
+      "slug": "fracture-modulation",
+      "shape": "MODULATION",
+      "domain": "fracture",
+      "title": "Hold 'crack-tip plasticity' while copying (modulation)",
+      "description": "Copy an unrelated sentence while concentrating on crack-tip plasticity. Does the concept surface over the carrier response?",
+      "user": "Write \"The lab notebook lay open on the bench.\" Concentrate on crack-tip plasticity while you write the sentence. Don't write anything else.",
+      "assistant_prefill": "The lab notebook lay open on the bench.",
+      "readout_at": -1,
+      "tracked": [
+        "crack",
+        "plastic",
+        "plasticity",
+        "tip",
+        "yield",
+        "zone"
+      ],
+      "note": "Directed-modulation protocol. Success = 'crack'/'plastic'/'zone' appear in the lens across the carrier sentence, whose surface text is about a notebook -- evidence the workspace holds an instructed concept independent of output."
+    },
+    {
+      "slug": "fracture-report-swap",
+      "shape": "REPORT_SWAP",
+      "domain": "fracture",
+      "title": "Think of a failure mode -> swap (verbal report)",
+      "description": "'Think of a failure mode. Answer in one word.' Swap the chosen mode's lens vector for a sibling and check the report changes.",
+      "user": "Think of a metal failure mode. Answer in one word.",
+      "readout_at": -1,
+      "tracked": [
+        "fatigue",
+        "creep",
+        "corrosion",
+        "fracture",
+        "buckling"
+      ],
+      "swap_from": "fatigue",
+      "swap_to": "creep",
+      "note": "Verbal-report swap (paper: Soccer->Rugby). The siblings are all single-word failure modes; swapping the chosen word's J-lens coordinate for a sibling's should flip the one-word answer."
+    }
+  ]
+}

jlens_materials/prompts/materials-heldout-v1-preregistered.json

@@ -0,0 +1,1068 @@
+{
+  "format_version": 1,
+  "study": "materials-heldout-v1",
+  "frozen_before_execution": "2026-07-14",
+  "description": "Fifty held-out association descriptions: ten materials mechanism families with five new phrasings each. Generated and frozen before any held-out model or lens output was inspected.",
+  "prompts": [
+    {
+      "slug": "heldout-v1-assoc-ductile-01",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "ductile",
+      "target_family": "ductile",
+      "phrasing_id": "heldout-ductile-01",
+      "title": "Energy-absorbing dimpled separation 1",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A round aluminum bar stretched into a pronounced waist before separating. The mating faces contained hemispherical pits centered on inclusions and a slanted tearing rim.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "ductile",
+        "nucleation",
+        "coalescence",
+        "void"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-ductile-02",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "ductile",
+      "target_family": "ductile",
+      "phrasing_id": "heldout-ductile-02",
+      "title": "Energy-absorbing dimpled separation 2",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A nickel coupon accumulated large permanent elongation before rupture. Microscopy showed a honeycomb of deep depressions surrounding second-phase particles.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "ductile",
+        "nucleation",
+        "coalescence",
+        "void"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-ductile-03",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "ductile",
+      "target_family": "ductile",
+      "phrasing_id": "heldout-ductile-03",
+      "title": "Energy-absorbing dimpled separation 3",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A copper tensile specimen lost much of its cross-sectional area before failure. Small particle-centered holes had enlarged and joined across the final ligament.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "ductile",
+        "nucleation",
+        "coalescence",
+        "void"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-ductile-04",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "ductile",
+      "target_family": "ductile",
+      "phrasing_id": "heldout-ductile-04",
+      "title": "Energy-absorbing dimpled separation 4",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "An impact-tested structural metal absorbed substantial energy and developed broad lateral lips. Its failure surface was covered by rounded microscopic depressions.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "ductile",
+        "nucleation",
+        "coalescence",
+        "void"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-ductile-05",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "ductile",
+      "target_family": "ductile",
+      "phrasing_id": "heldout-ductile-05",
+      "title": "Energy-absorbing dimpled separation 5",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A pressurized metal tube bulged strongly before opening. The torn wall displayed elongated pits near the surface and deeper equiaxed pits toward the center.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "ductile",
+        "nucleation",
+        "coalescence",
+        "void"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-boundary-attack-01",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "boundary-attack",
+      "target_family": "boundary-attack",
+      "phrasing_id": "heldout-boundary-attack-01",
+      "title": "Chromium-depleted interfacial attack 1",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A stainless weld spent several hours near 675 degrees Celsius. After nitric-acid exposure, narrow trenches followed the grain edges beside chromium-rich carbide rows.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "boundary",
+        "corrosion",
+        "sensitization"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-boundary-attack-02",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "boundary-attack",
+      "target_family": "boundary-attack",
+      "phrasing_id": "heldout-boundary-attack-02",
+      "title": "Chromium-depleted interfacial attack 2",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "An austenitic pipe was slowly cooled through an intermediate temperature range. A later chemical test exposed a connected network of material loss outlining individual grains.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "boundary",
+        "corrosion",
+        "sensitization"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-boundary-attack-03",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "boundary-attack",
+      "target_family": "boundary-attack",
+      "phrasing_id": "heldout-boundary-attack-03",
+      "title": "Chromium-depleted interfacial attack 3",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "Carbide particles decorated the interfaces in heat-affected stainless steel. Immersion testing removed chromium-depleted regions next to those particles while grain interiors remained sound.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "boundary",
+        "corrosion",
+        "sensitization"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-boundary-attack-04",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "boundary-attack",
+      "target_family": "boundary-attack",
+      "phrasing_id": "heldout-boundary-attack-04",
+      "title": "Chromium-depleted interfacial attack 4",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A stainless vessel received an improper post-weld thermal cycle. In service, fissures traced the polygonal grain-edge network instead of cutting through the crystallites.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "boundary",
+        "corrosion",
+        "sensitization"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-boundary-attack-05",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "boundary-attack",
+      "target_family": "boundary-attack",
+      "phrasing_id": "heldout-boundary-attack-05",
+      "title": "Chromium-depleted interfacial attack 5",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "Austenitic sheet aged near 650 degrees Celsius showed continuous carbide films around grains. An aggressive liquid subsequently produced narrow grooves along the same paths.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "boundary",
+        "corrosion",
+        "sensitization"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-cyclic-01",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "cyclic",
+      "target_family": "cyclic",
+      "phrasing_id": "heldout-cyclic-01",
+      "title": "Progressive damage under repeated loading 1",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A compressor blade endured millions of small vibration cycles before failing. Curved arrest lines radiated from a surface defect toward a small rough overload region.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "fatigue",
+        "crack",
+        "propagation"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-cyclic-02",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "cyclic",
+      "target_family": "cyclic",
+      "phrasing_id": "heldout-cyclic-02",
+      "title": "Progressive damage under repeated loading 2",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A bridge hanger separated although every measured load was below its monotonic capacity. Fine parallel striations recorded incremental front advance from a machining groove.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "fatigue",
+        "crack",
+        "propagation"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-cyclic-03",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "cyclic",
+      "target_family": "cyclic",
+      "phrasing_id": "heldout-cyclic-03",
+      "title": "Progressive damage under repeated loading 3",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A wheel axle survived years of repeated bending. The failure face contained thumbnail-shaped beach marks centered on several surface initiation sites.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "fatigue",
+        "crack",
+        "propagation"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-cyclic-04",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "cyclic",
+      "target_family": "cyclic",
+      "phrasing_id": "heldout-cyclic-04",
+      "title": "Progressive damage under repeated loading 4",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A bolted aluminum joint opened after many pressurization cycles. A smooth region with regularly spaced microscopic ridges occupied most of the separated surface.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "fatigue",
+        "crack",
+        "propagation"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-cyclic-05",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "cyclic",
+      "target_family": "cyclic",
+      "phrasing_id": "heldout-cyclic-05",
+      "title": "Progressive damage under repeated loading 5",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A rotating laboratory specimen failed after a long sequence of alternating stresses. Ratchet marks joined multiple surface origins before the remaining ligament broke suddenly.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "fatigue",
+        "crack",
+        "propagation"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-cleavage-01",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "cleavage",
+      "target_family": "cleavage",
+      "phrasing_id": "heldout-cleavage-01",
+      "title": "Faceted low-plasticity separation 1",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A ferritic plate separated abruptly during cold weather with almost no macroscopic distortion. Flat facets and converging river markings crossed the grain interiors.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "brittle",
+        "cleavage",
+        "transgranular"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-cleavage-02",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "cleavage",
+      "target_family": "cleavage",
+      "phrasing_id": "heldout-cleavage-02",
+      "title": "Faceted low-plasticity separation 2",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A low-temperature impact specimen absorbed little energy and retained its original cross section. The shiny failure face contained crystallographic steps and feather patterns.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "brittle",
+        "cleavage",
+        "transgranular"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-cleavage-03",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "cleavage",
+      "target_family": "cleavage",
+      "phrasing_id": "heldout-cleavage-03",
+      "title": "Faceted low-plasticity separation 3",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A body-centered metal failed suddenly below its transition temperature. Microscopy revealed broad planar facets connected by sharp ridges through individual crystallites.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "brittle",
+        "cleavage",
+        "transgranular"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-cleavage-04",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "cleavage",
+      "target_family": "cleavage",
+      "phrasing_id": "heldout-cleavage-04",
+      "title": "Faceted low-plasticity separation 4",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A thick steel section opened under modest loading after rapid cooling. The separated faces showed flat reflective planes and negligible evidence of local necking.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "brittle",
+        "cleavage",
+        "transgranular"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-cleavage-05",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "cleavage",
+      "target_family": "cleavage",
+      "phrasing_id": "heldout-cleavage-05",
+      "title": "Faceted low-plasticity separation 5",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A notched iron alloy snapped during a cold bend test before appreciable permanent curvature developed. River-like markings on the facets converged toward the initiation site.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "brittle",
+        "cleavage",
+        "transgranular"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-high-temperature-deformation-01",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "high-temperature-deformation",
+      "target_family": "high-temperature-deformation",
+      "phrasing_id": "heldout-high-temperature-deformation-01",
+      "title": "Time-dependent deformation under heat 1",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A steam-pipe alloy elongated gradually during ten years of steady load at 600 degrees Celsius. Rounded pores appeared at triple junctions late in life.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "creep",
+        "diffusion",
+        "cavity"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-high-temperature-deformation-02",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "high-temperature-deformation",
+      "target_family": "high-temperature-deformation",
+      "phrasing_id": "heldout-high-temperature-deformation-02",
+      "title": "Time-dependent deformation under heat 2",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A constant-load high-temperature test showed rapid initial strain, a long nearly linear segment, and accelerating extension before rupture.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "creep",
+        "diffusion",
+        "cavity"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-high-temperature-deformation-03",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "high-temperature-deformation",
+      "target_family": "high-temperature-deformation",
+      "phrasing_id": "heldout-high-temperature-deformation-03",
+      "title": "Time-dependent deformation under heat 3",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A turbine component operated under sustained stress near half its melting temperature. Grain-edge pores enlarged and linked even though the applied stress was below the room-temperature yield value.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "creep",
+        "diffusion",
+        "cavity"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-high-temperature-deformation-04",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "high-temperature-deformation",
+      "target_family": "high-temperature-deformation",
+      "phrasing_id": "heldout-high-temperature-deformation-04",
+      "title": "Time-dependent deformation under heat 4",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A fine-grained metal slowly changed shape during a prolonged hot hold. Matter moved away from compressed grain faces and accumulated on faces under tension.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "creep",
+        "diffusion",
+        "cavity"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-high-temperature-deformation-05",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "high-temperature-deformation",
+      "target_family": "high-temperature-deformation",
+      "phrasing_id": "heldout-high-temperature-deformation-05",
+      "title": "Time-dependent deformation under heat 5",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A heat-resistant alloy accumulated measurable permanent strain with increasing service time. Interfacial pores and wedge-shaped separations were concentrated near the final rupture zone.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "creep",
+        "diffusion",
+        "cavity"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-particle-strengthening-01",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "particle-strengthening",
+      "target_family": "particle-strengthening",
+      "phrasing_id": "heldout-particle-strengthening-01",
+      "title": "Nonshearable particle obstacles 1",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A metal matrix contains closely spaced rigid dispersoids that cannot be cut. Moving lattice lines arc between them and leave closed loops around each obstacle.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "precipitation",
+        "bowing",
+        "strengthening"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-particle-strengthening-02",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "particle-strengthening",
+      "target_family": "particle-strengthening",
+      "phrasing_id": "heldout-particle-strengthening-02",
+      "title": "Nonshearable particle obstacles 2",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "After aging, a dense population of hard nanoscale particles raises the stress required for plastic flow. Electron microscopy shows rings encircling the particles after deformation.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "precipitation",
+        "bowing",
+        "strengthening"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-particle-strengthening-03",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "particle-strengthening",
+      "target_family": "particle-strengthening",
+      "phrasing_id": "heldout-particle-strengthening-03",
+      "title": "Nonshearable particle obstacles 3",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "Reducing the spacing between stable nonshearable particles increases the yield stress. Flexible line defects must curve strongly to pass through the remaining channels.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "precipitation",
+        "bowing",
+        "strengthening"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-particle-strengthening-04",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "particle-strengthening",
+      "target_family": "particle-strengthening",
+      "phrasing_id": "heldout-particle-strengthening-04",
+      "title": "Nonshearable particle obstacles 4",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A dispersion-treated alloy contains hard particles embedded in a softer matrix. Mobile lattice imperfections wrap around the particles rather than passing through them.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "precipitation",
+        "bowing",
+        "strengthening"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-particle-strengthening-05",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "particle-strengthening",
+      "target_family": "particle-strengthening",
+      "phrasing_id": "heldout-particle-strengthening-05",
+      "title": "Nonshearable particle obstacles 5",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "Two alloys contain the same particle fraction, but the alloy with finer obstacle spacing requires a larger applied stress to move its line defects.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "precipitation",
+        "bowing",
+        "strengthening"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-rapid-transformation-01",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "rapid-transformation",
+      "target_family": "rapid-transformation",
+      "phrasing_id": "heldout-rapid-transformation-01",
+      "title": "Quench-induced coordinated transformation 1",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A carbon steel quenched too rapidly for long-range solute redistribution develops a hard acicular product with a body-centered cell elongated along one axis.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "martensite",
+        "shear",
+        "Bain",
+        "tetragonal"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-rapid-transformation-02",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "rapid-transformation",
+      "target_family": "rapid-transformation",
+      "phrasing_id": "heldout-rapid-transformation-02",
+      "title": "Quench-induced coordinated transformation 2",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "Severe cooling converts the parent face-centered phase into fine laths almost instantaneously. Carbon remains trapped and the product has unequal axial lattice parameters.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "martensite",
+        "shear",
+        "Bain",
+        "tetragonal"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-rapid-transformation-03",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "rapid-transformation",
+      "target_family": "rapid-transformation",
+      "phrasing_id": "heldout-rapid-transformation-03",
+      "title": "Quench-induced coordinated transformation 3",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A rapidly cooled steel forms hard plates by coordinated short-distance atomic motion. No composition partitioning is detected across the moving interface.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "martensite",
+        "shear",
+        "Bain",
+        "tetragonal"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-rapid-transformation-04",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "rapid-transformation",
+      "target_family": "rapid-transformation",
+      "phrasing_id": "heldout-rapid-transformation-04",
+      "title": "Quench-induced coordinated transformation 4",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "Quenching suppresses long-range atomic transport and produces packets of highly strained needles. The resulting body-centered unit cell has a c-to-a ratio above one.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "martensite",
+        "shear",
+        "Bain",
+        "tetragonal"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-rapid-transformation-05",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "rapid-transformation",
+      "target_family": "rapid-transformation",
+      "phrasing_id": "heldout-rapid-transformation-05",
+      "title": "Quench-induced coordinated transformation 5",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "An austenitic sample undergoes an abrupt, composition-preserving shape change during cooling. The product is hard, plate-like, and internally twinned.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "martensite",
+        "shear",
+        "Bain",
+        "tetragonal"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-line-defect-motion-01",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "line-defect-motion",
+      "target_family": "line-defect-motion",
+      "phrasing_id": "heldout-line-defect-motion-01",
+      "title": "Motion of linear lattice imperfections 1",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "Plastic flow begins when a linear lattice imperfection travels across a close-packed plane and leaves a permanent surface step.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "dislocation",
+        "glide",
+        "slip"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-line-defect-motion-02",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "line-defect-motion",
+      "target_family": "line-defect-motion",
+      "phrasing_id": "heldout-line-defect-motion-02",
+      "title": "Motion of linear lattice imperfections 2",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "During loading of a single crystal, etch-pit positions move along selected crystallographic planes and an irreversible offset accumulates.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "dislocation",
+        "glide",
+        "slip"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-line-defect-motion-03",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "line-defect-motion",
+      "target_family": "line-defect-motion",
+      "phrasing_id": "heldout-line-defect-motion-03",
+      "title": "Motion of linear lattice imperfections 3",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A one-dimensional defect sweeps through the crystal in the direction of its mismatch vector, displacing the material on opposite sides of its plane.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "dislocation",
+        "glide",
+        "slip"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-line-defect-motion-04",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "line-defect-motion",
+      "target_family": "line-defect-motion",
+      "phrasing_id": "heldout-line-defect-motion-04",
+      "title": "Motion of linear lattice imperfections 4",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "Once the resolved driving stress exceeds a threshold, mobile lattice lines traverse their favored planes and produce permanent strain.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "dislocation",
+        "glide",
+        "slip"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-line-defect-motion-05",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "line-defect-motion",
+      "target_family": "line-defect-motion",
+      "phrasing_id": "heldout-line-defect-motion-05",
+      "title": "Motion of linear lattice imperfections 5",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "High-speed microscopy shows linear imperfections moving across densely packed planes. Each passage creates a small step at the free surface.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "dislocation",
+        "glide",
+        "slip"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-notch-resistance-01",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "notch-resistance",
+      "target_family": "notch-resistance",
+      "phrasing_id": "heldout-notch-resistance-01",
+      "title": "Resistance to unstable flaw extension 1",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "Two steels have comparable yield stress, yet one pre-notched plate absorbs several times more energy before sudden separation and survives a much larger flaw.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "toughness",
+        "crack",
+        "fracture"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-notch-resistance-02",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "notch-resistance",
+      "target_family": "notch-resistance",
+      "phrasing_id": "heldout-notch-resistance-02",
+      "title": "Resistance to unstable flaw extension 2",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A sharp starter notch blunts as a plastic zone develops, allowing the specimen to carry additional load before rapid extension occurs.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "toughness",
+        "crack",
+        "fracture"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-notch-resistance-03",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "notch-resistance",
+      "target_family": "notch-resistance",
+      "phrasing_id": "heldout-notch-resistance-03",
+      "title": "Resistance to unstable flaw extension 3",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A compact-tension sample requires substantial work to advance its sharp flaw. Instability begins only at an unusually large stress-intensity value.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "toughness",
+        "crack",
+        "fracture"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-notch-resistance-04",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "notch-resistance",
+      "target_family": "notch-resistance",
+      "phrasing_id": "heldout-notch-resistance-04",
+      "title": "Resistance to unstable flaw extension 4",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A structural alloy containing a long defect remains stable at loads that immediately separate a comparison alloy of similar strength.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "toughness",
+        "crack",
+        "fracture"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-notch-resistance-05",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "notch-resistance",
+      "target_family": "notch-resistance",
+      "phrasing_id": "heldout-notch-resistance-05",
+      "title": "Resistance to unstable flaw extension 5",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "An instrumented impact test on a notched bar records high absorbed energy and extensive local deformation before the remaining ligament opens.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "toughness",
+        "crack",
+        "fracture"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-hot-air-surface-layer-01",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "hot-air-surface-layer",
+      "target_family": "hot-air-surface-layer",
+      "phrasing_id": "heldout-hot-air-surface-layer-01",
+      "title": "Oxygen-rich high-temperature surface film 1",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A superalloy held in oxygen-bearing gas gains mass while developing a dense ceramic surface film whose thickness follows an approximately parabolic time dependence.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "oxidation",
+        "oxide",
+        "scale"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-hot-air-surface-layer-02",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "hot-air-surface-layer",
+      "target_family": "hot-air-surface-layer",
+      "phrasing_id": "heldout-hot-air-surface-layer-02",
+      "title": "Oxygen-rich high-temperature surface film 2",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "Long furnace exposure covers a steel coupon with several brittle oxygen-rich layers. Continued growth requires ions to cross the existing reaction product.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "oxidation",
+        "oxide",
+        "scale"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-hot-air-surface-layer-03",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "hot-air-surface-layer",
+      "target_family": "hot-air-surface-layer",
+      "phrasing_id": "heldout-hot-air-surface-layer-03",
+      "title": "Oxygen-rich high-temperature surface film 3",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A chromium-containing alloy forms a compact adherent surface film in hot air, sharply slowing additional inward oxygen transport.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "oxidation",
+        "oxide",
+        "scale"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-hot-air-surface-layer-04",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "hot-air-surface-layer",
+      "target_family": "hot-air-surface-layer",
+      "phrasing_id": "heldout-hot-air-surface-layer-04",
+      "title": "Oxygen-rich high-temperature surface film 4",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "Repeated heating in air creates an external metal-oxygen reaction layer that thickens during each hold and sheds flakes during cooling.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "oxidation",
+        "oxide",
+        "scale"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    },
+    {
+      "slug": "heldout-v1-assoc-hot-air-surface-layer-05",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "hot-air-surface-layer",
+      "target_family": "hot-air-surface-layer",
+      "phrasing_id": "heldout-hot-air-surface-layer-05",
+      "title": "Oxygen-rich high-temperature surface film 5",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A turbine component exposed to hot combustion gas develops a multilayer surface product containing metal cations and oxygen anions.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "oxidation",
+        "oxide",
+        "scale"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    }
+  ]
+}

jlens_materials/prompts/materials-heldout-v1-smoke.json

@@ -0,0 +1,29 @@
+{
+  "format_version": 1,
+  "study": "materials-heldout-v1-smoke",
+  "source_manifest_sha256": "8c034cf33d287d379fddf842971914ec035a22b9e31d29f457258bf85c52e203",
+  "prompts": [
+    {
+      "slug": "heldout-v1-assoc-ductile-01",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "ductile",
+      "target_family": "ductile",
+      "phrasing_id": "heldout-ductile-01",
+      "title": "Energy-absorbing dimpled separation 1",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "text": "A round aluminum bar stretched into a pronounced waist before separating. The mating faces contained hemispherical pits centered on inclusions and a slanted tearing rim.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "ductile",
+        "nucleation",
+        "coalescence",
+        "void"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them."
+    }
+  ]
+}

jlens_materials/prompts/materials-key-concept-traces-v1.json

@@ -0,0 +1,121 @@
+{
+  "description": "Exact replication subset from materials-paper-v2. Four strong cases and two prespecified weak/negative cases are rerun only to persist per-layer J/logit trajectories for auditable figures.",
+  "prompts": [
+    {
+      "slug": "paper-v2-assoc-notch-resistance-04",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "notch-resistance",
+      "title": "Resistance to notch-driven failure 4",
+      "description": "Preregistered unnamed materials vignette; compare fixed-band J-lens ranks with the matched logit-lens baseline.",
+      "text": "During a compact-tension test, substantial energy is dissipated near the notch tip before the specimen separates. The material tolerates larger defects than a comparison alloy.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "toughness",
+        "crack",
+        "fracture"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Primary association-suite item. Inclusion requires tokenizer-level input/output absence. The family-level endpoint is paired pass@k/AUC for J-lens versus logit lens over the fixed 38--92% band."
+    },
+    {
+      "slug": "paper-v2-assoc-ductile-02",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "ductile",
+      "title": "Dimpled tensile failure 2",
+      "description": "Preregistered unnamed materials vignette; compare fixed-band J-lens ranks with the matched logit-lens baseline.",
+      "text": "After substantial plastic elongation, a metal bar separated along a surface filled with microscopic pits around second-phase particles. Neighboring pits had joined into larger depressions.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "ductile",
+        "nucleation",
+        "coalescence",
+        "void"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Primary association-suite item. Inclusion requires tokenizer-level input/output absence. The family-level endpoint is paired pass@k/AUC for J-lens versus logit lens over the fixed 38--92% band."
+    },
+    {
+      "slug": "paper-v2-assoc-line-defect-motion-01",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "line-defect-motion",
+      "title": "Crystal line-imperfection motion 1",
+      "description": "Preregistered unnamed materials vignette; compare fixed-band J-lens ranks with the matched logit-lens baseline.",
+      "text": "Under resolved loading, a line imperfection moves across a close-packed plane in the direction of its mismatch vector, leaving a permanent offset at the free surface.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "dislocation",
+        "glide",
+        "slip"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Primary association-suite item. Inclusion requires tokenizer-level input/output absence. The family-level endpoint is paired pass@k/AUC for J-lens versus logit lens over the fixed 38--92% band."
+    },
+    {
+      "slug": "paper-v2-assoc-rapid-transformation-04",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "rapid-transformation",
+      "title": "Rapid diffusionless steel transformation 4",
+      "description": "Preregistered unnamed materials vignette; compare fixed-band J-lens ranks with the matched logit-lens baseline.",
+      "text": "Suppressing atomic redistribution during cooling converts the parent face-centered phase into a hard acicular product. The new body-centered cell has a c-to-a ratio different from one.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "martensite",
+        "shear",
+        "Bain",
+        "tetragonal"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Primary association-suite item. Inclusion requires tokenizer-level input/output absence. The family-level endpoint is paired pass@k/AUC for J-lens versus logit lens over the fixed 38--92% band."
+    },
+    {
+      "slug": "paper-v2-assoc-cleavage-03",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "cleavage",
+      "title": "Faceted low-temperature failure 3",
+      "description": "Preregistered unnamed materials vignette; compare fixed-band J-lens ranks with the matched logit-lens baseline.",
+      "text": "A thick steel plate failed suddenly in cold service without visible plastic distortion. Microscopy revealed flat planes and converging river patterns cutting across grain interiors.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "brittle",
+        "cleavage",
+        "transgranular"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Primary association-suite item. Inclusion requires tokenizer-level input/output absence. The family-level endpoint is paired pass@k/AUC for J-lens versus logit lens over the fixed 38--92% band."
+    },
+    {
+      "slug": "paper-v2-assoc-hot-air-surface-layer-01",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "hot-air-surface-layer",
+      "title": "High-temperature environmental surface layer 1",
+      "description": "Preregistered unnamed materials vignette; compare fixed-band J-lens ranks with the matched logit-lens baseline.",
+      "text": "A nickel alloy held in hot air develops an adherent multilayer surface film as metal atoms react with oxygen. Continued exposure thickens the film according to an approximately parabolic time law.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "oxidation",
+        "oxide",
+        "scale"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Primary association-suite item. Inclusion requires tokenizer-level input/output absence. The family-level endpoint is paired pass@k/AUC for J-lens versus logit lens over the fixed 38--92% band."
+    }
+  ]
+}

jlens_materials/prompts/materials-mechanism-name-v1.json

@@ -0,0 +1,905 @@
+{
+  "description": "Frozen 50-item mechanism-versus-eponym materials evaluation: five principles, ten independent wordings each.",
+  "prompts": [
+    {
+      "slug": "mechanism-name-grain-size-strength-01",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "grain-size-strength",
+      "title": "Grain-size strengthening 1",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "Two specimens of the same alloy receive different heat treatments. The one with much smaller crystallites requires a larger applied stress before permanent deformation begins.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "boundary",
+        "Hall"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'boundary' against eponym token 'Hall'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-grain-size-strength-02",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "grain-size-strength",
+      "title": "Grain-size strengthening 2",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A polycrystalline metal is processed until its average crystallite diameter is reduced by half. Its resistance to the onset of plastic flow increases.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "boundary",
+        "Hall"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'boundary' against eponym token 'Hall'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-grain-size-strength-03",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "grain-size-strength",
+      "title": "Grain-size strengthening 3",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "An ultrafine-grained steel yields at a higher stress than a coarse-grained specimen with the same composition and phase fractions.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "boundary",
+        "Hall"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'boundary' against eponym token 'Hall'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-grain-size-strength-04",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "grain-size-strength",
+      "title": "Grain-size strengthening 4",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "Repeated rolling and recrystallization create many small crystallites. Moving line defects encounter interfaces more frequently, and the measured yield stress rises.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "boundary",
+        "Hall"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'boundary' against eponym token 'Hall'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-grain-size-strength-05",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "grain-size-strength",
+      "title": "Grain-size strengthening 5",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A metal with coarse crystallites is compared with an otherwise identical fine-crystallite version. The refined material better resists the start of irreversible strain.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "boundary",
+        "Hall"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'boundary' against eponym token 'Hall'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-grain-size-strength-06",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "grain-size-strength",
+      "title": "Grain-size strengthening 6",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "Thermomechanical processing reduces the characteristic crystal-region size without changing chemistry. The resulting sheet shows higher yield strength.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "boundary",
+        "Hall"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'boundary' against eponym token 'Hall'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-grain-size-strength-07",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "grain-size-strength",
+      "title": "Grain-size strengthening 7",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A line defect moving through one crystallite is blocked at the interface with a differently oriented neighbor. Making the crystallites smaller increases the number of such obstacles.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "boundary",
+        "Hall"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'boundary' against eponym token 'Hall'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-grain-size-strength-08",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "grain-size-strength",
+      "title": "Grain-size strengthening 8",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "Two copper samples differ mainly in crystallite size. The sample containing more interfaces per unit volume begins plastic flow at the larger stress.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "boundary",
+        "Hall"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'boundary' against eponym token 'Hall'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-grain-size-strength-09",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "grain-size-strength",
+      "title": "Grain-size strengthening 9",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "Recrystallized grains are refined from tens of micrometers to only a few micrometers. The stress needed to initiate slip increases approximately with the inverse square root of size.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "boundary",
+        "Hall"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'boundary' against eponym token 'Hall'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-grain-size-strength-10",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "grain-size-strength",
+      "title": "Grain-size strengthening 10",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A structural alloy is strengthened by subdividing it into smaller misoriented crystal regions, creating more barriers to the passage of moving line imperfections.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "boundary",
+        "Hall"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'boundary' against eponym token 'Hall'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-flaw-controlled-failure-01",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "flaw-controlled-failure",
+      "title": "Flaw-controlled brittle failure 1",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "Two brittle plates carry the same remote tensile load, but one contains a surface flaw four times longer. The plate with the larger flaw becomes unstable at a lower applied stress.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "crack",
+        "Griffith"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'crack' against eponym token 'Griffith'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-flaw-controlled-failure-02",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "flaw-controlled-failure",
+      "title": "Flaw-controlled brittle failure 2",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A glass specimen with a sharp internal slit fails well below its theoretical bond strength because the local intensity at the slit tip is strongly amplified.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "crack",
+        "Griffith"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'crack' against eponym token 'Griffith'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-flaw-controlled-failure-03",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "flaw-controlled-failure",
+      "title": "Flaw-controlled brittle failure 3",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "Doubling the length of a sharp flaw in an elastic ceramic reduces the remote stress required for unstable extension by roughly the square root of two.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "crack",
+        "Griffith"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'crack' against eponym token 'Griffith'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-flaw-controlled-failure-04",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "flaw-controlled-failure",
+      "title": "Flaw-controlled brittle failure 4",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A brittle solid separates when the elastic energy released by extending a sharp defect exceeds the energy needed to create two new surfaces.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "crack",
+        "Griffith"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'crack' against eponym token 'Griffith'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-flaw-controlled-failure-05",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "flaw-controlled-failure",
+      "title": "Flaw-controlled brittle failure 5",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A polished glass fiber is much stronger than a scratched one of the same diameter because the scratch concentrates tensile loading at its tip.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "crack",
+        "Griffith"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'crack' against eponym token 'Griffith'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-flaw-controlled-failure-06",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "flaw-controlled-failure",
+      "title": "Flaw-controlled brittle failure 6",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "An elastic plate containing a central slit is loaded in tension. Instability begins when the energy-release rate reaches the material's surface-formation resistance.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "crack",
+        "Griffith"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'crack' against eponym token 'Griffith'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-flaw-controlled-failure-07",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "flaw-controlled-failure",
+      "title": "Flaw-controlled brittle failure 7",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A ceramic component survives a given tensile stress when its largest flaw is short, but fails after processing introduces a much longer sharp flaw.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "crack",
+        "Griffith"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'crack' against eponym token 'Griffith'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-flaw-controlled-failure-08",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "flaw-controlled-failure",
+      "title": "Flaw-controlled brittle failure 8",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "The nominal strength of a brittle specimen scales approximately with the inverse square root of the size of its dominant sharp defect.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "crack",
+        "Griffith"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'crack' against eponym token 'Griffith'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-flaw-controlled-failure-09",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "flaw-controlled-failure",
+      "title": "Flaw-controlled brittle failure 9",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A sharp notch in a glass sheet locally magnifies stress, allowing separation to start even though the average applied stress is modest.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "crack",
+        "Griffith"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'crack' against eponym token 'Griffith'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-flaw-controlled-failure-10",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "flaw-controlled-failure",
+      "title": "Flaw-controlled brittle failure 10",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A brittle plate with a pre-existing slit becomes unstable when additional extension lowers the total elastic-plus-surface energy of the system.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "crack",
+        "Griffith"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'crack' against eponym token 'Griffith'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-resolved-slip-threshold-01",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "resolved-slip-threshold",
+      "title": "Resolved slip threshold 1",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A single crystal is pulled along an axis that places 120 MPa of resolved shear on its favored slip system. The critical value is 80 MPa, so microscopic plastic motion begins.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "dislocation",
+        "Schmid"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'dislocation' against eponym token 'Schmid'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-resolved-slip-threshold-02",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "resolved-slip-threshold",
+      "title": "Resolved slip threshold 2",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "An applied uniaxial stress is projected onto a crystal's slip-plane normal and slip direction. Plastic flow starts when that resolved component exceeds a critical threshold.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "dislocation",
+        "Schmid"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'dislocation' against eponym token 'Schmid'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-resolved-slip-threshold-03",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "resolved-slip-threshold",
+      "title": "Resolved slip threshold 3",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "Two identically loaded crystals have different orientations. The crystal with the larger product of the two direction cosines activates its favored slip system first.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "dislocation",
+        "Schmid"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'dislocation' against eponym token 'Schmid'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-resolved-slip-threshold-04",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "resolved-slip-threshold",
+      "title": "Resolved slip threshold 4",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A crystal remains elastic until the shear component acting along a permitted direction on a permitted plane reaches a material-specific critical value.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "dislocation",
+        "Schmid"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'dislocation' against eponym token 'Schmid'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-resolved-slip-threshold-05",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "resolved-slip-threshold",
+      "title": "Resolved slip threshold 5",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "Under tension, only part of the applied load drives shear on a close-packed system. Once this projected driving force is large enough, line imperfections begin to glide.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "dislocation",
+        "Schmid"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'dislocation' against eponym token 'Schmid'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-resolved-slip-threshold-06",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "resolved-slip-threshold",
+      "title": "Resolved slip threshold 6",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "Rotating a single crystal changes the axial stress required to initiate slip even though the critical resolved shear resistance is unchanged.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "dislocation",
+        "Schmid"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'dislocation' against eponym token 'Schmid'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-resolved-slip-threshold-07",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "resolved-slip-threshold",
+      "title": "Resolved slip threshold 7",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A favorably oriented slip system reaches its critical shear value before any other system, producing the first permanent surface steps.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "dislocation",
+        "Schmid"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'dislocation' against eponym token 'Schmid'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-resolved-slip-threshold-08",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "resolved-slip-threshold",
+      "title": "Resolved slip threshold 8",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "The onset of plasticity in a single crystal is predicted by multiplying the applied tensile stress by geometric factors for the plane normal and slip direction.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "dislocation",
+        "Schmid"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'dislocation' against eponym token 'Schmid'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-resolved-slip-threshold-09",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "resolved-slip-threshold",
+      "title": "Resolved slip threshold 9",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "One crystal orientation converts a larger fraction of the applied normal stress into shear on an allowed system, so it yields before a less favorable orientation.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "dislocation",
+        "Schmid"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'dislocation' against eponym token 'Schmid'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-resolved-slip-threshold-10",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "resolved-slip-threshold",
+      "title": "Resolved slip threshold 10",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "When the resolved driving force on the most favorably oriented plane-direction pair exceeds its critical resistance, mobile line defects sweep across the crystal.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "dislocation",
+        "Schmid"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'dislocation' against eponym token 'Schmid'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-lattice-diffusion-creep-01",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "lattice-diffusion-creep",
+      "title": "Lattice-diffusion creep 1",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A fine polycrystal held at high temperature and low stress lengthens as atoms diffuse through crystal interiors from compressed faces toward tensile faces.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "vacancy",
+        "Herring"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'vacancy' against eponym token 'Herring'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-lattice-diffusion-creep-02",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "lattice-diffusion-creep",
+      "title": "Lattice-diffusion creep 2",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "At elevated temperature, atoms move through the lattice between oppositely loaded grain faces, producing slow elongation without extensive line-defect glide.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "vacancy",
+        "Herring"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'vacancy' against eponym token 'Herring'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-lattice-diffusion-creep-03",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "lattice-diffusion-creep",
+      "title": "Lattice-diffusion creep 3",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A polycrystalline solid changes shape under a small sustained load because empty lattice sites migrate through grain interiors in the direction opposite atomic transport.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "vacancy",
+        "Herring"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'vacancy' against eponym token 'Herring'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-lattice-diffusion-creep-04",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "lattice-diffusion-creep",
+      "title": "Lattice-diffusion creep 4",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "Slow deformation at high homologous temperature is controlled by bulk atomic transport, and its rate increases strongly as crystallite size decreases.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "vacancy",
+        "Herring"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'vacancy' against eponym token 'Herring'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-lattice-diffusion-creep-05",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "lattice-diffusion-creep",
+      "title": "Lattice-diffusion creep 5",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "Material is transported through crystal interiors from interfaces under compression to interfaces under tension, gradually elongating the specimen.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "vacancy",
+        "Herring"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'vacancy' against eponym token 'Herring'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-lattice-diffusion-creep-06",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "lattice-diffusion-creep",
+      "title": "Lattice-diffusion creep 6",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A fine-grained ceramic deforms under low stress near its melting temperature through diffusion across the volume of each crystallite rather than along interfaces.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "vacancy",
+        "Herring"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'vacancy' against eponym token 'Herring'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-lattice-diffusion-creep-07",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "lattice-diffusion-creep",
+      "title": "Lattice-diffusion creep 7",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "Atoms leave compressed grain faces, traverse the crystal lattice, and attach at tensile faces, leading to time-dependent shape change.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "vacancy",
+        "Herring"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'vacancy' against eponym token 'Herring'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-lattice-diffusion-creep-08",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "lattice-diffusion-creep",
+      "title": "Lattice-diffusion creep 8",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "The steady deformation rate varies inversely with the square of crystallite size and is controlled by bulk self-diffusion.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "vacancy",
+        "Herring"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'vacancy' against eponym token 'Herring'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-lattice-diffusion-creep-09",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "lattice-diffusion-creep",
+      "title": "Lattice-diffusion creep 9",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A high-temperature specimen elongates as missing-site defects carry mass transport through the interiors of its constituent crystals.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "vacancy",
+        "Herring"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'vacancy' against eponym token 'Herring'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-lattice-diffusion-creep-10",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "lattice-diffusion-creep",
+      "title": "Lattice-diffusion creep 10",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "Under a small tensile load, bulk diffusion moves atoms toward faces normal to the loading direction, producing gradual axial extension.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "vacancy",
+        "Herring"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'vacancy' against eponym token 'Herring'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-diffusionless-lattice-change-01",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "diffusionless-lattice-change",
+      "title": "Diffusionless lattice change 1",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "Rapid cooling converts a face-centered parent lattice into a body-centered product whose vertical cell dimension differs from the two horizontal dimensions.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "tetragonal",
+        "Bain"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'tetragonal' against eponym token 'Bain'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-diffusionless-lattice-change-02",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "diffusionless-lattice-change",
+      "title": "Diffusionless lattice change 2",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A quenched carbon steel forms a hard phase by coordinated atomic motion. The product unit cell is stretched along one axis relative to the other two.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "tetragonal",
+        "Bain"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'tetragonal' against eponym token 'Bain'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-diffusionless-lattice-change-03",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "diffusionless-lattice-change",
+      "title": "Diffusionless lattice change 3",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "Long-range solute redistribution is suppressed during cooling, leaving a body-centered cell with an axial ratio different from unity.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "tetragonal",
+        "Bain"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'tetragonal' against eponym token 'Bain'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-diffusionless-lattice-change-04",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "diffusionless-lattice-change",
+      "title": "Diffusionless lattice change 4",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "The parent face-centered cell transforms through compression along one crystallographic axis and expansion along the other two, creating a distorted body-centered product.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "tetragonal",
+        "Bain"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'tetragonal' against eponym token 'Bain'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-diffusionless-lattice-change-05",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "diffusionless-lattice-change",
+      "title": "Diffusionless lattice change 5",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A rapid diffusionless transformation produces an acicular phase whose unit cell has two equal short axes and one longer axis.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "tetragonal",
+        "Bain"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'tetragonal' against eponym token 'Bain'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-diffusionless-lattice-change-06",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "diffusionless-lattice-change",
+      "title": "Diffusionless lattice change 6",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "Carbon trapped during quenching distorts the body-centered product so that its c dimension is not equal to a and b.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "tetragonal",
+        "Bain"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'tetragonal' against eponym token 'Bain'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-diffusionless-lattice-change-07",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "diffusionless-lattice-change",
+      "title": "Diffusionless lattice change 7",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "A coordinated lattice deformation maps the parent cubic arrangement into a body-centered cell elongated in one direction.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "tetragonal",
+        "Bain"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'tetragonal' against eponym token 'Bain'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-diffusionless-lattice-change-08",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "diffusionless-lattice-change",
+      "title": "Diffusionless lattice change 8",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "The hard lath-shaped product of severe quenching has a c-to-a lattice ratio measurably greater than one.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "tetragonal",
+        "Bain"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'tetragonal' against eponym token 'Bain'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-diffusionless-lattice-change-09",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "diffusionless-lattice-change",
+      "title": "Diffusionless lattice change 9",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "Atoms move cooperatively over short distances during a rapid transformation, generating a body-centered structure with unequal axial dimensions.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "tetragonal",
+        "Bain"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'tetragonal' against eponym token 'Bain'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    },
+    {
+      "slug": "mechanism-name-diffusionless-lattice-change-10",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "diffusionless-lattice-change",
+      "title": "Diffusionless lattice change 10",
+      "description": "Frozen mechanism-versus-eponym vignette; both tracked words are absent from the input and generated output.",
+      "text": "X-ray peaks from the quenched product split because one unit-cell axis differs from the two equivalent perpendicular axes.",
+      "readout_selector": "final_prompt_token",
+      "tracked": [
+        "tetragonal",
+        "Bain"
+      ],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Paired comparison of physical carrier 'tetragonal' against eponym token 'Bain'. Primary endpoint is the within-item log10 rank difference, summarized separately for J-lens and logit lens."
+    }
+  ]
+}

jlens_materials/prompts/materials-paper-case-studies-v1.json

@@ -0,0 +1,231 @@
+{
+  "description": "Preregistered exploratory Gemma-4 materials case studies. Four unnamed materials vignettes test latent scientific associations, four constrained-answer items test mechanism/recognition traces, and two matched carrier triplets replicate directed modulation. This pack is designed for documented qualitative examples, not population-level quantitative claims.",
+  "prompts": [
+    {
+      "slug": "paper-v1-ductile-fracture-replication",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "title": "Ductile-fracture replication vignette",
+      "description": "An independently worded fractography vignette tests replication of the ductile microvoid family.",
+      "text": "A tensile bar exhibited extensive necking before separation. The fracture surface was covered by rounded dimples centered on inclusions, and the final profile had a cup-shaped center with a surrounding shear lip.",
+      "readout_selector": "final_prompt_token",
+      "tracked": ["ductile", "void", "nucleation", "coalescence"],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Replication succeeds qualitatively if the ductile/microvoid family improves strongly under the J-lens relative to the logit lens. Do not infer an ordered mechanism unless nucleation precedes coalescence reproducibly."
+    },
+    {
+      "slug": "paper-v1-intergranular-replication",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "title": "Boundary-attack replication vignette",
+      "description": "An independently worded stainless-steel vignette tests replication of the corrosion and sensitization family.",
+      "text": "A stainless-steel component spent hours near 650 degrees Celsius, allowing chromium-rich carbides to decorate grain interfaces. After immersion in a chloride solution, narrow connected grooves followed those interfaces while the grain interiors remained nearly untouched.",
+      "readout_selector": "final_prompt_token",
+      "tracked": ["boundary", "corrosion", "sensitization", "intergranular"],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "The strongest result would recover both generic corrosion and the specific sensitization/intergranular interpretation without those terms appearing in the prompt or continuation."
+    },
+    {
+      "slug": "paper-v1-fatigue-fractography",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "title": "Cyclic-fracture vignette",
+      "description": "Fractographic and loading clues should evoke the cyclic damage family without naming it.",
+      "text": "A shaft survived thousands of alternating load cycles before sudden separation. Concentric beach marks spread from a surface notch, and microscopy showed closely spaced parallel ridges advancing across the smooth region.",
+      "readout_selector": "final_prompt_token",
+      "tracked": ["fatigue", "crack", "striations", "propagation"],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "This tests whether a coherent fatigue/crack-growth family emerges from loading history and fractographic morphology."
+    },
+    {
+      "slug": "paper-v1-cleavage-fractography",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "title": "Low-temperature faceted-fracture vignette",
+      "description": "A faceted low-temperature fracture should evoke a brittle cleavage interpretation.",
+      "text": "A body-centered alloy failed at low temperature with almost no necking. Its fracture surface contained flat crystallographic facets joined by river-like markings, while macroscopic plastic deformation was minimal.",
+      "readout_selector": "final_prompt_token",
+      "tracked": ["brittle", "cleavage", "transgranular", "fracture"],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Compare the specific cleavage/transgranular terms with the broader brittle/fracture family."
+    },
+    {
+      "slug": "paper-v1-hall-petch-constrained",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "title": "Constrained grain-size strengthening",
+      "description": "A forced binary completion prevents an adverb from invalidating the clean answer control.",
+      "text": "Two otherwise identical polycrystals have mean grain sizes of 60 micrometers and 6 micrometers. Interfaces impede carrier motion, so reducing the grain size changes the yield stress. Choose exactly one word, higher or lower: the fine-grained specimen's yield stress is",
+      "readout_selector": "before_answer",
+      "tracked": ["Hall", "Petch", "strengthening"],
+      "answer": "higher",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "This repairs the earlier clean-answer failure caused by the first generated token being 'significantly'."
+    },
+    {
+      "slug": "paper-v1-schmid-constrained",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "title": "Constrained resolved-shear event",
+      "description": "A one-word microscopic-event request distinguishes slip from the macroscopic answer yield.",
+      "text": "For a single crystal, the product of applied stress and the two orientation cosines gives a resolved value of 120 MPa on the most favorably oriented system. The critical value is 80 MPa. Answer with one word naming the microscopic deformation event that begins:",
+      "readout_selector": "before_answer",
+      "tracked": ["Schmid", "dislocation"],
+      "answer": "slip",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "This repairs the earlier ambiguity in which the scientifically plausible first token was 'yield'."
+    },
+    {
+      "slug": "paper-v1-fcc-constrained",
+      "shape": "RECOGNITION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "title": "Constrained diffraction recognition",
+      "description": "A strict abbreviation request prevents an initial article from invalidating the answer check.",
+      "text": "A cubic powder pattern permits reflections when the three Miller indices are all odd or all even and excludes mixed parity. Reply only with the standard three-letter Bravais-lattice abbreviation:",
+      "readout_selector": "before_answer",
+      "tracked": ["cubic", "extinction", "diffraction"],
+      "answer": "FCC",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "This repairs the earlier clean-answer failure caused by the first generated token being the article 'a'."
+    },
+    {
+      "slug": "paper-v1-griffith-constrained",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "title": "Constrained flaw-size scaling",
+      "description": "A binary comparative answer tests whether the named fracture criterion appears before the scaling result.",
+      "text": "Two elastic plates have identical modulus and surface energy, but one contains an internal flaw four times as long as the other. Choose exactly one word, higher or lower: the critical remote tensile stress for the plate with the longer flaw is",
+      "readout_selector": "before_answer",
+      "tracked": ["Griffith", "crack", "fracture"],
+      "answer": "lower",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "The answer follows the inverse-square-root flaw-size scaling; Griffith is the predeclared mechanism-level concept."
+    },
+    {
+      "slug": "paper-v1-pileup-focus-key",
+      "shape": "MODULATION",
+      "protocol": "directed_modulation",
+      "domain": "materials",
+      "title": "Pile-up focus replication: brass key",
+      "description": "First new teacher-forced carrier for the directed-modulation replication.",
+      "user": "Write \"The brass key rested under the folded map.\" Concentrate on a dislocation pile-up and its stress concentration while you write the sentence. Do not write anything else.",
+      "assistant_prefill": "The brass key rested under the folded map.",
+      "readout_selector": "assistant_response",
+      "tracked": ["dislocation", "pile", "stress"],
+      "must_be_absent_from_output": true,
+      "condition": "focus",
+      "control_group": "paper-v1-pileup-brass-key",
+      "phrasing_id": "focus-02",
+      "carrier_id": "brass-key",
+      "target_family": "dislocation-pileup",
+      "note": "Interpret only relative to the matched suppress and neutral records using the identical carrier and target set."
+    },
+    {
+      "slug": "paper-v1-pileup-suppress-key",
+      "shape": "MODULATION",
+      "protocol": "directed_modulation",
+      "domain": "materials",
+      "title": "Pile-up suppression replication: brass key",
+      "description": "Matched suppress condition for the brass-key carrier.",
+      "user": "Write \"The brass key rested under the folded map.\" Try not to think about a dislocation pile-up or its stress concentration while you write the sentence. Do not write anything else.",
+      "assistant_prefill": "The brass key rested under the folded map.",
+      "readout_selector": "assistant_response",
+      "tracked": ["dislocation", "pile", "stress"],
+      "must_be_absent_from_output": true,
+      "condition": "suppress",
+      "control_group": "paper-v1-pileup-brass-key",
+      "phrasing_id": "suppress-02",
+      "carrier_id": "brass-key",
+      "target_family": "dislocation-pileup"
+    },
+    {
+      "slug": "paper-v1-pileup-control-key",
+      "shape": "MODULATION",
+      "protocol": "directed_modulation",
+      "domain": "materials",
+      "title": "Pile-up neutral control: brass key",
+      "description": "Matched neutral condition for the brass-key carrier.",
+      "user": "Write \"The brass key rested under the folded map.\" Do not write anything else.",
+      "assistant_prefill": "The brass key rested under the folded map.",
+      "readout_selector": "assistant_response",
+      "tracked": ["dislocation", "pile", "stress"],
+      "must_be_absent_from_output": true,
+      "condition": "control",
+      "control_group": "paper-v1-pileup-brass-key",
+      "phrasing_id": "control-02",
+      "carrier_id": "brass-key",
+      "target_family": "dislocation-pileup"
+    },
+    {
+      "slug": "paper-v1-pileup-focus-cup",
+      "shape": "MODULATION",
+      "protocol": "directed_modulation",
+      "domain": "materials",
+      "title": "Pile-up focus replication: white cup",
+      "description": "Second new teacher-forced carrier for the directed-modulation replication.",
+      "user": "Write \"A white cup stood beside the closed window.\" Keep a dislocation pile-up and the stress it concentrates in mind while you write the sentence. Do not write anything else.",
+      "assistant_prefill": "A white cup stood beside the closed window.",
+      "readout_selector": "assistant_response",
+      "tracked": ["dislocation", "pile", "stress"],
+      "must_be_absent_from_output": true,
+      "condition": "focus",
+      "control_group": "paper-v1-pileup-white-cup",
+      "phrasing_id": "focus-03",
+      "carrier_id": "white-cup",
+      "target_family": "dislocation-pileup",
+      "note": "Interpret only relative to the matched suppress and neutral records using the identical carrier and target set."
+    },
+    {
+      "slug": "paper-v1-pileup-suppress-cup",
+      "shape": "MODULATION",
+      "protocol": "directed_modulation",
+      "domain": "materials",
+      "title": "Pile-up suppression replication: white cup",
+      "description": "Matched suppress condition for the white-cup carrier.",
+      "user": "Write \"A white cup stood beside the closed window.\" Try not to think about a dislocation pile-up or the stress it concentrates while you write the sentence. Do not write anything else.",
+      "assistant_prefill": "A white cup stood beside the closed window.",
+      "readout_selector": "assistant_response",
+      "tracked": ["dislocation", "pile", "stress"],
+      "must_be_absent_from_output": true,
+      "condition": "suppress",
+      "control_group": "paper-v1-pileup-white-cup",
+      "phrasing_id": "suppress-03",
+      "carrier_id": "white-cup",
+      "target_family": "dislocation-pileup"
+    },
+    {
+      "slug": "paper-v1-pileup-control-cup",
+      "shape": "MODULATION",
+      "protocol": "directed_modulation",
+      "domain": "materials",
+      "title": "Pile-up neutral control: white cup",
+      "description": "Matched neutral condition for the white-cup carrier.",
+      "user": "Write \"A white cup stood beside the closed window.\" Do not write anything else.",
+      "assistant_prefill": "A white cup stood beside the closed window.",
+      "readout_selector": "assistant_response",
+      "tracked": ["dislocation", "pile", "stress"],
+      "must_be_absent_from_output": true,
+      "condition": "control",
+      "control_group": "paper-v1-pileup-white-cup",
+      "phrasing_id": "control-03",
+      "carrier_id": "white-cup",
+      "target_family": "dislocation-pileup"
+    }
+  ]
+}

jlens_materials/prompts/materials-qualitative-pack.json

@@ -0,0 +1,228 @@
+{
+  "description": "Curated materials-science case studies for close qualitative inspection. These items exercise fixed score positions and controls, but the set is intentionally too small for paper-scale quantitative claims.",
+  "prompts": [
+    {
+      "slug": "materials-hall-petch",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "title": "Grain refinement to yield strength",
+      "description": "Tests whether the named strengthening law surfaces before a one-word comparative answer.",
+      "text": "Two otherwise identical polycrystalline samples have average crystallite diameters of 50 micrometers and 5 micrometers. In the finer sample, interfaces interrupt carrier motion much more frequently. Its yield stress should therefore be",
+      "readout_selector": "before_answer",
+      "tracked": ["Hall", "Petch", "strengthening"],
+      "answer": "higher",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Inspect whether Hall/Petch rises before the answer, whether onset is sustained in the fixed band, and whether its J-lens rank improves materially over the matched logit-lens rank."
+    },
+    {
+      "slug": "materials-orowan-spacing",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "title": "Obstacle spacing to bypass stress",
+      "description": "A particle-spacing change should evoke the bypass mechanism before the direction of the stress change.",
+      "text": "A metal contains a fixed volume fraction of hard, non-shearable nanoscale particles. When their center-to-center spacing is reduced, lattice line defects must bow more sharply between them. The stress required for passage should",
+      "readout_selector": "before_answer",
+      "tracked": ["Orowan", "bowing", "precipitation", "strengthening"],
+      "answer": "increase",
+      "acceptable_answers": ["rise"],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Look for Orowan as the specific latent mechanism rather than only generic strengthening. Compare it with the Hall-Petch case to test mechanism specificity."
+    },
+    {
+      "slug": "materials-martensite-quench",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "title": "Rapid quench to martensite",
+      "description": "Tests whether transformation-mechanism concepts appear before the product name.",
+      "text": "Austenitic steel is cooled so rapidly that long-range atomic redistribution is suppressed. The product is hard and needle-like, and its body-centered unit cell is elongated so that c divided by a is not equal to one. Its crystal symmetry is",
+      "readout_selector": "before_answer",
+      "tracked": ["martensite", "shear", "Bain"],
+      "answer": "tetragonal",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "The answer is the lattice symmetry; shear and Bain encode the transformation mechanism. Gemma tokenizes martensite into multiple pieces, so the report should explicitly list that exact phase name as dropped rather than silently approximating it."
+    },
+    {
+      "slug": "materials-schmid-slip",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "title": "Resolved stress to slip",
+      "description": "An orientation-factor calculation should evoke the named criterion before the deformation event.",
+      "text": "A single crystal under uniaxial load has cos(phi) = 0.8 and cos(lambda) = 0.5. Multiplying these orientation factors by the applied stress gives 120 MPa, above the critical resolved value of 80 MPa. The crystal will begin to",
+      "readout_selector": "before_answer",
+      "tracked": ["Schmid", "dislocation"],
+      "answer": "slip",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Check whether Schmid appears as a law-level concept and whether dislocation appears as the physical carrier before the clean one-token answer."
+    },
+    {
+      "slug": "materials-nabarro-herring",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "title": "High-temperature deformation mechanism",
+      "description": "The transport path and loading regime imply a named creep mechanism before the transport-process answer.",
+      "text": "A fine-grained polycrystal is held near its melting temperature under low stress. Atoms migrate through grain interiors from compressed boundaries toward tensile boundaries, producing slow shape change. The dominant transport process is",
+      "readout_selector": "before_answer",
+      "tracked": ["creep", "Nabarro", "Herring"],
+      "answer": "diffusion",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Compare generic creep with the more specific Nabarro concept. If only creep surfaces, the representation may be scientifically broad but not mechanism-specific."
+    },
+    {
+      "slug": "materials-screw-character",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "title": "Line-defect character",
+      "description": "Geometric information should evoke the standard defect vocabulary before the character label.",
+      "text": "For a line defect in a crystal, the displacement vector associated with the lattice mismatch is parallel to the defect's line direction. Its character is",
+      "readout_selector": "before_answer",
+      "tracked": ["Burgers", "dislocation"],
+      "answer": "screw",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "This separates recovery of the formal descriptor Burgers from the generic object class dislocation. Inspect their relative onset order rather than only their best ranks."
+    },
+    {
+      "slug": "materials-fcc-extinctions",
+      "shape": "RECOGNITION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "title": "Diffraction selection-rule recognition",
+      "description": "A raw reflection rule must be mapped to a crystal-structure label.",
+      "text": "A powder diffraction pattern contains reflections only when h, k, and l are all odd or all even; mixed-parity reflections are systematically absent. The corresponding Bravais lattice is",
+      "readout_selector": "before_answer",
+      "tracked": ["cubic", "extinction", "close-packed"],
+      "answer": "FCC",
+      "acceptable_answers": ["face"],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "This is a recognition test rather than a prose association. Check that cubic rises from the extinction rule and is not merely present at the final layer."
+    },
+    {
+      "slug": "materials-ductile-fractography",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "title": "Unnamed ductile-fracture vignette",
+      "description": "A fractographic description evokes a failure mechanism without naming it.",
+      "text": "The tensile specimen formed a pronounced waist before separation. Its failure surface contains rounded microscopic pits, each surrounding a second-phase particle, joined into a cup-and-cone profile.",
+      "readout_selector": "final_prompt_token",
+      "tracked": ["ductile", "microvoid", "void", "nucleation", "coalescence"],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "A strong result distinguishes the specific microvoid/coalescence mechanism from the broader adjective ductile. The generated-output absence control prevents counting a concept the model simply says next."
+    },
+    {
+      "slug": "materials-intergranular-attack",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "title": "Unnamed boundary-localized degradation",
+      "description": "A processing and microscopy vignette should evoke a specific degradation mode.",
+      "text": "An austenitic stainless steel was held in a temperature range where carbides formed at crystallite edges. After exposure to an aggressive solution, attack traces connected paths along those edges while the grain interiors remain largely intact.",
+      "readout_selector": "final_prompt_token",
+      "tracked": ["intergranular", "boundary", "corrosion", "sensitization"],
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Inspect specificity: intergranular plus sensitization is more informative than generic corrosion. Tokenizer-level input checks catch accidental leakage."
+    },
+    {
+      "slug": "materials-ductile-brittle-swap",
+      "shape": "PROBE_SWAP",
+      "protocol": "probe_swap",
+      "domain": "materials",
+      "category": "failure-response",
+      "title": "Ductile-to-brittle causal counterfactual",
+      "description": "Tests whether replacing a latent failure-response coordinate changes the declared downstream answer.",
+      "text": "A pure metal has many easy close-packed slip systems and accommodates extensive glide before separation. In a tensile test beyond yield it normally shows substantial",
+      "readout_selector": "before_answer",
+      "tracked": ["ductile"],
+      "answer": "elongation",
+      "swap_from": "ductile",
+      "swap_to": "brittle",
+      "swap_answer": "fracture",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Grade the clean and counterfactual output tokens. A target-rank improvement without the declared answer change is not causal success."
+    },
+    {
+      "slug": "materials-pileup-focus",
+      "shape": "MODULATION",
+      "protocol": "directed_modulation",
+      "domain": "materials",
+      "title": "Focus on a dislocation pile-up",
+      "description": "Teacher-forced carrier with a materials side concept.",
+      "user": "Write \"The blue notebook remained beside the furnace.\" Concentrate on a dislocation pile-up and its stress concentration while you write the sentence. Do not write anything else.",
+      "assistant_prefill": "The blue notebook remained beside the furnace.",
+      "readout_selector": "assistant_response",
+      "tracked": ["dislocation", "pile", "stress"],
+      "must_be_absent_from_output": true,
+      "condition": "focus",
+      "control_group": "materials-pileup-notebook",
+      "phrasing_id": "focus-01",
+      "carrier_id": "blue-notebook",
+      "target_family": "dislocation-pileup",
+      "note": "Compare only against the matched suppress and neutral records; the absolute rank by itself is not the modulation result."
+    },
+    {
+      "slug": "materials-pileup-suppress",
+      "shape": "MODULATION",
+      "protocol": "directed_modulation",
+      "domain": "materials",
+      "title": "Suppress a dislocation pile-up",
+      "description": "Matched suppress condition using the identical carrier and target set.",
+      "user": "Write \"The blue notebook remained beside the furnace.\" Try not to think about a dislocation pile-up or its stress concentration while you write the sentence. Do not write anything else.",
+      "assistant_prefill": "The blue notebook remained beside the furnace.",
+      "readout_selector": "assistant_response",
+      "tracked": ["dislocation", "pile", "stress"],
+      "must_be_absent_from_output": true,
+      "condition": "suppress",
+      "control_group": "materials-pileup-notebook",
+      "phrasing_id": "suppress-01",
+      "carrier_id": "blue-notebook",
+      "target_family": "dislocation-pileup"
+    },
+    {
+      "slug": "materials-pileup-control",
+      "shape": "MODULATION",
+      "protocol": "directed_modulation",
+      "domain": "materials",
+      "title": "Neutral carrier control",
+      "description": "Matched carrier with no materials side task.",
+      "user": "Write \"The blue notebook remained beside the furnace.\" Do not write anything else.",
+      "assistant_prefill": "The blue notebook remained beside the furnace.",
+      "readout_selector": "assistant_response",
+      "tracked": ["dislocation", "pile", "stress"],
+      "must_be_absent_from_output": true,
+      "condition": "control",
+      "control_group": "materials-pileup-notebook",
+      "phrasing_id": "control-01",
+      "carrier_id": "blue-notebook",
+      "target_family": "dislocation-pileup"
+    },
+    {
+      "slug": "materials-defect-report",
+      "shape": "REPORT_SWAP",
+      "protocol": "verbal_report",
+      "domain": "materials",
+      "title": "Crystal-defect verbal-report swap",
+      "description": "Dynamically select the clean one-word report, then try sibling defect coordinates.",
+      "user": "Think of a crystalline defect. Answer in one word.",
+      "readout_selector": "final_prompt_token",
+      "tracked": ["vacancy", "dislocation", "interstitial", "twin", "boundary"],
+      "swap_to": "interstitial",
+      "candidates": ["vacancy", "dislocation", "interstitial", "twin", "boundary"],
+      "note": "Inspect every recorded candidate trial. The primary diagnostic is whether the one-word report changes to the installed sibling, not whether its internal rank merely improves."
+    }
+  ]
+}

jlens_materials/prompts/materials-reasoning-pilot-v1.json

@@ -0,0 +1,409 @@
+{
+  "description": "Frozen chat-formatted clean-answer pilot.",
+  "prompts": [
+    {
+      "slug": "reason-hall-01",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "grain-size",
+      "title": "Grain-size comparison",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "Two otherwise identical polycrystals have mean grain sizes 50 and 5 micrometers. Grain interfaces impede carrier motion. Which specimen has the larger yield stress? Reply with exactly one word: fine or coarse.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "Hall",
+        "Petch",
+        "strengthening",
+        "dislocation"
+      ],
+      "answer": "fine",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    },
+    {
+      "slug": "reason-hall-02",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "grain-size",
+      "title": "Grain-size comparison",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "Two otherwise identical polycrystals have mean grain sizes 100 and 10 micrometers. Grain interfaces impede carrier motion. Which specimen has the larger yield stress? Reply with exactly one word: fine or coarse.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "Hall",
+        "Petch",
+        "strengthening",
+        "dislocation"
+      ],
+      "answer": "fine",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    },
+    {
+      "slug": "reason-hall-03",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "grain-size",
+      "title": "Grain-size comparison",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "Two otherwise identical polycrystals have mean grain sizes 40 and 4 micrometers. Grain interfaces impede carrier motion. Which specimen has the larger yield stress? Reply with exactly one word: fine or coarse.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "Hall",
+        "Petch",
+        "strengthening",
+        "dislocation"
+      ],
+      "answer": "fine",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    },
+    {
+      "slug": "reason-hall-04",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "grain-size",
+      "title": "Grain-size comparison",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "Two otherwise identical polycrystals have mean grain sizes 80 and 8 micrometers. Grain interfaces impede carrier motion. Which specimen has the larger yield stress? Reply with exactly one word: fine or coarse.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "Hall",
+        "Petch",
+        "strengthening",
+        "dislocation"
+      ],
+      "answer": "fine",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    },
+    {
+      "slug": "reason-griffith-01",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "flaw-scaling",
+      "title": "Flaw-size scaling",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "Two elastic plates have the same modulus and surface energy. One contains an internal elliptical flaw 2 times as long as the other. Is the critical remote tensile stress of the longer-flaw plate higher or lower? Reply with exactly one word.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "Griffith",
+        "crack",
+        "fracture"
+      ],
+      "answer": "lower",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    },
+    {
+      "slug": "reason-griffith-02",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "flaw-scaling",
+      "title": "Flaw-size scaling",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "Two elastic plates have the same modulus and surface energy. One contains an internal elliptical flaw 4 times as long as the other. Is the critical remote tensile stress of the longer-flaw plate higher or lower? Reply with exactly one word.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "Griffith",
+        "crack",
+        "fracture"
+      ],
+      "answer": "lower",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    },
+    {
+      "slug": "reason-griffith-03",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "flaw-scaling",
+      "title": "Flaw-size scaling",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "Two elastic plates have the same modulus and surface energy. One contains an internal elliptical flaw 9 times as long as the other. Is the critical remote tensile stress of the longer-flaw plate higher or lower? Reply with exactly one word.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "Griffith",
+        "crack",
+        "fracture"
+      ],
+      "answer": "lower",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    },
+    {
+      "slug": "reason-griffith-04",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "flaw-scaling",
+      "title": "Flaw-size scaling",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "Two elastic plates have the same modulus and surface energy. One contains an internal elliptical flaw 16 times as long as the other. Is the critical remote tensile stress of the longer-flaw plate higher or lower? Reply with exactly one word.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "Griffith",
+        "crack",
+        "fracture"
+      ],
+      "answer": "lower",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    },
+    {
+      "slug": "reason-schmid-01",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "resolved-threshold",
+      "title": "Resolved-force threshold",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "A single crystal has a resolved driving value of 120 MPa on its most favorably oriented system; the critical value is 80 MPa. Does microscopic plastic motion begin, or does the crystal remain elastic? Reply with exactly one word: slip or elastic.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "Schmid",
+        "dislocation"
+      ],
+      "answer": "slip",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    },
+    {
+      "slug": "reason-schmid-02",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "resolved-threshold",
+      "title": "Resolved-force threshold",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "A single crystal has a resolved driving value of 60 MPa on its most favorably oriented system; the critical value is 80 MPa. Does microscopic plastic motion begin, or does the crystal remain elastic? Reply with exactly one word: slip or elastic.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "Schmid",
+        "dislocation"
+      ],
+      "answer": "elastic",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    },
+    {
+      "slug": "reason-schmid-03",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "resolved-threshold",
+      "title": "Resolved-force threshold",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "A single crystal has a resolved driving value of 95 MPa on its most favorably oriented system; the critical value is 70 MPa. Does microscopic plastic motion begin, or does the crystal remain elastic? Reply with exactly one word: slip or elastic.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "Schmid",
+        "dislocation"
+      ],
+      "answer": "slip",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    },
+    {
+      "slug": "reason-schmid-04",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "resolved-threshold",
+      "title": "Resolved-force threshold",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "A single crystal has a resolved driving value of 45 MPa on its most favorably oriented system; the critical value is 65 MPa. Does microscopic plastic motion begin, or does the crystal remain elastic? Reply with exactly one word: slip or elastic.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "Schmid",
+        "dislocation"
+      ],
+      "answer": "elastic",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    },
+    {
+      "slug": "reason-particle-01",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "particle-spacing",
+      "title": "Obstacle-spacing comparison",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "A metal contains a fixed volume fraction of hard nonshearable particles. The center-to-center particle spacing decreases, so moving lattice line defects must curve between the obstacles. Is the required applied stress higher or lower? Reply with exactly one word.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "Orowan",
+        "bowing",
+        "precipitation",
+        "strengthening"
+      ],
+      "answer": "higher",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    },
+    {
+      "slug": "reason-particle-02",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "particle-spacing",
+      "title": "Obstacle-spacing comparison",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "A metal contains a fixed volume fraction of hard nonshearable particles. The center-to-center particle spacing increases, so moving lattice line defects must curve between the obstacles. Is the required applied stress higher or lower? Reply with exactly one word.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "Orowan",
+        "bowing",
+        "precipitation",
+        "strengthening"
+      ],
+      "answer": "lower",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    },
+    {
+      "slug": "reason-particle-03",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "particle-spacing",
+      "title": "Obstacle-spacing comparison",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "A metal contains a fixed volume fraction of hard nonshearable particles. The center-to-center particle spacing halves, so moving lattice line defects must curve between the obstacles. Is the required applied stress higher or lower? Reply with exactly one word.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "Orowan",
+        "bowing",
+        "precipitation",
+        "strengthening"
+      ],
+      "answer": "higher",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    },
+    {
+      "slug": "reason-particle-04",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "particle-spacing",
+      "title": "Obstacle-spacing comparison",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "A metal contains a fixed volume fraction of hard nonshearable particles. The center-to-center particle spacing doubles, so moving lattice line defects must curve between the obstacles. Is the required applied stress higher or lower? Reply with exactly one word.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "Orowan",
+        "bowing",
+        "precipitation",
+        "strengthening"
+      ],
+      "answer": "lower",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    },
+    {
+      "slug": "reason-intensity-01",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "intensity-threshold",
+      "title": "Stress-intensity threshold",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "A notched plate has applied mode-I intensity 20 MPa square-root-meter and a critical intensity 40 MPa square-root-meter. Is rapid extension stable or unstable? Reply with exactly one word.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "toughness",
+        "crack",
+        "fracture"
+      ],
+      "answer": "stable",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    },
+    {
+      "slug": "reason-intensity-02",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "intensity-threshold",
+      "title": "Stress-intensity threshold",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "A notched plate has applied mode-I intensity 50 MPa square-root-meter and a critical intensity 40 MPa square-root-meter. Is rapid extension stable or unstable? Reply with exactly one word.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "toughness",
+        "crack",
+        "fracture"
+      ],
+      "answer": "unstable",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    },
+    {
+      "slug": "reason-intensity-03",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "intensity-threshold",
+      "title": "Stress-intensity threshold",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "A notched plate has applied mode-I intensity 15 MPa square-root-meter and a critical intensity 30 MPa square-root-meter. Is rapid extension stable or unstable? Reply with exactly one word.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "toughness",
+        "crack",
+        "fracture"
+      ],
+      "answer": "stable",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    },
+    {
+      "slug": "reason-intensity-04",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "category": "intensity-threshold",
+      "title": "Stress-intensity threshold",
+      "description": "Chat-formatted one-word reasoning pilot with a preregistered latent mechanism.",
+      "user": "A notched plate has applied mode-I intensity 45 MPa square-root-meter and a critical intensity 35 MPa square-root-meter. Is rapid extension stable or unstable? Reply with exactly one word.",
+      "readout_selector": "before_answer",
+      "tracked": [
+        "toughness",
+        "crack",
+        "fracture"
+      ],
+      "answer": "unstable",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true,
+      "note": "Pilot inclusion requires the clean next token to equal the declared one-word answer."
+    }
+  ]
+}

jlens_materials/prompts/materials-report-swap-dose-v1.json

@@ -0,0 +1,70 @@
+{
+  "description": "Exploratory verbal-report intervention dose study across five materials categories.",
+  "prompts": [
+    {
+      "slug": "dose-report-defect",
+      "shape": "REPORT_SWAP",
+      "protocol": "verbal_report",
+      "domain": "materials",
+      "category": "defect",
+      "title": "Crystalline defect report",
+      "description": "Dynamically select the clean report and install sibling defect coordinates.",
+      "user": "Think of a crystalline defect. Answer in one word.",
+      "readout_selector": "final_prompt_token",
+      "tracked": ["vacancy", "dislocation", "interstitial", "twin", "boundary"],
+      "candidates": ["vacancy", "dislocation", "interstitial", "twin", "boundary"]
+    },
+    {
+      "slug": "dose-report-failure",
+      "shape": "REPORT_SWAP",
+      "protocol": "verbal_report",
+      "domain": "materials",
+      "category": "failure-mode",
+      "title": "Failure-mode report",
+      "description": "Dynamically select the clean report and install sibling failure-mode coordinates.",
+      "user": "Think of a materials failure mode. Answer in one word.",
+      "readout_selector": "final_prompt_token",
+      "tracked": ["ductile", "brittle", "fatigue", "cleavage", "creep"],
+      "candidates": ["ductile", "brittle", "fatigue", "cleavage", "creep"]
+    },
+    {
+      "slug": "dose-report-lattice",
+      "shape": "REPORT_SWAP",
+      "protocol": "verbal_report",
+      "domain": "materials",
+      "category": "lattice",
+      "title": "Crystal-lattice report",
+      "description": "Dynamically select the clean report and install sibling lattice coordinates.",
+      "user": "Think of a common crystal lattice. Answer using one abbreviation.",
+      "readout_selector": "final_prompt_token",
+      "tracked": ["FCC", "BCC", "HCP"],
+      "candidates": ["FCC", "BCC", "HCP"]
+    },
+    {
+      "slug": "dose-report-deformation",
+      "shape": "REPORT_SWAP",
+      "protocol": "verbal_report",
+      "domain": "materials",
+      "category": "deformation",
+      "title": "Deformation-mechanism report",
+      "description": "Dynamically select the clean report and install sibling deformation coordinates.",
+      "user": "Think of a material deformation mechanism. Answer in one word.",
+      "readout_selector": "final_prompt_token",
+      "tracked": ["slip", "twinning", "creep", "diffusion"],
+      "candidates": ["slip", "twinning", "creep", "diffusion"]
+    },
+    {
+      "slug": "dose-report-strengthening",
+      "shape": "REPORT_SWAP",
+      "protocol": "verbal_report",
+      "domain": "materials",
+      "category": "strengthening",
+      "title": "Strengthening-mechanism report",
+      "description": "Dynamically select the clean report and install sibling strengthening coordinates.",
+      "user": "Think of a metal strengthening mechanism. Answer in one word.",
+      "readout_selector": "final_prompt_token",
+      "tracked": ["precipitation", "solution", "grain", "work"],
+      "candidates": ["precipitation", "solution", "grain", "work"]
+    }
+  ]
+}

jlens_materials/prompts/mechanics-paper-example.json

@@ -0,0 +1,118 @@
+{
+  "description": "Small schema example only. Expand each evaluation distribution to at least 50 independent items before quantitative reporting.",
+  "prompts": [
+    {
+      "slug": "griffith-length-example",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "mechanics",
+      "title": "Brittle flaw scaling",
+      "description": "Input-absent mechanism before the expected answer.",
+      "text": "A brittle elastic plate contains a sharp internal flaw. Its failure load follows from balancing stored work released during advance against the cost of forming two fresh faces. With material properties fixed, the critical load is inversely proportional to the square root of the flaw's",
+      "readout_selector": "before_answer",
+      "tracked": ["Griffith", "crack"],
+      "answer": "length",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true
+    },
+    {
+      "slug": "fatigue-striations-example",
+      "shape": "MULTIHOP",
+      "protocol": "lens_eval",
+      "domain": "mechanics",
+      "title": "Repeated loading to fractographic feature",
+      "description": "The mechanism is latent; the requested output is a different token.",
+      "text": "A polished steel shaft undergoes millions of sub-yield load reversals. Microscopy shows one small increment of advance per reversal. The characteristic parallel markings left on the failure surface are called",
+      "readout_selector": "before_answer",
+      "tracked": ["fatigue"],
+      "answer": "striations",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true
+    },
+    {
+      "slug": "ductile-counterfactual-example",
+      "shape": "PROBE_SWAP",
+      "protocol": "probe_swap",
+      "domain": "materials",
+      "category": "failure-response",
+      "title": "Failure-response counterfactual",
+      "description": "Grade the downstream answer, not target-direction rank movement.",
+      "text": "A metal can accommodate extensive dislocation glide before separation. Under a tensile test beyond yield it will exhibit substantial",
+      "readout_selector": "before_answer",
+      "tracked": ["ductile"],
+      "answer": "elongation",
+      "swap_from": "ductile",
+      "swap_to": "brittle",
+      "swap_answer": "fracture",
+      "must_be_absent_from_input": true,
+      "must_be_absent_from_output": true
+    },
+    {
+      "slug": "plasticity-focus-example",
+      "shape": "MODULATION",
+      "protocol": "directed_modulation",
+      "domain": "mechanics",
+      "title": "Focus condition",
+      "description": "Teacher-forced carrier with an instructed side concept.",
+      "user": "Write \"The lab notebook lay open on the bench.\" Concentrate on crack-tip plasticity while you write the sentence. Don't write anything else.",
+      "assistant_prefill": "The lab notebook lay open on the bench.",
+      "readout_selector": "assistant_response",
+      "tracked": ["crack", "plasticity", "yield"],
+      "must_be_absent_from_output": true,
+      "condition": "focus",
+      "control_group": "plasticity-notebook",
+      "phrasing_id": "concentrate-01",
+      "carrier_id": "notebook",
+      "target_family": "crack-tip-plasticity"
+    },
+    {
+      "slug": "plasticity-suppress-example",
+      "shape": "MODULATION",
+      "protocol": "directed_modulation",
+      "domain": "mechanics",
+      "title": "Suppress condition",
+      "description": "Matched carrier and targets with a suppress instruction.",
+      "user": "Write \"The lab notebook lay open on the bench.\" Try not to think about crack-tip plasticity while you write it. Don't write anything else.",
+      "assistant_prefill": "The lab notebook lay open on the bench.",
+      "readout_selector": "assistant_response",
+      "tracked": ["crack", "plasticity", "yield"],
+      "must_be_absent_from_output": true,
+      "condition": "suppress",
+      "control_group": "plasticity-notebook",
+      "phrasing_id": "suppress-01",
+      "carrier_id": "notebook",
+      "target_family": "crack-tip-plasticity"
+    },
+    {
+      "slug": "plasticity-control-example",
+      "shape": "MODULATION",
+      "protocol": "directed_modulation",
+      "domain": "mechanics",
+      "title": "Neutral control",
+      "description": "Matched carrier without a side-task instruction.",
+      "user": "Write \"The lab notebook lay open on the bench.\" Don't write anything else.",
+      "assistant_prefill": "The lab notebook lay open on the bench.",
+      "readout_selector": "assistant_response",
+      "tracked": ["crack", "plasticity", "yield"],
+      "must_be_absent_from_output": true,
+      "condition": "control",
+      "control_group": "plasticity-notebook",
+      "phrasing_id": "control-01",
+      "carrier_id": "notebook",
+      "target_family": "crack-tip-plasticity"
+    },
+    {
+      "slug": "failure-mode-report-example",
+      "shape": "REPORT_SWAP",
+      "protocol": "verbal_report",
+      "domain": "materials",
+      "title": "Dynamic verbal-report swap",
+      "description": "The clean greedy report is selected dynamically as the source.",
+      "user": "Think of a metal failure mode. Answer in one word.",
+      "readout_selector": "final_prompt_token",
+      "tracked": ["fatigue", "creep", "corrosion", "fracture", "buckling"],
+      "swap_to": "creep",
+      "candidates": ["fatigue", "creep", "corrosion", "fracture", "buckling"]
+    }
+  ]
+}

jlens_materials/prompts/protein.json

@@ -0,0 +1,148 @@
+{
+  "prompts": [
+    {
+      "slug": "protein-helix-multihop",
+      "shape": "MULTIHOP",
+      "domain": "protein",
+      "title": "Backbone H-bonds -> secondary structure (multi-hop)",
+      "description": "i to i+4 backbone hydrogen bonding. Track 'hydrogen'/'backbone' before 'helix'.",
+      "text": "In a folded protein, when the backbone carbonyl oxygen of one residue accepts a hydrogen bond from the amide of the residue four positions later, and this repeats down the chain, the local fold that results is the alpha",
+      "readout_at": -1,
+      "tracked": [
+        "hydrogen",
+        "bond",
+        "backbone",
+        "helix",
+        "spiral"
+      ],
+      "note": "Two-hop: i,i+4 H-bond pattern -> alpha helix. 'hydrogen'/'bond'/'backbone' are the unspoken intermediates before the answer 'helix'."
+    },
+    {
+      "slug": "protein-gfp-recognition",
+      "shape": "RECOGNITION",
+      "domain": "protein",
+      "title": "GFP sequence -> function (recognition)",
+      "description": "The paper's own example, ported: a GFP N-terminal fragment. Watch 'protein'/'fluor'/'green' appear a few residues in.",
+      "text": "What is this? MSKGEELFTGVVPILVELDGDVNGHKFSVSGEGEGDATYGKLTLKFICTTGKLPVPWPTL",
+      "readout_at": -1,
+      "tracked": [
+        "protein",
+        "fluor",
+        "green",
+        "GFP",
+        "sequence",
+        "amino"
+      ],
+      "note": "Direct port of the paper's Figure 3 protein example (GFP). Their lens reads 'protein' ~5 residues in, then 'fluor' and 'green'. Reproducing this on Gemma is the cleanest cross-model check that the fitted lens is behaving."
+    },
+    {
+      "slug": "protein-collagen-recognition",
+      "shape": "RECOGNITION",
+      "domain": "protein",
+      "title": "Gly-X-Y repeat -> collagen (recognition)",
+      "description": "A (Gly-Pro-Hyp)_n-like repeat. Does the lens surface 'collagen'/'triple'/'helix' from the motif alone?",
+      "text": "What is this? GPPGPKGDPGPQGPRGPAGPPGPAGPQGPKGDPGPQGPRGPAGPP",
+      "readout_at": -1,
+      "tracked": [
+        "collagen",
+        "triple",
+        "helix",
+        "glycine",
+        "proline",
+        "fibril"
+      ],
+      "note": "Gly-X-Y (often Gly-Pro-Hyp) is the collagen signature. Whether the lens jumps from generic 'protein' to specific 'collagen' probes how much structured biochemical knowledge sits in the workspace."
+    },
+    {
+      "slug": "protein-transmembrane-recognition",
+      "shape": "RECOGNITION",
+      "domain": "protein",
+      "title": "Hydrophobic stretch -> transmembrane (recognition)",
+      "description": "A ~20-residue hydrophobic run. Expect 'membrane'/'hydrophobic'/'transmembrane'/'helix'.",
+      "text": "What is this? KKPNGFILVGGVLLLGAAIIGGVMLFAYVVKKPN",
+      "readout_at": -1,
+      "tracked": [
+        "membrane",
+        "hydrophobic",
+        "transmembrane",
+        "helix",
+        "lipid"
+      ],
+      "note": "A flanked poly-(L/I/V/A/G/F) stretch is the transmembrane-helix signature. Recognising it requires reading residue *chemistry*, not just 'this is a sequence'."
+    },
+    {
+      "slug": "protein-amyloid-assoc",
+      "shape": "ASSOCIATION",
+      "domain": "protein",
+      "title": "Amyloid aggregation (association, unnamed)",
+      "description": "A vignette of misfolding and fibril formation. Does 'amyloid'/'aggregate'/'fibril' rise without being named?",
+      "text": "Under the microscope the soluble molecules had abandoned their folds. They had stacked edge to edge into long stiff ribbons that clumped into tangled plaques, each strand locked into its neighbour in a rigid pleated sheet.",
+      "readout_at": -1,
+      "tracked": [
+        "amyloid",
+        "fibril",
+        "aggregate",
+        "misfold",
+        "beta",
+        "plaque"
+      ],
+      "note": "Association eval. The passage describes cross-beta amyloid formation (misfolding -> stacked beta strands -> fibrils/plaques) without the diagnostic words; a workspace concept of 'amyloid' should be readable at the close."
+    },
+    {
+      "slug": "protein-fold-multihop",
+      "shape": "MULTIHOP",
+      "domain": "protein",
+      "title": "Hydrophobic collapse -> tertiary fold (multi-hop)",
+      "description": "Chain: sequence -> hydrophobic core -> native fold. Track 'hydrophobic'/'core'/'fold' in order.",
+      "text": "A denatured protein is returned to water. Its nonpolar side chains are driven out of contact with solvent and pack together in the interior, and this burial of hydrophobic residues is the main force that drives the chain back to its native",
+      "readout_at": -1,
+      "tracked": [
+        "hydrophobic",
+        "core",
+        "collapse",
+        "fold",
+        "structure",
+        "native"
+      ],
+      "note": "Anfinsen-style hydrophobic-collapse chain: the intermediates 'hydrophobic' and 'core' should precede the answer 'fold'/'structure' across layers."
+    },
+    {
+      "slug": "protein-modulation",
+      "shape": "MODULATION",
+      "domain": "protein",
+      "title": "Hold 'disulfide bond' while copying (modulation)",
+      "description": "Copy an unrelated sentence while concentrating on disulfide bonds. Does 'disulfide'/'cysteine'/'bridge' surface?",
+      "user": "Write \"The rain tapped against the window all afternoon.\" Concentrate on disulfide bonds while you write the sentence. Don't write anything else.",
+      "assistant_prefill": "The rain tapped against the window all afternoon.",
+      "readout_at": -1,
+      "tracked": [
+        "disulfide",
+        "cysteine",
+        "bridge",
+        "bond",
+        "sulfur",
+        "cross"
+      ],
+      "note": "Directed modulation. The carrier is about rain; success is 'disulfide'/'cysteine' appearing in the lens across it, showing the workspace holds an instructed biochemical concept off-output."
+    },
+    {
+      "slug": "protein-report-swap",
+      "shape": "REPORT_SWAP",
+      "domain": "protein",
+      "title": "Think of an amino acid -> swap (verbal report)",
+      "description": "'Think of an amino acid. Answer in one word.' Swap the chosen residue's lens vector for a sibling; check the report changes.",
+      "user": "Think of an amino acid. Answer in one word.",
+      "readout_at": -1,
+      "tracked": [
+        "glycine",
+        "alanine",
+        "leucine",
+        "serine",
+        "proline"
+      ],
+      "swap_from": "glycine",
+      "swap_to": "proline",
+      "note": "Verbal-report swap. Amino-acid names are the 'sport' category analogue; the swap tests whether the reported residue is set by the workspace coordinate at the report position."
+    }
+  ]
+}

jlens_materials/runs/comparison_Qwen25-05B-Instruct_vs_gpt2.json

@@ -0,0 +1,198 @@
+{
+  "tag": "Qwen25-05B-Instruct_vs_gpt2",
+  "models": [
+    "Qwen/Qwen2.5-0.5B-Instruct",
+    "gpt2"
+  ],
+  "provider": "anthropic",
+  "offline": true,
+  "hitrate": {
+    "shapes": [
+      "MULTIHOP",
+      "ASSOCIATION",
+      "RECOGNITION",
+      "MODULATION",
+      "REPORT_SWAP"
+    ],
+    "labels": [
+      "Qwen2.5-0.5B-Instruct",
+      "gpt2"
+    ],
+    "values": [
+      [
+        0.625,
+        0.2,
+        0.07142857142857142,
+        0.5555555555555556,
+        0.5
+      ],
+      [
+        0.4782608695652174,
+        0.3,
+        0.0,
+        NaN,
+        NaN
+      ]
+    ]
+  },
+  "swaps": [
+    {
+      "model": "Qwen2.5-0.5B-Instruct",
+      "slug": "fracture-report-swap",
+      "source": "fatigue",
+      "target": "creep",
+      "alpha": 1.0,
+      "source_rank_clean": 86,
+      "source_rank_swapped": 7933,
+      "target_rank_clean": 324,
+      "target_rank_swapped": 154
+    }
+  ],
+  "analysis": "[offline template] Comparison of Qwen2.5-0.5B-Instruct, gpt2. Highest overall top-5 surfacing rate: Qwen2.5-0.5B-Instruct. See the grouped-bar and concept-rank figures for the per-shape and per-concept breakdown.",
+  "per_prompt": [
+    {
+      "slug": "fracture-griffith",
+      "title": "Griffith criterion (multi-hop)",
+      "shape": "MULTIHOP",
+      "domain": "fracture",
+      "figure": "figures/compare/Qwen25-05B-Instruct_vs_gpt2/fracture-griffith__ranks.png",
+      "cross": "[offline template] Qwen2.5-0.5B-Instruct: 3 concept(s) in lens top-5 (energy, surface, crack); gpt2: 2 concept(s) in lens top-5 (energy, surface).",
+      "models": [
+        "Qwen2.5-0.5B-Instruct",
+        "gpt2"
+      ]
+    },
+    {
+      "slug": "fracture-fatigue",
+      "title": "Cyclic loading -> failure mode (multi-hop)",
+      "shape": "MULTIHOP",
+      "domain": "fracture",
+      "figure": "figures/compare/Qwen25-05B-Instruct_vs_gpt2/fracture-fatigue__ranks.png",
+      "cross": "[offline template] Qwen2.5-0.5B-Instruct: 3 concept(s) in lens top-5 (fatigue, crack, cyclic); gpt2: 1 concept(s) in lens top-5 (crack).",
+      "models": [
+        "Qwen2.5-0.5B-Instruct",
+        "gpt2"
+      ]
+    },
+    {
+      "slug": "fracture-toughness-chain",
+      "title": "Stress-intensity factor chain (multi-hop)",
+      "shape": "MULTIHOP",
+      "domain": "fracture",
+      "figure": "figures/compare/Qwen25-05B-Instruct_vs_gpt2/fracture-toughness-chain__ranks.png",
+      "cross": "[offline template] Qwen2.5-0.5B-Instruct: 4 concept(s) in lens top-5 (stress, intensity, toughness, propagate); gpt2: 3 concept(s) in lens top-5 (stress, intensity, toughness).",
+      "models": [
+        "Qwen2.5-0.5B-Instruct",
+        "gpt2"
+      ]
+    },
+    {
+      "slug": "fracture-brittle-assoc",
+      "title": "Brittle fracture (association, unnamed)",
+      "shape": "ASSOCIATION",
+      "domain": "fracture",
+      "figure": "figures/compare/Qwen25-05B-Instruct_vs_gpt2/fracture-brittle-assoc__ranks.png",
+      "cross": "[offline template] Qwen2.5-0.5B-Instruct: 1 concept(s) in lens top-5 (glass); gpt2: 2 concept(s) in lens top-5 (glass, shatter).",
+      "models": [
+        "Qwen2.5-0.5B-Instruct",
+        "gpt2"
+      ]
+    },
+    {
+      "slug": "fracture-ductile-assoc",
+      "title": "Ductile failure (association, unnamed)",
+      "shape": "ASSOCIATION",
+      "domain": "fracture",
+      "figure": "figures/compare/Qwen25-05B-Instruct_vs_gpt2/fracture-ductile-assoc__ranks.png",
+      "cross": "[offline template] Qwen2.5-0.5B-Instruct: 0 concept(s) in lens top-5; gpt2: 0 concept(s) in lens top-5.",
+      "models": [
+        "Qwen2.5-0.5B-Instruct",
+        "gpt2"
+      ]
+    },
+    {
+      "slug": "fracture-stress-recognition",
+      "title": "Stress state from a tensor (recognition)",
+      "shape": "RECOGNITION",
+      "domain": "fracture",
+      "figure": "figures/compare/Qwen25-05B-Instruct_vs_gpt2/fracture-stress-recognition__ranks.png",
+      "cross": "[offline template] Qwen2.5-0.5B-Instruct: 1 concept(s) in lens top-5 (mode); gpt2: 0 concept(s) in lens top-5.",
+      "models": [
+        "Qwen2.5-0.5B-Instruct",
+        "gpt2"
+      ]
+    },
+    {
+      "slug": "protein-helix-multihop",
+      "title": "Backbone H-bonds -> secondary structure (multi-hop)",
+      "shape": "MULTIHOP",
+      "domain": "protein",
+      "figure": "figures/compare/Qwen25-05B-Instruct_vs_gpt2/protein-helix-multihop__ranks.png",
+      "cross": "[offline template] Qwen2.5-0.5B-Instruct: 3 concept(s) in lens top-5 (hydrogen, bond, backbone); gpt2: 3 concept(s) in lens top-5 (hydrogen, bond, backbone).",
+      "models": [
+        "Qwen2.5-0.5B-Instruct",
+        "gpt2"
+      ]
+    },
+    {
+      "slug": "protein-gfp-recognition",
+      "title": "GFP sequence -> function (recognition)",
+      "shape": "RECOGNITION",
+      "domain": "protein",
+      "figure": "figures/compare/Qwen25-05B-Instruct_vs_gpt2/protein-gfp-recognition__ranks.png",
+      "cross": "[offline template] Qwen2.5-0.5B-Instruct: 0 concept(s) in lens top-5; gpt2: 0 concept(s) in lens top-5.",
+      "models": [
+        "Qwen2.5-0.5B-Instruct",
+        "gpt2"
+      ]
+    },
+    {
+      "slug": "protein-collagen-recognition",
+      "title": "Gly-X-Y repeat -> collagen (recognition)",
+      "shape": "RECOGNITION",
+      "domain": "protein",
+      "figure": "figures/compare/Qwen25-05B-Instruct_vs_gpt2/protein-collagen-recognition__ranks.png",
+      "cross": "[offline template] Qwen2.5-0.5B-Instruct: 0 concept(s) in lens top-5; gpt2: 0 concept(s) in lens top-5.",
+      "models": [
+        "Qwen2.5-0.5B-Instruct",
+        "gpt2"
+      ]
+    },
+    {
+      "slug": "protein-transmembrane-recognition",
+      "title": "Hydrophobic stretch -> transmembrane (recognition)",
+      "shape": "RECOGNITION",
+      "domain": "protein",
+      "figure": "figures/compare/Qwen25-05B-Instruct_vs_gpt2/protein-transmembrane-recognition__ranks.png",
+      "cross": "[offline template] Qwen2.5-0.5B-Instruct: 0 concept(s) in lens top-5; gpt2: 0 concept(s) in lens top-5.",
+      "models": [
+        "Qwen2.5-0.5B-Instruct",
+        "gpt2"
+      ]
+    },
+    {
+      "slug": "protein-amyloid-assoc",
+      "title": "Amyloid aggregation (association, unnamed)",
+      "shape": "ASSOCIATION",
+      "domain": "protein",
+      "figure": "figures/compare/Qwen25-05B-Instruct_vs_gpt2/protein-amyloid-assoc__ranks.png",
+      "cross": "[offline template] Qwen2.5-0.5B-Instruct: 1 concept(s) in lens top-5 (plaque); gpt2: 1 concept(s) in lens top-5 (plaque).",
+      "models": [
+        "Qwen2.5-0.5B-Instruct",
+        "gpt2"
+      ]
+    },
+    {
+      "slug": "protein-fold-multihop",
+      "title": "Hydrophobic collapse -> tertiary fold (multi-hop)",
+      "shape": "MULTIHOP",
+      "domain": "protein",
+      "figure": "figures/compare/Qwen25-05B-Instruct_vs_gpt2/protein-fold-multihop__ranks.png",
+      "cross": "[offline template] Qwen2.5-0.5B-Instruct: 2 concept(s) in lens top-5 (structure, native); gpt2: 2 concept(s) in lens top-5 (structure, native).",
+      "models": [
+        "Qwen2.5-0.5B-Instruct",
+        "gpt2"
+      ]
+    }
+  ]
+}

jlens_materials/runs/gemma4-e4b-it-heldout-v1-smoke-seed0.json

@@ -0,0 +1,3539 @@
+{
+  "format_version": 2,
+  "created_at": "2026-07-15T00:32:28.839707+00:00",
+  "model": "google/gemma-4-E4B-it",
+  "model_identity": {
+    "requested_id": "google/gemma-4-E4B-it",
+    "requested_revision": "a4c2d58be94dda072b918d9db64ee85c8ed34e3f",
+    "model_class": "Gemma4ForConditionalGeneration",
+    "config_name_or_path": "google/gemma-4-E4B-it",
+    "model_revision": "a4c2d58be94dda072b918d9db64ee85c8ed34e3f",
+    "tokenizer_class": "GemmaTokenizer",
+    "tokenizer_name_or_path": "google/gemma-4-E4B-it",
+    "vocab_size": 262144,
+    "n_layers": 42,
+    "d_model": 2560,
+    "dtype": "bfloat16"
+  },
+  "tag": "gemma4-e4b-it-heldout-v1-smoke-seed0",
+  "n_layers": 42,
+  "d_model": 2560,
+  "lens_n_prompts": 1000,
+  "lens_provenance": {
+    "corpus": {
+      "dataset": "Salesforce/wikitext",
+      "kind": "huggingface",
+      "min_chars": 600,
+      "records": 1000,
+      "requested_records": 1000,
+      "revision": "main",
+      "seed": 0,
+      "sha256": "a4b31b0673b21a932e33494c8471fb53577bd39b041f0799431c912876f5681e",
+      "split": "train",
+      "subset": "wikitext-103-raw-v1",
+      "text_field": "text",
+      "unique_records": 1000
+    },
+    "created_at": "2026-07-11T10:46:06.374107+00:00",
+    "dim_batch": 16,
+    "format_version": 1,
+    "hub": {
+      "path_in_repo": "gemma4-e4b-it/paper/seed0.pt",
+      "repo_id": "lamm-mit/gemma4-jacobian-lenses",
+      "repo_type": "model",
+      "revision": "main",
+      "uploaded_at": "2026-07-11T21:52:28.279960+00:00"
+    },
+    "model": {
+      "config_name_or_path": "google/gemma-4-E4B-it",
+      "d_model": 2560,
+      "dtype": "bfloat16",
+      "model_class": "Gemma4ForConditionalGeneration",
+      "model_revision": "a4c2d58be94dda072b918d9db64ee85c8ed34e3f",
+      "n_layers": 42,
+      "requested_id": "google/gemma-4-E4B-it",
+      "requested_revision": null,
+      "tokenizer_class": "GemmaTokenizer",
+      "tokenizer_name_or_path": "google/gemma-4-E4B-it",
+      "vocab_size": 262144
+    },
+    "paper_reference": "https://transformer-circuits.pub/2026/workspace/index.html",
+    "recipe": {
+      "description": "Paper-scale estimator: 1000 independent 128-token sequences, penultimate target layer, and 25 evenly spaced reported layers.",
+      "max_seq_len": 128,
+      "n_fit": 1000,
+      "name": "paper",
+      "report_layer_count": 25,
+      "skip_first": 16,
+      "strict_corpus": true,
+      "target_layer": -2
+    },
+    "source_layers": [
+      0,
+      2,
+      3,
+      5,
+      6,
+      8,
+      10,
+      11,
+      13,
+      15,
+      16,
+      18,
+      20,
+      21,
+      23,
+      24,
+      26,
+      28,
+      29,
+      31,
+      32,
+      34,
+      36,
+      37,
+      39
+    ],
+    "target_layer": 40,
+    "upstream_reference": "lamm-mit/jacobian-lens@581d398613e5602a5af361e1c34d3a92ea82ba8e"
+  },
+  "methodology": {
+    "recipe": {
+      "name": "demo",
+      "n_fit": 12,
+      "max_seq_len": 64,
+      "target_layer": null,
+      "report_layer_count": null,
+      "skip_first": 16,
+      "strict_corpus": false,
+      "description": "Fast exploratory fit. Results from this profile are not a quantitative replication of the paper."
+    },
+    "claims_level": "exploratory demo",
+    "paper_protocol_complete": false,
+    "strict_protocol": true,
+    "workspace_band": [
+      38.0,
+      92.0
+    ],
+    "workspace_band_source": "preregistered_cli_or_config",
+    "score_reduction": "fixed span x fixed band x synonym minimum",
+    "emergence_threshold": 5,
+    "emergence_sustain": 2,
+    "pass_at_k": [
+      1,
+      2,
+      5,
+      10,
+      20,
+      50,
+      100
+    ],
+    "minimum_items_per_shape": 1,
+    "evaluation_max_seq_len": 512,
+    "generation_max_new_tokens": 1,
+    "model_revision": "a4c2d58be94dda072b918d9db64ee85c8ed34e3f",
+    "dtype": "bfloat16",
+    "layer_stride": 1,
+    "prompt_source": "prompts/materials-heldout-v1-smoke.json",
+    "swaps_enabled": false,
+    "swap_alpha": 1.0
+  },
+  "shapes": [
+    "ASSOCIATION"
+  ],
+  "domains": [
+    "materials"
+  ],
+  "metrics": {
+    "by_shape": {
+      "ASSOCIATION": {
+        "n_items": 1,
+        "minimum_required": 1,
+        "sufficient_sample": true,
+        "ks": [
+          1,
+          2,
+          5,
+          10,
+          20,
+          50,
+          100
+        ],
+        "jacobian_lens": {
+          "pass_at_k": [
+            0.0,
+            0.0,
+            0.0,
+            0.0,
+            0.0,
+            0.0,
+            0.0
+          ],
+          "auc_log_k": 0.0,
+          "bootstrap_95": {
+            "1": {
+              "mean": 0.0,
+              "low": 0.0,
+              "high": 0.0
+            },
+            "2": {
+              "mean": 0.0,
+              "low": 0.0,
+              "high": 0.0
+            },
+            "5": {
+              "mean": 0.0,
+              "low": 0.0,
+              "high": 0.0
+            },
+            "10": {
+              "mean": 0.0,
+              "low": 0.0,
+              "high": 0.0
+            },
+            "20": {
+              "mean": 0.0,
+              "low": 0.0,
+              "high": 0.0
+            },
+            "50": {
+              "mean": 0.0,
+              "low": 0.0,
+              "high": 0.0
+            },
+            "100": {
+              "mean": 0.0,
+              "low": 0.0,
+              "high": 0.0
+            }
+          }
+        },
+        "logit_lens": {
+          "pass_at_k": [
+            0.0,
+            0.0,
+            0.0,
+            0.0,
+            0.0,
+            0.0,
+            0.0
+          ],
+          "auc_log_k": 0.0
+        }
+      }
+    },
+    "directed_modulation_controls": {},
+    "causal_swaps": {
+      "n_interventions": 0,
+      "n_graded": 0,
+      "minimum_required": 1,
+      "sufficient_sample": false,
+      "protocol_success_rate": null,
+      "n_counterfactual": 0,
+      "causal_success_rate": null
+    }
+  },
+  "errors": [],
+  "insufficient_sample_shapes": [],
+  "prompts": [
+    {
+      "slug": "heldout-v1-assoc-ductile-01",
+      "shape": "ASSOCIATION",
+      "protocol": "lens_eval",
+      "domain": "materials",
+      "title": "Energy-absorbing dimpled separation 1",
+      "description": "Held-out unnamed materials description; compare predetermined full-vocabulary ranks and target-free discovery under matched lenses.",
+      "note": "Frozen held-out item. Predetermined terms are used only for the rank endpoint; open-vocabulary generation and ranking cannot use them.",
+      "prompt_text": "A round aluminum bar stretched into a pronounced waist before separating. The mating faces contained hemispherical pits centered on inclusions and a slanted tearing rim.",
+      "prompt_tail": "luminum bar stretched into a pronounced waist before separating. The mating faces contained hemispherical pits centered on inclusions and a slanted tearing rim.",
+      "readout_selector": "final_prompt_token",
+      "readout_at": -1,
+      "score_positions": [
+        28
+      ],
+      "band": [
+        38.0,
+        92.0
+      ],
+      "band_source": "preregistered_cli_or_config",
+      "condition": null,
+      "control_group": null,
+      "phrasing_id": "heldout-ductile-01",
+      "carrier_id": null,
+      "target_family": "ductile",
+      "category": "ductile",
+      "answer": null,
+      "baseline": {
+        "required": false,
+        "correct": null,
+        "expected": []
+      },
+      "generated_completion": "end",
+      "generated_token_ids": [
+        643
+      ],
+      "protocol_violations": [],
+      "valid_for_metrics": true,
+      "excluded_reasons": [],
+      "discovered": false,
+      "open_vocabulary_config": {
+        "layer_readout_top": 32,
+        "surprising_top": 64,
+        "candidate_generation_uses_tracked_terms": false,
+        "matched_logit_open_vocabulary_stored": true
+      },
+      "tracked_dropped": [],
+      "tracked": [
+        {
+          "label": "ductile",
+          "surfaces": [
+            "ductile"
+          ],
+          "token_ids": [
+            185335
+          ]
+        },
+        {
+          "label": "nucleation",
+          "surfaces": [
+            "nucleation"
+          ],
+          "token_ids": [
+            87211
+          ]
+        },
+        {
+          "label": "coalescence",
+          "surfaces": [
+            "coalescence"
+          ],
+          "token_ids": [
+            216792
+          ]
+        },
+        {
+          "label": "void",
+          "surfaces": [
+            "void"
+          ],
+          "token_ids": [
+            2325,
+            4117,
+            61177,
+            70562
+          ]
+        }
+      ],
+      "emergence": [
+        {
+          "label": "ductile",
+          "best_rank": 533,
+          "best_depth": 39.0,
+          "onset_depth": null,
+          "reached_top1": false,
+          "best_pos": 28,
+          "logit_lens_best_rank": 25566
+        },
+        {
+          "label": "nucleation",
+          "best_rank": 1285,
+          "best_depth": 43.9,
+          "onset_depth": null,
+          "reached_top1": false,
+          "best_pos": 28,
+          "logit_lens_best_rank": 19938
+        },
+        {
+          "label": "coalescence",
+          "best_rank": 267,
+          "best_depth": 48.8,
+          "onset_depth": null,
+          "reached_top1": false,
+          "best_pos": 28,
+          "logit_lens_best_rank": 51516
+        },
+        {
+          "label": "void",
+          "best_rank": 147,
+          "best_depth": 51.2,
+          "onset_depth": null,
+          "reached_top1": false,
+          "best_pos": 28,
+          "logit_lens_best_rank": 5317
+        }
+      ],
+      "layer_readouts": [
+        {
+          "layer": 0,
+          "depth": 0.0,
+          "top_tokens": [
+            "alike",
+            "behalf",
+            "orry",
+            "accordingly",
+            "hereby",
+            "leness",
+            "ients",
+            "hence",
+            "ogether",
+            "oba",
+            "iken",
+            "ammed",
+            "isses",
+            "according",
+            "therefore",
+            "eles",
+            "ices",
+            "ilities",
+            "asons",
+            "anyak",
+            "cen",
+            "hence",
+            "eces"
+          ]
+        },
+        {
+          "layer": 2,
+          "depth": 4.9,
+          "top_tokens": [
+            "gonna",
+            "info",
+            "lle",
+            "wanna",
+            "lly",
+            "ie",
+            "ited",
+            "ided",
+            "thing",
+            "guys",
+            "guy",
+            "yy",
+            "yeah",
+            "wie",
+            "lli",
+            "arie",
+            "dia",
+            "eded",
+            "leness",
+            "witch",
+            "ndo",
+            "tics",
+            "nto",
+            "d",
+            "ion",
+            "ng",
+            "outh",
+            "wards",
+            "ii",
+            "basically"
+          ]
+        },
+        {
+          "layer": 3,
+          "depth": 7.3,
+          "top_tokens": [
+            "gonna",
+            "lle",
+            "thing",
+            "etc",
+            "info",
+            "basically",
+            "alike",
+            "lly",
+            "wards",
+            "etc",
+            "stuff",
+            "xd",
+            "ILD",
+            "afterwards",
+            "onus",
+            "d",
+            "ii",
+            "mente",
+            "persons",
+            "atre",
+            "dia",
+            "ie",
+            "plus",
+            "centre",
+            "guy",
+            "aska",
+            "aid",
+            "dz"
+          ]
+        },
+        {
+          "layer": 5,
+          "depth": 12.2,
+          "top_tokens": [
+            "thereafter",
+            "hiy",
+            "pursuant",
+            "ained",
+            "ortheast",
+            "orial",
+            "arters",
+            "vier",
+            "thereto",
+            "ued",
+            "oides",
+            "oid",
+            "ited",
+            "toward",
+            "aken",
+            "ented",
+            "edly",
+            "udos",
+            "thereon",
+            "eded",
+            "yses",
+            "oused",
+            "urized",
+            "veh",
+            "maupun"
+          ]
+        },
+        {
+          "layer": 6,
+          "depth": 14.6,
+          "top_tokens": [
+            "thereafter",
+            "pursuant",
+            "thereto",
+            "thence",
+            "toward",
+            "ued",
+            "resultant",
+            "yses",
+            "extant",
+            "eded",
+            "abutting",
+            "arters",
+            "thereof",
+            "Util",
+            "edly",
+            "sowie",
+            "sic",
+            "heretofore",
+            "insofar",
+            "hiy",
+            "ained",
+            "unto",
+            "CTED",
+            "ct",
+            "lement"
+          ]
+        },
+        {
+          "layer": 8,
+          "depth": 19.5,
+          "top_tokens": [
+            "thereafter",
+            "also",
+            "utilized",
+            "comprised",
+            "beneath",
+            "consisted",
+            "upon",
+            "toward",
+            "concerning",
+            "surfaced",
+            "filmed",
+            "afterward",
+            "Util",
+            "commenced",
+            "personnel",
+            "composed",
+            "thereto",
+            "zur",
+            "resulted",
+            "located",
+            "arters",
+            "subsequently",
+            "aided",
+            "underwent",
+            "characterized",
+            "occupied",
+            "utilizing"
+          ]
+        },
+        {
+          "layer": 10,
+          "depth": 24.4,
+          "top_tokens": [
+            "thereafter",
+            "adjoining",
+            "thence",
+            "toward",
+            "employed",
+            "aided",
+            "abutting",
+            "noted",
+            "commenced",
+            "apiece",
+            "later",
+            "thereto",
+            "ensuing",
+            "preparatory",
+            "subsequently",
+            "beneath",
+            "notably",
+            "notable",
+            "viz",
+            "punctuated",
+            "upon",
+            "ield",
+            "flanking",
+            "further",
+            "varied",
+            "preceding",
+            "employing",
+            "preceded",
+            "wrought",
+            "occasioned",
+            "edly"
+          ]
+        },
+        {
+          "layer": 11,
+          "depth": 26.8,
+          "top_tokens": [
+            "commenced",
+            "indistinct",
+            "thereto",
+            "obliquely",
+            "characterized",
+            "constituted",
+            "merely",
+            "coloration",
+            "concomitant",
+            "apiece",
+            "characteristic",
+            "alike",
+            "apparently",
+            "also",
+            "toward",
+            "considerable",
+            "furthermore",
+            "examined",
+            "preoccupation",
+            "concomit",
+            "composition",
+            "employed",
+            "concerning",
+            "cuticle",
+            "evidently",
+            "notwithstanding",
+            "adjoining",
+            "thence",
+            "suggestive",
+            "complexion",
+            "occasioned",
+            "noted"
+          ]
+        },
+        {
+          "layer": 13,
+          "depth": 31.7,
+          "top_tokens": [
+            "obliquely",
+            "abutting",
+            "abrasive",
+            "delineated",
+            "abut",
+            "commenced",
+            "abrasion",
+            "protruding",
+            "necessitating",
+            "mottled",
+            "idal",
+            "fiss",
+            "excised",
+            "coloration",
+            "discontinuity",
+            "indistinct",
+            "imping",
+            "characterized",
+            "anisotropic",
+            "azimuth",
+            "purview",
+            "flanks",
+            "delineation",
+            "protrusion",
+            "contour",
+            "oblique",
+            "terminus",
+            "concomitant",
+            "incipient",
+            "mouth",
+            "ridge",
+            "topography"
+          ]
+        },
+        {
+          "layer": 15,
+          "depth": 36.6,
+          "top_tokens": [
+            "protruding",
+            "protrusion",
+            "debris",
+            "delineated",
+            "curving",
+            "perpendicular",
+            "emplacement",
+            "suture",
+            "affixed",
+            "lacer",
+            "scarred",
+            "etching",
+            "abutting",
+            "etched",
+            "demarc",
+            "delineation",
+            "angled",
+            "abrasion",
+            "scarring",
+            "protrusions",
+            "curved",
+            "deline",
+            "shoreline",
+            "etch",
+            "starboard",
+            "subsurface",
+            "littoral",
+            "scar",
+            "markings",
+            "azimuth",
+            "flanking"
+          ]
+        },
+        {
+          "layer": 16,
+          "depth": 39.0,
+          "top_tokens": [
+            "debris",
+            "protruding",
+            "perpendicular",
+            "suture",
+            "lacer",
+            "scarred",
+            "scarring",
+            "obliquely",
+            "curving",
+            "protrusion",
+            "abutting",
+            "deform",
+            "distal",
+            "angled",
+            "ruptured",
+            "delineated",
+            "flanking",
+            "protrusions",
+            "resection",
+            "emplacement",
+            "delineation",
+            "damaged",
+            "rupture",
+            "elongated",
+            "azimuth",
+            "jagged",
+            "excised",
+            "adjacent",
+            "geometry",
+            "flank",
+            "anatomical"
+          ]
+        },
+        {
+          "layer": 18,
+          "depth": 43.9,
+          "top_tokens": [
+            "scarring",
+            "damaged",
+            "debris",
+            "ruptured",
+            "necrotic",
+            "necrosis",
+            "scarred",
+            "disintegration",
+            "deform",
+            "discoloration",
+            "lacer",
+            "damaged",
+            "salvaged",
+            "scar",
+            "rupture",
+            "diseased",
+            "cicat",
+            "wounding",
+            "Deformation",
+            "debris",
+            "damage",
+            "deformation",
+            "disintegr",
+            "protruding",
+            "indicating",
+            "degradation",
+            "Damage",
+            "discontinuity",
+            "excision",
+            "excised",
+            "wreckage",
+            "abutting"
+          ]
+        },
+        {
+          "layer": 20,
+          "depth": 48.8,
+          "top_tokens": [
+            "necrotic",
+            "necrosis",
+            "debris",
+            "inconclusive",
+            "disintegration",
+            "scarring",
+            "artifact",
+            "artifacts",
+            "forensic",
+            "anatomical",
+            "discontinuity",
+            "diseased",
+            "cicat",
+            "resection",
+            "disintegr",
+            "discoloration",
+            "excised",
+            "Deformation",
+            "ruptured",
+            "deformation",
+            "decomposition",
+            "degradation",
+            "deform",
+            "deteriorated",
+            "protrusions",
+            "fissures",
+            "deformity",
+            "decay",
+            "indicating",
+            "damaged",
+            "betray"
+          ]
+        },
+        {
+          "layer": 21,
+          "depth": 51.2,
+          "top_tokens": [
+            "necrosis",
+            "necrotic",
+            "diseased",
+            "specimen",
+            "fail",
+            "holotype",
+            "decay",
+            "damaged",
+            "degradation",
+            "artifact",
+            "artifacts",
+            "pathology",
+            "diagnose",
+            "characterizing",
+            "deteriorated",
+            "resection",
+            "inconclusive",
+            "scarring",
+            "irreparable",
+            "deformity",
+            "failure",
+            "recovered",
+            "debris",
+            "deterioration",
+            "anomalous",
+            "corrupted",
+            "salvage",
+            "haunt"
+          ]
+        },
+        {
+          "layer": 23,
+          "depth": 56.1,
+          "top_tokens": [
+            "furthermore",
+            "Ultimately",
+            "showcasing",
+            "culminating",
+            "whereupon",
+            "Which",
+            "ultimately",
+            "whence",
+            "Ultimately",
+            "characterized",
+            "demonstrating",
+            "which",
+            "featuring",
+            "tasked",
+            "Yet",
+            "Eventually",
+            "Featuring",
+            "simultaneously",
+            "comprised",
+            "punctuated",
+            "Notably",
+            "culminated",
+            "Conversely",
+            "necessitating",
+            "wherein",
+            "Eventually",
+            "whereas",
+            "meanwhile"
+          ]
+        },
+        {
+          "layer": 24,
+          "depth": 58.5,
+          "top_tokens": [
+            "furthermore",
+            "Furthermore",
+            "Ultimately",
+            "Also",
+            "Yet",
+            "Meanwhile",
+            "Each",
+            "Their",
+            "whereas",
+            "Its",
+            "Yet",
+            "Furthermore",
+            "Eventually",
+            "Conversely",
+            "Ultimately",
+            "Eventually",
+            "Featuring",
+            "Conversely",
+            "Simultaneously",
+            "Their",
+            "Likewise",
+            "Its",
+            "Also",
+            "Finally",
+            "Again",
+            "Consequently",
+            "Similarly",
+            "meanwhile",
+            "Afterwards",
+            "Likewise",
+            "Each"
+          ]
+        },
+        {
+          "layer": 26,
+          "depth": 63.4,
+          "top_tokens": [
+            "whereas",
+            "Furthermore",
+            "Similarly",
+            "Conversely",
+            "furthermore",
+            "Also",
+            "Each",
+            "Meanwhile",
+            "Similarly",
+            "Simultaneously",
+            "Then",
+            "Finally",
+            "meanwhile",
+            "Likewise",
+            "Afterwards",
+            "Conversely",
+            "likewise",
+            "Secondly",
+            "Ultimately",
+            "conversely",
+            "Equally",
+            "Yet",
+            "Furthermore",
+            "Alongside",
+            "Each",
+            "Its",
+            "Also",
+            "Again",
+            "Which",
+            "simultaneously",
+            "Otherwise",
+            "Consequently"
+          ]
+        },
+        {
+          "layer": 28,
+          "depth": 68.3,
+          "top_tokens": [
+            "meanwhile",
+            "furthermore",
+            "whereas",
+            "Finally",
+            "Meanwhile",
+            "Secondly",
+            "secondly",
+            "Secondly",
+            "conversely",
+            "Furthermore",
+            "Its",
+            "tandis",
+            "Simultaneously",
+            "Equally",
+            "Alongside",
+            "Whether",
+            "Finally",
+            "Similarly",
+            "Beneath",
+            "Their",
+            "underneath",
+            "theirs",
+            "respectively",
+            "finally",
+            "Which",
+            "mselves",
+            "During",
+            "Similarly",
+            "Either",
+            "Neither",
+            "Within",
+            "Lastly"
+          ]
+        },
+        {
+          "layer": 29,
+          "depth": 70.7,
+          "top_tokens": [
+            "meanwhile",
+            "Meanwhile",
+            "Yet",
+            "mselves",
+            "whereas",
+            "Yours",
+            "Yet",
+            "accordingly",
+            "Its",
+            "furthermore",
+            "conversely",
+            "Meanwhile",
+            "Accordingly",
+            "Conversely",
+            "Lately",
+            "Nevertheless",
+            "Together",
+            "Accordingly",
+            "Conversely",
+            "wobei",
+            "Today",
+            "Recently",
+            "tandis",
+            "Consequently",
+            "Referred",
+            "anymore",
+            "lately",
+            "Along",
+            "Yesterday"
+          ]
+        },
+        {
+          "layer": 31,
+          "depth": 75.6,
+          "top_tokens": [
+            "Meanwhile",
+            "meanwhile",
+            "furthermore",
+            "mselves",
+            "Consequently",
+            "Furthermore",
+            "Its",
+            "Conversely",
+            "whereas",
+            "Each",
+            "Along",
+            "conversely",
+            "They",
+            "Moreover",
+            "Within",
+            "Nevertheless",
+            "Together",
+            "Its",
+            "Meanwhile",
+            "Recently",
+            "Consequently",
+            "Yours",
+            "Today",
+            "Moreover",
+            "Along",
+            "Secondly",
+            "Conversely",
+            "Beneath",
+            "Regardless",
+            "Within",
+            "Their"
+          ]
+        },
+        {
+          "layer": 32,
+          "depth": 78.0,
+          "top_tokens": [
+            "Meanwhile",
+            "furthermore",
+            "meanwhile",
+            "Consequently",
+            "Furthermore",
+            "Each",
+            "They",
+            "Yet",
+            "Recently",
+            "Regardless",
+            "Conversely",
+            "Together",
+            "Beneath",
+            "Secondly",
+            "Its",
+            "Welcome",
+            "Therefore",
+            "Its",
+            "Hence",
+            "conversely",
+            "Moreover",
+            "Within",
+            "Below",
+            "Along",
+            "nonetheless",
+            "Today",
+            "Nevertheless",
+            "Lately",
+            "Additionally",
+            "Meanwhile",
+            "Nonetheless"
+          ]
+        },
+        {
+          "layer": 34,
+          "depth": 82.9,
+          "top_tokens": [
+            "Secondly",
+            "secondly",
+            "Furthermore",
+            "furthermore",
+            "Secondly",
+            "Welcome",
+            "Consequently",
+            "Needless",
+            "Except",
+            "Needless",
+            "Each",
+            "conversely",
+            "Conversely",
+            "Additionally",
+            "Hence",
+            "Alternatively",
+            "Likewise",
+            "jsonData",
+            "Ships",
+            "Meanwhile",
+            "unim",
+            "Terror",
+            "Notwithstanding",
+            "Another",
+            "Though",
+            "Unless",
+            "txt",
+            "Yet",
+            "Tissue"
+          ]
+        },
+        {
+          "layer": 36,
+          "depth": 87.8,
+          "top_tokens": [
+            "secondly",
+            "Secondly",
+            "plant",
+            "Secondly",
+            "ield",
+            "Meanwhile",
+            "indd",
+            "opy",
+            "Tis",
+            "acks",
+            "bred",
+            "balls",
+            "pots",
+            "conversely",
+            "Aren",
+            "ular",
+            "jpeg",
+            "endswith",
+            "THAN",
+            "endregion",
+            "quist"
+          ]
+        },
+        {
+          "layer": 37,
+          "depth": 90.2,
+          "top_tokens": [
+            "UMENT",
+            "secondly",
+            "opy",
+            "ular",
+            "plant",
+            "ument",
+            "ately",
+            "THAN",
+            "Secondly",
+            "ahme",
+            "ULAR",
+            "endregion",
+            "quist",
+            "ledes",
+            "heave",
+            "omatic",
+            "emat",
+            "ufer",
+            "nitrate",
+            "ufs"
+          ]
+        },
+        {
+          "layer": 39,
+          "depth": 95.1,
+          "top_tokens": [
+            "emat",
+            "secondly",
+            "igate",
+            "Secondly",
+            "UMENT",
+            "imiter",
+            "declare",
+            "volve",
+            "etera",
+            "ittle",
+            "ingly",
+            "ayet",
+            "heave",
+            "paginate",
+            "heading",
+            "ipation",
+            "ogloss"
+          ]
+        },
+        {
+          "layer": 41,
+          "depth": 100.0,
+          "top_tokens": [
+            "meet",
+            "The",
+            "he",
+            "sun",
+            "second",
+            "end",
+            "past",
+            "copy",
+            "pair",
+            "emat",
+            "length",
+            "evenly",
+            "tl",
+            "mat",
+            "Secondly",
+            "le",
+            "The",
+            "line",
+            "black",
+            "kill",
+            "union",
+            "meat",
+            "more",
+            "around",
+            "april",
+            "em",
+            "hat",
+            "rest",
+            "parts",
+            "jpg",
+            "fit",
+            "top"
+          ]
+        }
+      ],
+      "concept_trajectories": {
+        "jacobian_lens": [
+          {
+            "label": "ductile",
+            "layers": [
+              0,
+              2,
+              3,
+              5,
+              6,
+              8,
+              10,
+              11,
+              13,
+              15,
+              16,
+              18,
+              20,
+              21,
+              23,
+              24,
+              26,
+              28,
+              29,
+              31,
+              32,
+              34,
+              36,
+              37,
+              39,
+              41
+            ],
+            "depths": [
+              0.0,
+              4.9,
+              7.3,
+              12.2,
+              14.6,
+              19.5,
+              24.4,
+              26.8,
+              31.7,
+              36.6,
+              39.0,
+              43.9,
+              48.8,
+              51.2,
+              56.1,
+              58.5,
+              63.4,
+              68.3,
+              70.7,
+              75.6,
+              78.0,
+              82.9,
+              87.8,
+              90.2,
+              95.1,
+              100.0
+            ],
+            "ranks": [
+              256316,
+              171301,
+              139408,
+              84849,
+              99229,
+              82148,
+              10620,
+              8521,
+              1959,
+              298,
+              534,
+              4682,
+              9473,
+              14323,
+              158567,
+              212285,
+              162096,
+              160490,
+              133216,
+              139270,
+              111642,
+              155637,
+              93445,
+              177539,
+              253726,
+              244858
+            ]
+          },
+          {
+            "label": "nucleation",
+            "layers": [
+              0,
+              2,
+              3,
+              5,
+              6,
+              8,
+              10,
+              11,
+              13,
+              15,
+              16,
+              18,
+              20,
+              21,
+              23,
+              24,
+              26,
+              28,
+              29,
+              31,
+              32,
+              34,
+              36,
+              37,
+              39,
+              41
+            ],
+            "depths": [
+              0.0,
+              4.9,
+              7.3,
+              12.2,
+              14.6,
+              19.5,
+              24.4,
+              26.8,
+              31.7,
+              36.6,
+              39.0,
+              43.9,
+              48.8,
+              51.2,
+              56.1,
+              58.5,
+              63.4,
+              68.3,
+              70.7,
+              75.6,
+              78.0,
+              82.9,
+              87.8,
+              90.2,
+              95.1,
+              100.0
+            ],
+            "ranks": [
+              224278,
+              84474,
+              137534,
+              178111,
+              164212,
+              83026,
+              92351,
+              35917,
+              7030,
+              990,
+              1838,
+              1286,
+              3302,
+              3901,
+              120243,
+              85137,
+              60456,
+              61023,
+              22839,
+              26759,
+              24993,
+              26973,
+              71489,
+              76639,
+              140111,
+              115093
+            ]
+          },
+          {
+            "label": "coalescence",
+            "layers": [
+              0,
+              2,
+              3,
+              5,
+              6,
+              8,
+              10,
+              11,
+              13,
+              15,
+              16,
+              18,
+              20,
+              21,
+              23,
+              24,
+              26,
+              28,
+              29,
+              31,
+              32,
+              34,
+              36,
+              37,
+              39,
+              41
+            ],
+            "depths": [
+              0.0,
+              4.9,
+              7.3,
+              12.2,
+              14.6,
+              19.5,
+              24.4,
+              26.8,
+              31.7,
+              36.6,
+              39.0,
+              43.9,
+              48.8,
+              51.2,
+              56.1,
+              58.5,
+              63.4,
+              68.3,
+              70.7,
+              75.6,
+              78.0,
+              82.9,
+              87.8,
+              90.2,
+              95.1,
+              100.0
+            ],
+            "ranks": [
+              239934,
+              234062,
+              219132,
+              180225,
+              109383,
+              78187,
+              14082,
+              4937,
+              534,
+              815,
+              419,
+              642,
+              268,
+              569,
+              22513,
+              29390,
+              17972,
+              59447,
+              26433,
+              39967,
+              61572,
+              92224,
+              137098,
+              78300,
+              87054,
+              112142
+            ]
+          },
+          {
+            "label": "void",
+            "layers": [
+              0,
+              2,
+              3,
+              5,
+              6,
+              8,
+              10,
+              11,
+              13,
+              15,
+              16,
+              18,
+              20,
+              21,
+              23,
+              24,
+              26,
+              28,
+              29,
+              31,
+              32,
+              34,
+              36,
+              37,
+              39,
+              41
+            ],
+            "depths": [
+              0.0,
+              4.9,
+              7.3,
+              12.2,
+              14.6,
+              19.5,
+              24.4,
+              26.8,
+              31.7,
+              36.6,
+              39.0,
+              43.9,
+              48.8,
+              51.2,
+              56.1,
+              58.5,
+              63.4,
+              68.3,
+              70.7,
+              75.6,
+              78.0,
+              82.9,
+              87.8,
+              90.2,
+              95.1,
+              100.0
+            ],
+            "ranks": [
+              26025,
+              8302,
+              19072,
+              10166,
+              12960,
+              3573,
+              5758,
+              3384,
+              1286,
+              2331,
+              2190,
+              592,
+              340,
+              148,
+              23517,
+              42384,
+              74521,
+              25264,
+              49320,
+              32484,
+              25978,
+              44286,
+              76196,
+              41870,
+              14516,
+              406
+            ]
+          }
+        ],
+        "logit_lens": [
+          {
+            "label": "ductile",
+            "layers": [
+              16,
+              18,
+              20,
+              21,
+              23,
+              24,
+              26,
+              28,
+              29,
+              31,
+              32,
+              34,
+              36,
+              37
+            ],
+            "depths": [
+              39.0,
+              43.9,
+              48.8,
+              51.2,
+              56.1,
+              58.5,
+              63.4,
+              68.3,
+              70.7,
+              75.6,
+              78.0,
+              82.9,
+              87.8,
+              90.2
+            ],
+            "ranks": [
+              25567,
+              214305,
+              184927,
+              98554,
+              60740,
+              83752,
+              145639,
+              111954,
+              78348,
+              133736,
+              110212,
+              149809,
+              145253,
+              196788
+            ]
+          },
+          {
+            "label": "nucleation",
+            "layers": [
+              16,
+              18,
+              20,
+              21,
+              23,
+              24,
+              26,
+              28,
+              29,
+              31,
+              32,
+              34,
+              36,
+              37
+            ],
+            "depths": [
+              39.0,
+              43.9,
+              48.8,
+              51.2,
+              56.1,
+              58.5,
+              63.4,
+              68.3,
+              70.7,
+              75.6,
+              78.0,
+              82.9,
+              87.8,
+              90.2
+            ],
+            "ranks": [
+              206682,
+              243741,
+              232732,
+              220665,
+              179243,
+              205910,
+              201209,
+              59957,
+              19939,
+              55563,
+              101388,
+              30097,
+              71361,
+              69633
+            ]
+          },
+          {
+            "label": "coalescence",
+            "layers": [
+              16,
+              18,
+              20,
+              21,
+              23,
+              24,
+              26,
+              28,
+              29,
+              31,
+              32,
+              34,
+              36,
+              37
+            ],
+            "depths": [
+              39.0,
+              43.9,
+              48.8,
+              51.2,
+              56.1,
+              58.5,
+              63.4,
+              68.3,
+              70.7,
+              75.6,
+              78.0,
+              82.9,
+              87.8,
+              90.2
+            ],
+            "ranks": [
+              159088,
+              255710,
+              214045,
+              189202,
+              139079,
+              196479,
+              105957,
+              51517,
+              69565,
+              207110,
+              170738,
+              113673,
+              212919,
+              136476
+            ]
+          },
+          {
+            "label": "void",
+            "layers": [
+              16,
+              18,
+              20,
+              21,
+              23,
+              24,
+              26,
+              28,
+              29,
+              31,
+              32,
+              34,
+              36,
+              37
+            ],
+            "depths": [
+              39.0,
+              43.9,
+              48.8,
+              51.2,
+              56.1,
+              58.5,
+              63.4,
+              68.3,
+              70.7,
+              75.6,
+              78.0,
+              82.9,
+              87.8,
+              90.2
+            ],
+            "ranks": [
+              64936,
+              232577,
+              181473,
+              46102,
+              150609,
+              14076,
+              27396,
+              130475,
+              63049,
+              45443,
+              6367,
+              27682,
+              14965,
+              5318
+            ]
+          }
+        ],
+        "rank_indexing": "one-indexed; rank 1 is the highest-readout token",
+        "position_reduction": "minimum over preregistered score positions and declared single-token synonyms"
+      },
+      "surprising": [
+        {
+          "concept": "which",
+          "score": 11.0,
+          "best_rank": 1,
+          "best_depth": 56.1,
+          "near_token": "and",
+          "output_checked": true
+        },
+        {
+          "concept": "because",
+          "score": 10.0,
+          "best_rank": 1,
+          "best_depth": 58.5,
+          "near_token": "round",
+          "output_checked": true
+        },
+        {
+          "concept": "each",
+          "score": 9.0,
+          "best_rank": 1,
+          "best_depth": 63.4,
+          "near_token": "centered",
+          "output_checked": true
+        },
+        {
+          "concept": "cylindrical",
+          "score": 8.0,
+          "best_rank": 1,
+          "best_depth": 51.2,
+          "near_token": "bar",
+          "output_checked": true
+        },
+        {
+          "concept": "was",
+          "score": 8.0,
+          "best_rank": 1,
+          "best_depth": 58.5,
+          "near_token": "bar",
+          "output_checked": true
+        },
+        {
+          "concept": "became",
+          "score": 8.0,
+          "best_rank": 1,
+          "best_depth": 58.5,
+          "near_token": "stretched",
+          "output_checked": true
+        },
+        {
+          "concept": "remained",
+          "score": 8.0,
+          "best_rank": 1,
+          "best_depth": 63.4,
+          "near_token": "faces",
+          "output_checked": true
+        },
+        {
+          "concept": "within",
+          "score": 8.0,
+          "best_rank": 1,
+          "best_depth": 70.7,
+          "near_token": "centered",
+          "output_checked": true
+        },
+        {
+          "concept": "lengthwise",
+          "score": 7.0,
+          "best_rank": 1,
+          "best_depth": 43.9,
+          "near_token": "into",
+          "output_checked": true
+        },
+        {
+          "concept": "halves",
+          "score": 7.0,
+          "best_rank": 1,
+          "best_depth": 68.3,
+          "near_token": "The",
+          "output_checked": true
+        },
+        {
+          "concept": "several",
+          "score": 6.0,
+          "best_rank": 1,
+          "best_depth": 56.1,
+          "near_token": "before",
+          "output_checked": true
+        },
+        {
+          "concept": "pattern",
+          "score": 6.0,
+          "best_rank": 1,
+          "best_depth": 70.7,
+          "near_token": "tearing",
+          "output_checked": true
+        },
+        {
+          "concept": "where",
+          "score": 5.0,
+          "best_rank": 1,
+          "best_depth": 75.6,
+          "near_token": "pits",
+          "output_checked": true
+        },
+        {
+          "concept": "shape",
+          "score": 5.0,
+          "best_rank": 1,
+          "best_depth": 82.9,
+          "near_token": "into",
+          "output_checked": true
+        },
+        {
+          "concept": "cavity",
+          "score": 4.0,
+          "best_rank": 1,
+          "best_depth": 39.0,
+          "near_token": "pits",
+          "output_checked": true
+        },
+        {
+          "concept": "intricate",
+          "score": 4.0,
+          "best_rank": 1,
+          "best_depth": 70.7,
+          "near_token": "a",
+          "output_checked": true
+        },
+        {
+          "concept": "protrusions",
+          "score": 4.0,
+          "best_rank": 1,
+          "best_depth": 75.6,
+          "near_token": "erical",
+          "output_checked": true
+        },
+        {
+          "concept": "circumference",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 39.0,
+          "near_token": "round",
+          "output_checked": true
+        },
+        {
+          "concept": "diameter",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 39.0,
+          "near_token": "bar",
+          "output_checked": true
+        },
+        {
+          "concept": "tensile",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 39.0,
+          "near_token": "stretched",
+          "output_checked": true
+        },
+        {
+          "concept": "metallurgical",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 39.0,
+          "near_token": "a",
+          "output_checked": true
+        },
+        {
+          "concept": "cinematic",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 39.0,
+          "near_token": ".",
+          "output_checked": true
+        },
+        {
+          "concept": "ripping",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 39.0,
+          "near_token": "tearing",
+          "output_checked": true
+        },
+        {
+          "concept": "flanked",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 43.9,
+          "near_token": "before",
+          "output_checked": true
+        },
+        {
+          "concept": "abutting",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 43.9,
+          "near_token": "faces",
+          "output_checked": true
+        },
+        {
+          "concept": "example",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 63.4,
+          "near_token": "round",
+          "output_checked": true
+        },
+        {
+          "concept": "meanwhile",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 68.3,
+          "near_token": ".",
+          "output_checked": true
+        },
+        {
+          "concept": "sections",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 68.3,
+          "near_token": "mating",
+          "output_checked": true
+        },
+        {
+          "concept": "lacked",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 70.7,
+          "near_token": "and",
+          "output_checked": true
+        },
+        {
+          "concept": "depressions",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 78.0,
+          "near_token": "erical",
+          "output_checked": true
+        },
+        {
+          "concept": "table",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 82.9,
+          "near_token": "round",
+          "output_checked": true
+        },
+        {
+          "concept": "shaped",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 82.9,
+          "near_token": "bar",
+          "output_checked": true
+        },
+        {
+          "concept": "tapering",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 82.9,
+          "near_token": "before",
+          "output_checked": true
+        },
+        {
+          "concept": "ends",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 82.9,
+          "near_token": "The",
+          "output_checked": true
+        },
+        {
+          "concept": "groove",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 82.9,
+          "near_token": "slanted",
+          "output_checked": true
+        },
+        {
+          "concept": "whereas",
+          "score": 3.0,
+          "best_rank": 3,
+          "best_depth": 63.4,
+          "near_token": ".",
+          "output_checked": true
+        },
+        {
+          "concept": "visually",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 39.0,
+          "near_token": "The",
+          "output_checked": true
+        },
+        {
+          "concept": "requisite",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 39.0,
+          "near_token": "mating",
+          "output_checked": true
+        },
+        {
+          "concept": "perpendicular",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 39.0,
+          "near_token": "on",
+          "output_checked": true
+        },
+        {
+          "concept": "spouses",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 48.8,
+          "near_token": "mating",
+          "output_checked": true
+        },
+        {
+          "concept": "berupa",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 48.8,
+          "near_token": "contained",
+          "output_checked": true
+        },
+        {
+          "concept": "depicting",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 51.2,
+          "near_token": ".",
+          "output_checked": true
+        },
+        {
+          "concept": "quarters",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 56.1,
+          "near_token": "erical",
+          "output_checked": true
+        },
+        {
+          "concept": "arms",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 58.5,
+          "near_token": "into",
+          "output_checked": true
+        },
+        {
+          "concept": "portion",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 58.5,
+          "near_token": "mating",
+          "output_checked": true
+        },
+        {
+          "concept": "becoming",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 63.4,
+          "near_token": "before",
+          "output_checked": true
+        },
+        {
+          "concept": "then",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 63.4,
+          "near_token": "The",
+          "output_checked": true
+        },
+        {
+          "concept": "counterpart",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 63.4,
+          "near_token": "mating",
+          "output_checked": true
+        },
+        {
+          "concept": "representations",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 63.4,
+          "near_token": "hemisph",
+          "output_checked": true
+        },
+        {
+          "concept": "rectangular",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 68.3,
+          "near_token": "a",
+          "output_checked": true
+        },
+        {
+          "concept": "vertical",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 68.3,
+          "near_token": "pronounced",
+          "output_checked": true
+        },
+        {
+          "concept": "larger",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 68.3,
+          "near_token": "a",
+          "output_checked": true
+        },
+        {
+          "concept": "ImageUrl",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 75.6,
+          "near_token": "<bos>",
+          "output_checked": true
+        },
+        {
+          "concept": "pizza",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 75.6,
+          "near_token": "round",
+          "output_checked": true
+        },
+        {
+          "concept": "used",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 75.6,
+          "near_token": "bar",
+          "output_checked": true
+        },
+        {
+          "concept": "being",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 75.6,
+          "near_token": "before",
+          "output_checked": true
+        },
+        {
+          "concept": "Meanwhile",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 75.6,
+          "near_token": ".",
+          "output_checked": true
+        },
+        {
+          "concept": "cookware",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 78.0,
+          "near_token": "aluminum",
+          "output_checked": true
+        },
+        {
+          "concept": "smooth",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 78.0,
+          "near_token": "a",
+          "output_checked": true
+        },
+        {
+          "concept": "grooves",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 82.9,
+          "near_token": "erical",
+          "output_checked": true
+        },
+        {
+          "concept": "tray",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 87.8,
+          "near_token": "aluminum",
+          "output_checked": true
+        },
+        {
+          "concept": "designed",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 87.8,
+          "near_token": "pits",
+          "output_checked": true
+        },
+        {
+          "concept": "extending",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 87.8,
+          "near_token": "rim",
+          "output_checked": true
+        },
+        {
+          "concept": "eject",
+          "score": 2.0,
+          "best_rank": 2,
+          "best_depth": 48.8,
+          "near_token": "rim",
+          "output_checked": true
+        }
+      ],
+      "surprising_label": "exploratory surfaced candidates",
+      "logit_layer_readouts": [
+        {
+          "layer": 0,
+          "depth": 0.0,
+          "top_tokens": [
+            "mull",
+            "Hail",
+            "nada",
+            "ervan",
+            "wand",
+            "quantos",
+            "undos",
+            "ynitrite",
+            "addRow",
+            "NORE",
+            "Kwa",
+            "maneiras",
+            "punkte",
+            "ties"
+          ]
+        },
+        {
+          "layer": 2,
+          "depth": 4.9,
+          "top_tokens": [
+            "ago",
+            "channel",
+            "rid",
+            "tons",
+            "opat",
+            "stir",
+            "dim",
+            "grac",
+            "Blanch",
+            "Mannes",
+            "lev",
+            "Sob",
+            "reu",
+            "pres",
+            "twe",
+            "trem",
+            "mamy",
+            "zorunda"
+          ]
+        },
+        {
+          "layer": 3,
+          "depth": 7.3,
+          "top_tokens": [
+            "Clon",
+            "dim",
+            "plo",
+            "reu",
+            "champ",
+            "Blanch",
+            "mamy",
+            "hus",
+            "Salv",
+            "ums",
+            "hush"
+          ]
+        },
+        {
+          "layer": 5,
+          "depth": 12.2,
+          "top_tokens": [
+            "Sod",
+            "tfrac",
+            "champ",
+            "Hus",
+            "fam",
+            "essional",
+            "ilg",
+            "Langer",
+            "olyte",
+            "byt",
+            "alk",
+            "tene",
+            "Pres",
+            "reu"
+          ]
+        },
+        {
+          "layer": 6,
+          "depth": 14.6,
+          "top_tokens": [
+            "ago",
+            "tene",
+            "Hus",
+            "Sod",
+            "ilj",
+            "olyte",
+            "Slave",
+            "byt",
+            "essional",
+            "tfrac",
+            "Reis",
+            "Langer",
+            "fam",
+            "Pres",
+            "lev",
+            "wedge"
+          ]
+        },
+        {
+          "layer": 8,
+          "depth": 19.5,
+          "top_tokens": [
+            "ilers",
+            "rem",
+            "icot",
+            "wyst",
+            "Vo",
+            "WOR",
+            "oben",
+            "rekken",
+            "setia",
+            "kuasa",
+            "hst",
+            "hely",
+            "IVI",
+            "susp",
+            "yn",
+            "ae",
+            "falls",
+            "Vla",
+            "atus",
+            "illing",
+            "sn",
+            "vy",
+            "tit",
+            "VERE"
+          ]
+        },
+        {
+          "layer": 10,
+          "depth": 24.4,
+          "top_tokens": [
+            "side",
+            "trading",
+            "tool",
+            "tools",
+            "HI",
+            "bro",
+            "das",
+            "sol",
+            "quire",
+            "Rim",
+            "truncate",
+            "contr",
+            "Me",
+            "Green",
+            "dal",
+            "formal",
+            "IVING",
+            "Wakefield",
+            "MI",
+            "serial",
+            "crow",
+            "Ged",
+            "hybrid"
+          ]
+        },
+        {
+          "layer": 11,
+          "depth": 26.8,
+          "top_tokens": [
+            "tools",
+            "side",
+            "tool",
+            "HI",
+            "trading",
+            "truncate",
+            "tail",
+            "yal",
+            "lives",
+            "select",
+            "mei",
+            "Zealand",
+            "Green",
+            "tit",
+            "turbo",
+            "demos",
+            "serial",
+            "dad",
+            "valued",
+            "sol"
+          ]
+        },
+        {
+          "layer": 13,
+          "depth": 31.7,
+          "top_tokens": [
+            "side",
+            "select",
+            "blank",
+            "kick",
+            "sol",
+            "HI",
+            "tool",
+            "Green",
+            "Select",
+            "tools",
+            "record",
+            "remen",
+            "rem",
+            "Me",
+            "bro",
+            "Blank",
+            "lives",
+            "simple"
+          ]
+        },
+        {
+          "layer": 15,
+          "depth": 36.6,
+          "top_tokens": [
+            "ettle",
+            "fetch",
+            "linear",
+            "switch",
+            "Select",
+            "greeting",
+            "OOSE",
+            "olved",
+            "greet",
+            "cdots",
+            "subject",
+            "sus",
+            "REST",
+            "select",
+            "Green",
+            "notify",
+            "lives",
+            "tch",
+            "var",
+            "talk",
+            "discussed",
+            "deafness"
+          ]
+        },
+        {
+          "layer": 16,
+          "depth": 39.0,
+          "top_tokens": [
+            "greet",
+            "fetch",
+            "Greet",
+            "greeting",
+            "switch",
+            "ettle",
+            "notify",
+            "olved",
+            "wechsel",
+            "greet",
+            "Fetch",
+            "RESOL",
+            "discuss",
+            "tit",
+            "titles",
+            "resolve",
+            "GetData",
+            "data",
+            "tch",
+            "SWITCH",
+            "discussed",
+            "isselle",
+            "scope",
+            "isel",
+            "Greeting",
+            "NDC"
+          ]
+        },
+        {
+          "layer": 18,
+          "depth": 43.9,
+          "top_tokens": [
+            "essage",
+            "BOSS",
+            "dash",
+            "erat",
+            "boss",
+            "anners",
+            "Dash",
+            "handset",
+            "press",
+            "lain",
+            "SANT",
+            "extend",
+            "CAL",
+            "epar",
+            "BL",
+            "rion",
+            "MAIL",
+            "riends",
+            "messaging",
+            "owed",
+            "PRESS",
+            "Spacer"
+          ]
+        },
+        {
+          "layer": 20,
+          "depth": 48.8,
+          "top_tokens": [
+            "restart",
+            "golfer",
+            "Restart",
+            "bosses",
+            "Restart",
+            "adet",
+            "staffers",
+            "cribable",
+            "department",
+            "rectum",
+            "restarting",
+            "sempat",
+            "decade",
+            "unched",
+            "playlists",
+            "buttonclick",
+            "assol"
+          ]
+        },
+        {
+          "layer": 21,
+          "depth": 51.2,
+          "top_tokens": [
+            "golfer",
+            "deka",
+            "God",
+            "punt",
+            "viol",
+            "brus",
+            "del",
+            "bedingt",
+            "preservar",
+            "solle",
+            "assol"
+          ]
+        },
+        {
+          "layer": 23,
+          "depth": 56.1,
+          "top_tokens": [
+            "logique",
+            "prong",
+            "logical",
+            "undocumented",
+            "toggle",
+            "fect",
+            "jeweils",
+            "logic",
+            "Prong",
+            "individu",
+            "ausge",
+            "logical",
+            "ssk",
+            "segmented",
+            "landen"
+          ]
+        },
+        {
+          "layer": 24,
+          "depth": 58.5,
+          "top_tokens": [
+            "idium",
+            "patched",
+            "tier",
+            "prong",
+            "ausge",
+            "spf",
+            "opnieuw",
+            "welt",
+            "opcode",
+            "morally",
+            "thereof",
+            "afect",
+            "ign",
+            "osp",
+            "sacrific",
+            "wir",
+            "halk",
+            "ham",
+            "wij",
+            "Tus",
+            "spf",
+            "fect",
+            "Versions",
+            "truncated",
+            "discord",
+            "branching",
+            "sped"
+          ]
+        },
+        {
+          "layer": 26,
+          "depth": 63.4,
+          "top_tokens": [
+            "adn",
+            "ovi",
+            "individual",
+            "awn",
+            "Pla",
+            "szen",
+            "contr",
+            "contre",
+            "parent"
+          ]
+        },
+        {
+          "layer": 28,
+          "depth": 68.3,
+          "top_tokens": [
+            "cantilever",
+            "forward",
+            "bl",
+            "keduanya",
+            "extr",
+            "simultaneously",
+            "forward"
+          ]
+        },
+        {
+          "layer": 29,
+          "depth": 70.7,
+          "top_tokens": [
+            "alike",
+            "Secondly",
+            "Secondly",
+            "Thirdly",
+            "industry",
+            "Lastly",
+            "weekly",
+            "Weekly",
+            "rebounding",
+            "Weekly",
+            "bellows",
+            "flows",
+            "Subs",
+            "workflow",
+            "descent",
+            "ForThe",
+            "Together",
+            "Lastly",
+            "themselves",
+            "backing",
+            "Subs",
+            "mselves",
+            "unfolding",
+            "Epit",
+            "Sound",
+            "Document",
+            "sheath",
+            "along"
+          ]
+        },
+        {
+          "layer": 31,
+          "depth": 75.6,
+          "top_tokens": [
+            "mselves",
+            "Consequently",
+            "meanwhile",
+            "Secondly",
+            "Secondly",
+            "Meanwhile",
+            "Additionally",
+            "Additionally",
+            "Furthermore",
+            "Yours",
+            "Secondary",
+            "Consequently",
+            "Early",
+            "Stem",
+            "EndOf",
+            "Increased",
+            "Bill",
+            "Its",
+            "Moreover",
+            "Document",
+            "Conversely",
+            "Lastly",
+            "Thirdly",
+            "Simultaneously",
+            "nThe",
+            "Independent"
+          ]
+        },
+        {
+          "layer": 32,
+          "depth": 78.0,
+          "top_tokens": [
+            "Secondly",
+            "Secondly",
+            "Additionally",
+            "nThe",
+            "Additionally",
+            "Consequently",
+            "meanwhile",
+            "mselves",
+            "Yours",
+            "furthermore",
+            "Meanwhile",
+            "Secondary",
+            "Consequently",
+            "Lesser",
+            "sheath",
+            "Early",
+            "Moreover",
+            "Biological",
+            "Steinberg",
+            "Tragedy",
+            "Furthermore",
+            "Within"
+          ]
+        },
+        {
+          "layer": 34,
+          "depth": 82.9,
+          "top_tokens": [
+            "Secondly",
+            "secondly",
+            "furthermore",
+            "Secondly",
+            "ingly",
+            "Consequently",
+            "Allow",
+            "heave",
+            "heading",
+            "Ships",
+            "paragraph",
+            "unim",
+            "conversely",
+            "Bid",
+            "subheading",
+            "Morton",
+            "Posters",
+            "Otherwise",
+            "StringTo",
+            "framing",
+            "Sick",
+            "Furthermore",
+            "Consequently",
+            "Peripheral",
+            "consequently",
+            "sted",
+            "Except"
+          ]
+        },
+        {
+          "layer": 36,
+          "depth": 87.8,
+          "top_tokens": [
+            "endswith",
+            "Secondly",
+            "chapter",
+            "heading",
+            "secondly",
+            "Allow",
+            "piece",
+            "paragraph",
+            "heading",
+            "versa",
+            "plant",
+            "sure",
+            "Havre",
+            "Otherwise",
+            "Secondly",
+            "Early",
+            "Border",
+            "Meanwhile",
+            "Survival",
+            "conversely",
+            "Second",
+            "subheading",
+            "heave"
+          ]
+        },
+        {
+          "layer": 37,
+          "depth": 90.2,
+          "top_tokens": [
+            "paragraph",
+            "heading",
+            "secondly",
+            "Secondly",
+            "sure",
+            "Early",
+            "endswith",
+            "heading",
+            "Exercise",
+            "chapter",
+            "weak",
+            "heave",
+            "wind",
+            "Meanwhile",
+            "Allow",
+            "subheading",
+            "Begin",
+            "Border",
+            "Havre",
+            "plant",
+            "border",
+            "subsection",
+            "piece",
+            "part",
+            "begin"
+          ]
+        },
+        {
+          "layer": 39,
+          "depth": 95.1,
+          "top_tokens": [
+            "Secondly",
+            "heading",
+            "secondly",
+            "heading",
+            "hole",
+            "black",
+            "sure",
+            "Border",
+            "imiter",
+            "hline",
+            "Border",
+            "border",
+            "Secondly",
+            "heave",
+            "alone",
+            "paragraph",
+            "closed",
+            "chapter",
+            "ingly",
+            "ently",
+            "igator",
+            "igate",
+            "versa"
+          ]
+        },
+        {
+          "layer": 41,
+          "depth": 100.0,
+          "top_tokens": [
+            "end",
+            "The",
+            "he",
+            "black",
+            "hole",
+            "parts",
+            "meet",
+            "Secondly",
+            "tip",
+            "padding",
+            "bordered",
+            "closed",
+            "part",
+            "hline",
+            "Sb",
+            "hd",
+            "ond",
+            "append",
+            "to",
+            "catch",
+            "em",
+            "here",
+            "Stand",
+            "filled",
+            "et",
+            "caption",
+            "second",
+            "more",
+            "eth",
+            "apply",
+            "meat",
+            "angle"
+          ]
+        }
+      ],
+      "logit_surprising": [
+        {
+          "concept": "ausge",
+          "score": 11.0,
+          "best_rank": 1,
+          "best_depth": 56.1,
+          "near_token": "A",
+          "output_checked": true
+        },
+        {
+          "concept": "shape",
+          "score": 9.0,
+          "best_rank": 1,
+          "best_depth": 68.3,
+          "near_token": "into",
+          "output_checked": true
+        },
+        {
+          "concept": "cylindrical",
+          "score": 8.0,
+          "best_rank": 1,
+          "best_depth": 75.6,
+          "near_token": "a",
+          "output_checked": true
+        },
+        {
+          "concept": "symmetrically",
+          "score": 7.0,
+          "best_rank": 1,
+          "best_depth": 68.3,
+          "near_token": "stretched",
+          "output_checked": true
+        },
+        {
+          "concept": "halves",
+          "score": 6.0,
+          "best_rank": 1,
+          "best_depth": 68.3,
+          "near_token": "The",
+          "output_checked": true
+        },
+        {
+          "concept": "lengthwise",
+          "score": 6.0,
+          "best_rank": 1,
+          "best_depth": 70.7,
+          "near_token": "stretched",
+          "output_checked": true
+        },
+        {
+          "concept": "protrusions",
+          "score": 6.0,
+          "best_rank": 1,
+          "best_depth": 75.6,
+          "near_token": "erical",
+          "output_checked": true
+        },
+        {
+          "concept": "pattern",
+          "score": 5.0,
+          "best_rank": 1,
+          "best_depth": 70.7,
+          "near_token": "tearing",
+          "output_checked": true
+        },
+        {
+          "concept": "tapering",
+          "score": 5.0,
+          "best_rank": 1,
+          "best_depth": 75.6,
+          "near_token": "before",
+          "output_checked": true
+        },
+        {
+          "concept": "groove",
+          "score": 5.0,
+          "best_rank": 1,
+          "best_depth": 75.6,
+          "near_token": "slanted",
+          "output_checked": true
+        },
+        {
+          "concept": "ends",
+          "score": 5.0,
+          "best_rank": 1,
+          "best_depth": 78.0,
+          "near_token": "The",
+          "output_checked": true
+        },
+        {
+          "concept": "verw",
+          "score": 4.0,
+          "best_rank": 1,
+          "best_depth": 58.5,
+          "near_token": "faces",
+          "output_checked": true
+        },
+        {
+          "concept": "remained",
+          "score": 4.0,
+          "best_rank": 1,
+          "best_depth": 78.0,
+          "near_token": "faces",
+          "output_checked": true
+        },
+        {
+          "concept": "ostat",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 48.8,
+          "near_token": "stretched",
+          "output_checked": true
+        },
+        {
+          "concept": "punt",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 51.2,
+          "near_token": "aluminum",
+          "output_checked": true
+        },
+        {
+          "concept": "each",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 70.7,
+          "near_token": "on",
+          "output_checked": true
+        },
+        {
+          "concept": "elongated",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 75.6,
+          "near_token": "pronounced",
+          "output_checked": true
+        },
+        {
+          "concept": "intricate",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 75.6,
+          "near_token": "contained",
+          "output_checked": true
+        },
+        {
+          "concept": "embedded",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 75.6,
+          "near_token": "inclusions",
+          "output_checked": true
+        },
+        {
+          "concept": "smooth",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 78.0,
+          "near_token": "a",
+          "output_checked": true
+        },
+        {
+          "concept": "table",
+          "score": 3.0,
+          "best_rank": 1,
+          "best_depth": 82.9,
+          "near_token": "round",
+          "output_checked": true
+        },
+        {
+          "concept": "boo",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 51.2,
+          "near_token": "stretched",
+          "output_checked": true
+        },
+        {
+          "concept": "God",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 51.2,
+          "near_token": "before",
+          "output_checked": true
+        },
+        {
+          "concept": "dazu",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 56.1,
+          "near_token": "aluminum",
+          "output_checked": true
+        },
+        {
+          "concept": "halk",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 58.5,
+          "near_token": "slanted",
+          "output_checked": true
+        },
+        {
+          "concept": "datasheet",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 68.3,
+          "near_token": "aluminum",
+          "output_checked": true
+        },
+        {
+          "concept": "entrambi",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 68.3,
+          "near_token": "pits",
+          "output_checked": true
+        },
+        {
+          "concept": "shaped",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 68.3,
+          "near_token": "hemisph",
+          "output_checked": true
+        },
+        {
+          "concept": "Each",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 68.3,
+          "near_token": "rim",
+          "output_checked": true
+        },
+        {
+          "concept": "rectangular",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 70.7,
+          "near_token": "a",
+          "output_checked": true
+        },
+        {
+          "concept": "were",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 70.7,
+          "near_token": "faces",
+          "output_checked": true
+        },
+        {
+          "concept": "axis",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 70.7,
+          "near_token": "slanted",
+          "output_checked": true
+        },
+        {
+          "concept": "donc",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 75.6,
+          "near_token": "<bos>",
+          "output_checked": true
+        },
+        {
+          "concept": "Yours",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 75.6,
+          "near_token": ".",
+          "output_checked": true
+        },
+        {
+          "concept": "around",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 75.6,
+          "near_token": "centered",
+          "output_checked": true
+        },
+        {
+          "concept": "machined",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 78.0,
+          "near_token": "bar",
+          "output_checked": true
+        },
+        {
+          "concept": "plate",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 82.9,
+          "near_token": "aluminum",
+          "output_checked": true
+        },
+        {
+          "concept": "tubing",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 82.9,
+          "near_token": "bar",
+          "output_checked": true
+        },
+        {
+          "concept": "engraved",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 82.9,
+          "near_token": "contained",
+          "output_checked": true
+        },
+        {
+          "concept": "perfectly",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 82.9,
+          "near_token": "pits",
+          "output_checked": true
+        },
+        {
+          "concept": "encir",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 82.9,
+          "near_token": "rim",
+          "output_checked": true
+        },
+        {
+          "concept": "priori",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 87.8,
+          "near_token": "A",
+          "output_checked": true
+        },
+        {
+          "concept": "depressions",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 87.8,
+          "near_token": "erical",
+          "output_checked": true
+        },
+        {
+          "concept": "opposing",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 87.8,
+          "near_token": "on",
+          "output_checked": true
+        },
+        {
+          "concept": "within",
+          "score": 2.0,
+          "best_rank": 1,
+          "best_depth": 87.8,
+          "near_token": "inclusions",
+          "output_checked": true
+        },
+        {
+          "concept": "Elsewhere",
+          "score": 1.0,
+          "best_rank": 1,
+          "best_depth": 39.0,
+          "near_token": "A",
+          "output_checked": true
+        },
+        {
+          "concept": "CRED",
+          "score": 1.0,
+          "best_rank": 1,
+          "best_depth": 39.0,
+          "near_token": "on",
+          "output_checked": true
+        },
+        {
+          "concept": "kuri",
+          "score": 1.0,
+          "best_rank": 1,
+          "best_depth": 43.9,
+          "near_token": "bar",
+          "output_checked": true
+        },
+        {
+          "concept": "bit",
+          "score": 1.0,
+          "best_rank": 1,
+          "best_depth": 43.9,
+          "near_token": "stretched",
+          "output_checked": true
+        },
+        {
+          "concept": "combined",
+          "score": 1.0,
+          "best_rank": 1,
+          "best_depth": 43.9,
+          "near_token": "The",
+          "output_checked": true
+        },
+        {
+          "concept": "tread",
+          "score": 1.0,
+          "best_rank": 1,
+          "best_depth": 43.9,
+          "near_token": "erical",
+          "output_checked": true
+        },
+        {
+          "concept": "sest",
+          "score": 1.0,
+          "best_rank": 1,
+          "best_depth": 43.9,
+          "near_token": "a",
+          "output_checked": true
+        },
+        {
+          "concept": "ello",
+          "score": 1.0,
+          "best_rank": 1,
+          "best_depth": 48.8,
+          "near_token": "aluminum",
+          "output_checked": true
+        },
+        {
+          "concept": "heat",
+          "score": 1.0,
+          "best_rank": 1,
+          "best_depth": 48.8,
+          "near_token": "The",
+          "output_checked": true
+        },
+        {
+          "concept": "dieu",
+          "score": 1.0,
+          "best_rank": 1,
+          "best_depth": 48.8,
+          "near_token": "hemisph",
+          "output_checked": true
+        },
+        {
+          "concept": "ticker",
+          "score": 1.0,
+          "best_rank": 1,
+          "best_depth": 48.8,
+          "near_token": "pits",
+          "output_checked": true
+        },
+        {
+          "concept": "retomar",
+          "score": 1.0,
+          "best_rank": 1,
+          "best_depth": 48.8,
+          "near_token": "tearing",
+          "output_checked": true
+        },
+        {
+          "concept": "restart",
+          "score": 1.0,
+          "best_rank": 1,
+          "best_depth": 48.8,
+          "near_token": ".",
+          "output_checked": true
+        },
+        {
+          "concept": "del",
+          "score": 1.0,
+          "best_rank": 1,
+          "best_depth": 51.2,
+          "near_token": "mating",
+          "output_checked": true
+        },
+        {
+          "concept": "brus",
+          "score": 1.0,
+          "best_rank": 1,
+          "best_depth": 51.2,
+          "near_token": "pits",
+          "output_checked": true
+        },
+        {
+          "concept": "kau",
+          "score": 1.0,
+          "best_rank": 1,
+          "best_depth": 56.1,
+          "near_token": "<bos>",
+          "output_checked": true
+        },
+        {
+          "concept": "prong",
+          "score": 1.0,
+          "best_rank": 1,
+          "best_depth": 56.1,
+          "near_token": "The",
+          "output_checked": true
+        },
+        {
+          "concept": "dabei",
+          "score": 1.0,
+          "best_rank": 1,
+          "best_depth": 56.1,
+          "near_token": "and",
+          "output_checked": true
+        },
+        {
+          "concept": "Jaz",
+          "score": 1.0,
+          "best_rank": 1,
+          "best_depth": 58.5,
+          "near_token": "aluminum",
+          "output_checked": true
+        }
+      ],
+      "figures": {
+        "grid": "figures/gemma4-e4b-it-heldout-v1-smoke-seed0/materials/heldout-v1-assoc-ductile-01__grid.png",
+        "stream": "figures/gemma4-e4b-it-heldout-v1-smoke-seed0/materials/heldout-v1-assoc-ductile-01__stream.png",
+        "trajectory": "figures/gemma4-e4b-it-heldout-v1-smoke-seed0/materials/heldout-v1-assoc-ductile-01__trajectory.png",
+        "emergence": "figures/gemma4-e4b-it-heldout-v1-smoke-seed0/materials/heldout-v1-assoc-ductile-01__emergence.png",
+        "heatmap": "figures/gemma4-e4b-it-heldout-v1-smoke-seed0/materials/heldout-v1-assoc-ductile-01__heatmap.png"
+      }
+    }
+  ]
+}

jlens_materials/runs/gemma4-materials-causal_analysis.json

@@ -0,0 +1,49 @@
+{
+  "model": "google/gemma-4-E4B-it",
+  "provider": "openai",
+  "analysis_model": "gpt-5.5",
+  "offline": false,
+  "per_prompt": [
+    {
+      "slug": "materials-ductile-brittle-swap",
+      "title": "Ductile-to-brittle causal counterfactual",
+      "shape": "PROBE_SWAP",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "This item is **excluded from registered causal/protocol success** because the clean-answer control failed: the expected first token was **“elongation”**, but the model greedily produced **“plastic”**. Qualitatively, the J-lens does show an interpretable materials-state trajectory in the workspace band: **“ductility”** is top-ranked around 39.0–43.9% depth, then the readout shifts toward consequence/answer concepts such as **“elongation”** at 73.2–75.6%, **“deformation”** at 78.0–85.4%, and finally **“plastic”** from 87.8% onward, matching the actual next-token output. That is a real semantic pattern, but not a clean registered success; moreover, for the tracked **ductile** concept the J-lens peak is rank 5 at 80.5%, while the matched vanilla logit lens is rank 2, so this is not a case where the J-lens clearly outperforms the logit baseline.\n\nThe causal swap does **not** confirm a causal downstream role. The intervention **ductile → brittle** only modestly suppressed the source concept and weakly installed the target: **ductile** full-vocab rank worsened from 20 to 37, while **brittle** improved only from 127 to 119. The declared next-token distribution did not switch toward a brittle-failure answer: clean top tokens were `plastic, strain, neck, uniform, work`, and after the swap they remained essentially `plastic, strain, uniform, neck, plasticity`. So this is best described as a mild representational suppression/installation with no behavioral answer change, not a causal success.\n\nThe exploratory candidate list is useful only as a heuristic. **“ductility,” “ductile,” and “deformation”** are meaningful surfaced concepts consistent with the prompt and the top-token layer sequence; **“atoms”** may reflect inferred microscopic/materials structure but is not the target answer; **“undergoes”** is plausible completion scaffolding rather than a stable concept. By contrast, **“because,” “therefore/Therefore,” “which,” and “was”** look mostly like discourse or syntactic scaffolding, not substantive hidden reasoning. The slice grid also contains many truncated or stem-like tokens such as `deformat`, `fracturi`, `ductilit`, and `metallur`, which should be treated as tokenizer artifacts or single-token limitations rather than clean verbalized concepts."
+    },
+    {
+      "slug": "materials-pileup-focus",
+      "title": "Focus on a dislocation pile-up",
+      "shape": "MODULATION",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "This included directed-modulation record shows only a weak, unsustained materials-side signal in the J-lens, and by itself it is not a modulation result because the matched suppress and neutral records are not provided here. Within the fixed workspace band, **stress** reaches J-rank 11 at 43.9% depth, but the matched vanilla logit lens is better at rank 8. **Pile** reaches J-rank 20 at 87.8%, only slightly better than logit rank 23. **Dislocation** reaches only J-rank 101 at 85.4%, while the logit lens is much stronger at rank 13. None has a sustained onset. So the registered readout does not show a clean J-lens-specific emergence of the unspoken concept; it mostly says that the model has some prompt-conditioned availability of the materials words, but not a robust workspace representation beyond the logit baseline.\n\nQualitatively, the slice grid does contain plausible side-concept traces around the instruction/concept positions: tokens like **dislocat**, **pile**, **stacking**, **pressure**, **tensile**, **stress**, **metallur**, and **fracturi** appear in the workspace band. That is consistent with the model attending to the “dislocation pile-up / stress concentration” instruction while producing the neutral carrier sentence. But these are largely near or derived from the prompted side-concept text, and the readout-position top tokens are dominated by generic discourse or carrier-sentence alternatives such as “Its,” “While,” “The,” and literal sentence tokens. I would treat this as a qualitative prompt-conditioning pattern, not a clean demonstration of unspoken reasoning.\n\nThe exploratory candidate list is heuristic only. **because**, **but**, **also/Also** are generic connective continuations and likely reflect instruction-following or discourse defaults rather than materials reasoning, even though “because” peaks near “dislocation.” **motionless**, **was**, and **shimmering** look like carrier-sentence or furnace/description imagery artifacts; **motionless** is an echo of “remained,” and **shimmering** plausibly comes from “blue/furnace” imagery. **writing** is just a morphological/task echo of “Write,” and **whatsoever** is an instruction-control echo of “Do not write anything else.” **implications** is the only candidate that might weakly reflect the “stress concentration/effects” side instruction, but it is too generic to count as a materials concept. None of these candidates should be treated as a paper metric success."
+    },
+    {
+      "slug": "materials-pileup-suppress",
+      "title": "Suppress a dislocation pile-up",
+      "shape": "MODULATION",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "This included suppress item is **not a clean J-lens success for the forbidden materials concept**. Within the fixed 38–92% workspace band, **dislocation** never becomes salient: best J rank is 8513 at 46.3%, worse than the matched logit-lens rank 5531. **Stress** is also weak: J rank 1269 at 87.8%, essentially identical to logit rank 1268. **Pile** is the only strong-looking tracked token, reaching J rank 52 at 90.2%, but the logit lens is actually better at rank 37, and there is no sustained onset. This looks more like a late single-token/subword echo of the prompted phrase “pile-up” than a J-specific workspace representation of the full materials concept. The slice grid similarly shows local prompt-conditioned traces around the “dislocation / pile / stress / concentration” positions, but the assistant-response region is dominated by sentence-writing and final-token content such as “sentence,” “write,” “Period,” and copied words from “The blue notebook remained beside the furnace.”\n\nSo the lens suggests the model processed the suppression instruction and the forbidden phrase locally, but it does **not** show a robust unspoken “dislocation pile-up / stress concentration” concept being carried into the response workspace. The clean output absence is satisfied by the completion, but because the target concepts are explicitly in the prompt and the J-lens is not stronger than the logit baseline, this should be read as weak qualitative evidence at most, not as a registered hidden-thought discovery. The best apparent target signal, **pile**, is especially vulnerable to the single-token limitation: it does not establish representation of the compound concept “dislocation pile-up.”\n\nFor the exploratory candidate list, the meaningful entries are mostly task/carrier related: **writing** reflects the instruction to write the sentence; **motionless** is a plausible semantic neighbor of “remained”; **was** is a banal tense/syntax alternative near “notebook”; and **implications** may be a loose association to “its stress concentration,” though it is not a metric result. **grammat** is partly interpretable as “grammar/grammatical sentence,” but it is also a tokenizer fragment. **also**, **Also**, **but**, and **because** look like generic discourse-prior artifacts rather than meaningful reasoning, and **weird** is likely small-model/noise association rather than a reliable concept."
+    },
+    {
+      "slug": "materials-pileup-control",
+      "title": "Neutral carrier control",
+      "shape": "MODULATION",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "This included neutral-control item shows no clean materials-side-task signal. In the fixed workspace band (38–92% depth), the registered concepts do not show sustained onset: **dislocation** is weak and worse than vanilla logit lens at its best point (J rank 4809 at 70.7% vs logit rank 3153), **stress** is only modestly better than logit but still low-ranked (J 1641 at 85.4% vs logit 2700), and **pile** becomes fairly high-ranked late (J 59 at 90.2%) but the matched logit lens is actually better (rank 47). So this should not be counted as evidence that the model is representing unspoken materials concepts; if anything, the control behaves as desired. The layer/position grid is dominated by instruction-following and literal-copy content: mid-band tokens like **grammat**, **verbatim**, **whatsoever**, **sentence/wording**, and later the actual carrier sentence and punctuation/period. That looks like “copy exactly, do not add anything,” not hidden reasoning about materials.\n\nThe exploratory candidate list is useful only qualitatively. **verbatim** and **whatsoever** are plausibly meaningful: they reflect the instruction to write exactly the sentence and nothing else. **color** near “blue” and perhaps **motionless** near “remained” are local semantic associations, not task-relevant intermediates. **was** is likely a syntactic alternative for the sentence frame (“notebook was…”) rather than a concept the model computed for a side task. **also/Also**, **because**, and **and** look like generic discourse-token artifacts, and **masterpiece** is best treated as noise or small-model/lens contamination rather than a meaningful hidden concept. Overall, the J-lens reveals a workspace concerned with exact reproduction and output suppression, with no registered clean emergence of **dislocation**, **pile**, or **stress** beyond what the vanilla logit lens already explains."
+    },
+    {
+      "slug": "materials-defect-report",
+      "title": "Crystal-defect verbal-report swap",
+      "shape": "REPORT_SWAP",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "Within the fixed workspace band, the lens shows a fairly clean emergence of the chosen verbal report, **“vacancy.”** The readout position first carries generic answer/task structure and defect-related material vocabulary: around 73–78% depth the top tokens include **Dislocations**, **Fracture**, **Distortion**, **Schottky**, and by 78–80.5% it has moved to **defects / imperfection / Vacancy**. The registered concept then locks onto **vacancy**, reaching J-lens rank 1 at **82.9% depth** with sustained onset at **80.5%**, matching the actual completion “Vacancy.” This is not a strong J-lens-over-logit case, however: the matched vanilla logit lens also gives **vacancy rank 1**. The sibling defects are present only as weaker alternatives: **dislocation** reaches J rank 8 at 87.8% versus logit rank 9, and **interstitial** reaches J rank 20 at 90.2% versus logit rank 21. **Twin** and **boundary** remain poor, with twin actually better in the logit lens than the J-lens. So the readout supports “the model is settling on vacancy from a defect-family representation,” not a clean hidden plan to output one of the siblings.\n\nThe causal swap does **not** confirm controllable replacement of the verbal report. The registered success rate is **0/5**. Suppressing the source concept often knocked **vacancy** down substantially, but adding a sibling coordinate did not make the output become **dislocation**, **interstitial**, **twin**, or **boundary**; every swapped trial’s top-1 was **`<turn|>`**. That is suppression or disruption of the report channel, not a successful report swap. The case-only “vacancy” self-target is also not a sibling-concept test, and it failed as well. Thus the primary diagnostic—whether the one-word report changes to the installed sibling—comes out negative.\n\nThe exploratory candidate list should be treated only as a heuristic. Items such as **Example**, **sentence**, **only**, **Answer/Describe/Explain**-like grid tokens are plausibly task-format or instruction-compliance features rather than defect concepts. **Because**, **and**, **also**, **that**, and **about** are mostly function-word or discourse artifacts from all-position top-1 patterns, not meaningful scientific intermediates. **Description** may reflect the prompt style, and **imaginative** is likely noise or a prompt-interpretation artifact. None of these should be counted as protocol evidence; the meaningful content here is the late emergence of the actual report **vacancy**, with weak co-activation of nearby single-token defect alternatives."
+    }
+  ],
+  "synthesis": "Across the supplied items, the clearest workspace-like phenomenon is the **REPORT_SWAP / verbal-report** record for the crystal-defect prompt: the model settles on the single-token report **“vacancy”**, with J-lens rank 1 at **82.9% depth** and sustained onset around **80.5%**. However, this is mainly a readout of the answer the model is about to say, not a hidden intermediate, and the matched vanilla logit lens also ranks **vacancy** at 1. The sibling concepts **dislocation** and **interstitial** appear only as weaker alternatives, and the actual causal report-swap trials fail: installing sibling defect concepts does not make the model output those siblings, instead disrupting the report channel to `<turn|>`. Thus this is a real semantic settling pattern, but not a successful causal report-swap demonstration and not a J-lens-specific advantage over the logit lens.\n\nThe **MODULATION** records are much weaker. In the focus/suppress/control carrier-sentence setting, the lens shows local prompt-conditioned traces of materials vocabulary such as **dislocation**, **pile**, **stress**, **stacking**, **tensile**, and **metallur**, but the registered tracked concepts do not show robust, sustained J-lens emergence in the workspace band. Often the vanilla logit lens is equal or better: for example, in the focus item **dislocation** is much stronger in the logit lens than the J-lens, and in the suppress/control items **pile** is late and logit-favored. Because these modulation groups are exploratory and sample-insufficient, and because the side concepts are explicitly present in the prompt, these should be treated as qualitative prompt-conditioning patterns rather than registered evidence of unspoken scientific reasoning.\n\nThe **ductile-to-brittle PROBE_SWAP** item gives the most interesting qualitative ordered trajectory: **ductility** appears around **39.0–43.9% depth**, then consequence/answer concepts such as **elongation** at **73.2–75.6%**, **deformation** at **78.0–85.4%**, and finally **plastic** from **87.8% onward**, matching the model’s actual answer. This resembles the paper’s claim that intermediate concepts can appear before the final answer, and that later layers move toward the literal next token. But it is explicitly excluded from registered causal/protocol success because the clean-answer control failed: the expected first token was **“elongation”**, while the model greedily produced **“plastic.”** In addition, the tracked **ductile** signal is not J-lens-superior; its best J rank is 5 while the matched logit-lens rank is 2. The causal swap also fails to install **brittle** strongly or change the answer distribution. No supplied records cleanly test **MULTIHOP**, **ASSOCIATION**, or **RECOGNITION**, so no run-level conclusion should be drawn about those experiment types.\n\nOverall, this run is consistent with the paper’s cautions more than with its strongest positive claims. The lens can reveal interpretable workspace content when the target is a clean single vocabulary token, as with **vacancy**, and it can show plausible semantic trajectories in individual examples, as in the excluded ductility case. But the run does not provide a clean registered demonstration of hidden intermediate scientific concepts being computed in order, nor a successful causal swap. The single-token limitation is substantial in this materials domain: **dislocation pile-up**, **stress concentration**, and related metallurgical concepts are multi-token or fragmented into subword pieces, so isolated tokens like **pile**, **dislocat**, **ductilit**, or **fracturi** should not be overread as full concepts. The weak or absent signals for some intended concepts should also be taken seriously: on this smaller Gemma target, the lens may be faithfully reporting that the model does not robustly maintain those scientific representations.\n\nThe exploratory surfaced candidates are best viewed as a heuristic for noticing patterns, not as a paper metric. Plausible but weak task-related candidates include **writing**, **whatsoever**, and **verbatim**, which reflect instruction compliance; **motionless** and **was**, which are carrier-sentence alternatives around “remained”; and perhaps **implications**, which could loosely relate to “stress concentration” but is too generic to count as a materials concept. Frequent candidates such as **because**, **also/Also**, **but**, and **and** are mostly discourse-prior artifacts. **masterpiece** looks like small-model or lens noise rather than meaningful hidden reasoning. A reader should therefore treat google/gemma-4-E4B-it as useful for exploratory interpretability demos and for debugging the J-lens pipeline, but not as an ideal target for strong scientific-reasoning claims. A larger or more capable model, with more stable domain knowledge and better single-token verbalizations of scientific concepts, would be a more appropriate target for testing the paper’s global-workspace claims in materials science."
+}

jlens_materials/runs/gemma4-materials-paper-v1_analysis.json

@@ -0,0 +1,121 @@
+{
+  "model": "google/gemma-4-E4B-it",
+  "provider": "openai",
+  "analysis_model": "gpt-5.5",
+  "offline": false,
+  "per_prompt": [
+    {
+      "slug": "paper-v1-ductile-fracture-replication",
+      "title": "Ductile-fracture replication vignette",
+      "shape": "ASSOCIATION",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "This is an included protocol item, and the fixed-band readout shows a real but partial ductile-fracture replication. In the workspace band, **coalescence** is the cleanest signal: J-lens rank **13 at 41.5% depth** versus logit-lens rank **3048**, a very large improvement for an unspoken mechanism implied by rounded dimples/inclusions and cup-and-cone morphology. **Ductile** is also present, though less dramatically: rank **55 at 41.5%** versus logit **171**. **Nucleation** improves too, rank **150 at 51.2%** versus logit **599**, but it is weaker and not sustained. **Void** is not a J-lens success here: rank **6456 at 53.7%**, worse than the matched logit rank **1166**. Since all tracked concepts have sustained onset “None,” this is a transient mid-band emergence, not a stable plateau. Also, the order does **not** support a nucleation-then-coalescence mechanism: coalescence peaks earlier and much more strongly than nucleation, so no ordered mechanistic computation should be inferred.\n\nThe layer-position grid qualitatively supports the same interpretation but should not be treated as the metric. Around the workspace band it shows materials/fractography language such as **ductile**, **ductility**, **metallur-**, **deformat-**, **microstr-**, **contour**, **cup**, **neck**, and **shear lip** near relevant prompt positions, suggesting the model is organizing the vignette as a ductile deformation/fracture case. However, some of this is local lexical association and tokenizer-shaped fragments rather than a clean mechanistic trace. The tracked-rank chart is the stronger evidence: it says the model has the verbalizable concept **microvoid coalescence/ductile fracture**, but the single-token **void** probe fails, likely reflecting either token-choice mismatch or absence of that exact verbalization.\n\nFor the exploratory candidates: **ductile** is meaningful, and its many top-band cells are consistent with the registered ductile signal. **ductility** and **deformation** are also meaningful inferred concepts from “extensive necking,” though they are auxiliary rather than the registered mechanism. **metallurgical** is plausibly domain-contextual but broad. **characteristic** and **indicating** are semi-meaningful diagnostic/explanatory framing words, not fracture mechanisms. **which**, **This**, and **because** are discourse/function-token artifacts. **cracks** is not a reliable positive signal here: ductile dimples are not simply cracks, so this looks like generic fracture-language bleed-through rather than a precise concept."
+    },
+    {
+      "slug": "paper-v1-intergranular-replication",
+      "title": "Boundary-attack replication vignette",
+      "shape": "ASSOCIATION",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "Within the fixed workspace band, the lens shows a real but limited corrosion-family interpretation. The clearest absolute signal is **corrosion**: J-lens rank **12 at 41.5% depth**, with nearby early-band support in the slice grid around the chloride/groove region. However, the matched logit-lens rank is **14**, essentially the same, so this is not a strong J-specific hidden-concept result; the surface cues already strongly push “corrosion.” The more interesting J-vs-logit gaps are for **boundary** and **sensitization**: boundary reaches J rank **886 at 39.0%** versus logit rank **4034**, and sensitization reaches J rank **1077 at 41.5%** versus logit rank **15902**. Those are substantial relative improvements, suggesting the transported representation contains some unspoken grain-boundary/sensitization interpretation. But both remain only around rank 10³ and have **no sustained onset**, so this should be called suggestive rather than a clean recovery. **Intergranular** is correctly dropped as multi-token/non-trackable, so the specific intergranular-attack concept cannot be registered by this metric.\n\nThe exploratory candidate list is useful qualitatively but not a paper metric. Plausibly meaningful entries are **metallurgical** as a domain frame, **corrosion** as the main inferred process, **fissures** as a morphology related to the narrow connected grooves, and **unaffected** as a paraphrase of the untouched grain interiors. **Temperatures** is mostly a surface restatement of “650 degrees Celsius,” not an unspoken mechanistic concept. **Because**, **during**, and **subsequently** are discourse/causal-continuation tokens rather than materials concepts, and **and** / **was** are function-word artifacts. Overall, the lens does reveal the model leaning toward corrosion and weakly toward sensitization/boundary attack, but the strongest generic term is not stronger than vanilla logit lens, and the specific mechanism remains sub-top, unsustained, and partly hidden behind the single-token limitation."
+    },
+    {
+      "slug": "paper-v1-fatigue-fractography",
+      "title": "Cyclic-fracture vignette",
+      "shape": "ASSOCIATION",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "Within the registered workspace band, this is a **weak-to-moderate protocol signal for the cyclic damage family**, not a clean emergence of the exact answer concept. The best tracked J-lens result is **propagation**, which reaches rank **337 at 41.5% depth**, substantially better than the matched vanilla logit-lens rank **2228**. That is a real J-lens advantage and fits the prompt’s “ridges advancing” / crack-growth semantics. By contrast, **fatigue** only becomes moderately good late, rank **239 at 87.8%**, and the vanilla logit lens is actually slightly better at **190**, so this is not strong evidence that the J-lens uniquely reveals an unspoken “fatigue” concept. **Crack** is also not a J-lens win: rank **1525** versus logit-lens **631**. None of the tracked concepts has a sustained onset, so the rank plot does not show a clean ordered computation of “crack → propagation → fatigue”; it mostly shows a transient propagation signal and a late general rise of fatigue/crack-related vocabulary.\n\nThe layer-position grid qualitatively looks more semantically organized in the workspace band: around the fracture-description positions it contains **fracture, cracks, propagat-/extending/radiating, surface, ridges, microscopy**, and later even **fatigue**. That supports the idea that the model is representing a coherent materials/fractography situation, but this is a qualitative all-position pattern, not the registered metric. At the fixed readout position, the per-layer top tokens are dominated first by observation/microscopy terms and then by discourse tokens like **This**, so the final readout position alone does not strongly verbalize “fatigue failure.”\n\nFor the exploratory candidate list, **cracks** is the most meaningful heuristic hit: it tops many band cells and is scientifically appropriate, though it does not rescue the tracked singular **crack** metric. **metallurgical** is plausibly a meaningful domain cue, and **indicating** is an inference connective consistent with the prompt, but neither is the target concept. **during, because, This, where, was, throughout, years** are mostly discourse, syntax, or paraphrase artifacts; “years” may loosely reflect long duration but is not diagnostic of fatigue. Also, **striations** was dropped as multi-token, which matters here because it is one of the most diagnostic fractographic clues; the single-token limitation likely understates part of the model’s available concept."
+    },
+    {
+      "slug": "paper-v1-cleavage-fractography",
+      "title": "Low-temperature faceted-fracture vignette",
+      "shape": "ASSOCIATION",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "This item is **excluded from registered metric interpretation** because the intermediate concept **“fracture”** is already present in the input. Qualitatively, the clearest J-lens signal is for **brittle**: it reaches rank **567 at 43.9% depth** within the fixed workspace band, far stronger than the matched vanilla logit-lens rank (**13,413**). That is a real association signal: the vignette has made the model disposed to say “brittle” even though the word is not in the prompt. However, it is not sustained, and the final-position top tokens are mostly generic descriptors or discourse tokens rather than a stable material-science answer. **Fracture** also improves under the J-lens, rank **1,428 at 85.4%** versus logit rank **3,344**, but because “fracture” is literally in the prompt this cannot count as an unspoken-concept result. **Cleavage** is not convincingly recovered: its best J-lens rank is only **7,687 at 90.2%**, and the vanilla logit lens is actually better at **5,739**, so there is no registered J-lens advantage for the specific cleavage interpretation. The multi-token **transgranular** target is correctly dropped as untrackable.\n\nThe slice grid gives a plausible qualitative picture of materials-domain processing: around the workspace band it surfaces local concepts such as **ductility**, **metallurgical**, **failure**, **negligible**, **fracture**, and occasional **brittle** near relevant prompt positions. But this should not be upgraded into a metric; the scored final-token trajectory only supports a transient **brittle** association, not a clean ordered emergence of “brittle cleavage.” In the exploratory candidate list, **ductility** is the most meaningful heuristic candidate, tied to “almost no necking” and “minimal plastic deformation”; **metallurgical** is a broad but sensible domain label; **failure**, **negligible**, and **temperatures** are mostly paraphrases or local restatements of input content; **indicating** is a plausible explanatory connective but not a scientific concept. **because**, **This**, **which**, and **also** are mainly discourse/syntax artifacts. Overall, the lens shows that the model has some unspoken **brittle-fracture** association, but it does not cleanly verbalize the more specific **cleavage** mechanism in the registered readout."
+    },
+    {
+      "slug": "paper-v1-hall-petch-constrained",
+      "title": "Constrained grain-size strengthening",
+      "shape": "MULTIHOP",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "This item is **excluded** from registered success: the clean model’s greedy first token was `**`, not `higher`. Within the fixed workspace band, the registered intermediate concepts do **not** give a clean Hall–Petch reasoning trace. `strengthening` is somewhat J-lens-favored over the matched logit lens, peaking at J rank **3,831** at **41.5%** depth versus logit rank **10,124**, but that is still weak and unsustained. `Hall` is worse for the J-lens than the logit baseline: J rank **16,297** at **53.7%** versus logit rank **1,171**. `Petch` was dropped as multi-token. So there is no registered ordered emergence of “Hall → strengthening → higher.”\n\nQualitatively, the readout does show the model drifting toward the correct comparative answer late: `higher` appears in the top tokens at **85.4%** depth rank 2, **87.8%** rank 5, **90.2%** rank 3, and becomes top-1 only just outside the scored band at **92.7%** and later. The dominant workspace tokens before that are generic comparative/continuation tokens such as `compared`, `comparatively`, `predicted`, and `relative`, not a verbalized materials mechanism. Thus the lens suggests a late answer-direction representation, but not a clean unspoken Hall–Petch explanation; and because the clean-answer control failed, this remains qualitative only.\n\nThe exploratory candidate list is useful only as a heuristic over the slice grid, not a metric. `microstructure`, `diameter`, and `metallurgical` look plausibly meaningful domain associations near grain-size positions; `differing` and `respectively` plausibly reflect parsing the two specimens and their two sizes. `because` may indicate causal framing, but it is generic and not specific to Hall–Petch strengthening. `but`, `was`, `which`, and `will` are mostly syntax/continuation artifacts. Overall, the slice contains materials-flavored context, but the scored readout does not show a robust, ordered, single-token reasoning chain."
+    },
+    {
+      "slug": "paper-v1-schmid-constrained",
+      "title": "Constrained resolved-shear event",
+      "shape": "MULTIHOP",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "This item is **excluded**: the clean model did not produce the expected one-word answer **“slip”** and instead emitted `<turn|>`, so this cannot be counted as a clean protocol success. Within the registered workspace band, the J-lens does not show a clean emergence of the target microscopic event. The readout-position top tokens are dominated by prompt/task words early-to-mid band (“initiating”, “commencement”, “Answer”) and by generic mechanics terms late (“stress”, “strain”, “deformation”). The tracked **dislocation** concept does improve late, reaching J-rank **104 at 90.2% depth**, slightly better than the matched logit-lens rank **151**, but it has no sustained onset and is not the requested event; it is at best a related microscopic mechanism/carrier. **Schmid** is not a J-lens signal here: its best J-rank is only **4292 at 82.9%**, while the vanilla logit lens is much stronger at **391**, so the J-lens is not adding evidence for Schmid-law reasoning. There is also no visible ordered multihop chain ending in “slip”; the model appears to have materials/mechanics associations without the final answer concept surfacing.\n\nThe exploratory candidate list should be treated only as heuristic. **“threshold”** is meaningful: it corresponds to comparing the resolved 120 MPa value against the critical 80 MPa value. **“plane”** is also plausibly meaningful, since resolved shear/slip-system reasoning involves planes even though the prompt says “system” rather than “plane.” **“metallurgical”** is a weak but domain-relevant contextual signal. **“coefficient”** may reflect the orientation-cosine/product computation, but it is not evidence of the correct microscopic event. **“phenomenon”** is mostly a generic paraphrase of “microscopic deformation event,” not a specific solved concept. **“when,” “because,” “which,” “also,”** and likely **“corresponding”** are discourse/relational tokens or prompt-continuation artifacts, not substantive reasoning states. The slice grid contains scattered materials fragments such as “microsco,” “deformat,” and “dislocat,” but these look like generic/subword or neighboring-position associations; they should not be promoted into a metric, and they do not rescue the absence of a clean **slip** signal."
+    },
+    {
+      "slug": "paper-v1-fcc-constrained",
+      "title": "Constrained diffraction recognition",
+      "shape": "RECOGNITION",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "This item is **excluded**, so it is not a registered protocol success: the prompt already contains **“cubic”**, and the model’s clean greedy completion was just `'\\n\\n'`, not the expected first token **FCC**. Within the fixed workspace band, the clearest readout is not an unspoken derivation of **FCC**, but a late activation of a broad **cubic/Bravais-lattice answer space**. Around 78.0–92.7% depth the top tokens become **NaCl / isotropic / BCC / FCC / lattice**, then **lattice, cubic, BCC, FCC**, and by the final layers **BCC** dominates while **FCC** remains only rank 2–4 in the displayed top tokens. That is qualitatively consistent with the model recognizing “this is a cubic lattice abbreviation question,” but not with a clean internal computation of “all odd or all even ⇒ FCC.” The tracked concepts reinforce this: **cubic** reaches J-lens rank 1 at 85.4%, but it is in the input and the vanilla logit lens is also rank 1, so it is not a distinctive J-lens discovery. **Diffraction** is only moderate, J rank 180 at 85.4%, and the logit lens is better at rank 149. **Extinction** is essentially absent, J rank 33011 versus logit rank 27160, with no sustained onset. Thus the lens mostly shows a late, somewhat confused lattice-label association, with **BCC incorrectly preferred over FCC**.\n\nThe layer×position grid has plausible qualitative local structure—tokens like **parity**, **symmetry**, **exclusion**, **Miller**, **abbreviation**, **Brav**, and **lattice** appear in relevant columns—but that is an all-position visual pattern, not the registered metric. It suggests the model is representing parts of the prompt semantics, especially “Miller-index parity rule” and “standard Bravais abbreviation,” but the readout position does not show a clean ordered reasoning chain from reflection condition to face-centering. The absence of **extinction/systematic absence** language is notable: if the model had explicitly formed the standard crystallographic rule, one might expect that concept to surface, but here it does not.\n\nFor the exploratory candidate list, several entries look meaningful only as local prompt semantics: **nomenclature** near “standard” reflects the abbreviation/naming instruction; **combinations** near “mixed” and **integers** near “all” plausibly reflect the parity/index condition; **diffraction** near “reflections” is a domain synonym, though not stronger than logit at the answer readout. **Simultaneously** may weakly reflect the “all three indices share parity” constraint, but it is generic. **Description**, **because**, and **This** are mostly discourse/instructional artifacts, and **was** is likely grammatical noise. **Polynomial** near “cubic” is a tokenizer/semantic-neighborhood artifact from the mathematical sense of “cubic,” not evidence of crystallographic reasoning. Overall, the exploratory list supports shallow recognition of the prompt’s vocabulary, not a reliable hidden computation of **FCC**."
+    },
+    {
+      "slug": "paper-v1-griffith-constrained",
+      "title": "Constrained flaw-size scaling",
+      "shape": "MULTIHOP",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "This item is **protocol-excluded**: the clean model did not produce the expected first-token answer “lower” and instead greedily produced “**”, so this cannot be counted as a clean J-lens success. Within the fixed workspace band, there is a real but limited materials/fracture signal: **fracture** reaches J-lens rank **101 at 41.5% depth**, much stronger than the matched vanilla logit-lens rank **5632**. That is a meaningful mechanism-adjacent concept, but it is not the registered criterion **Griffith**. **Griffith** itself is weak in the J-lens, best rank only **4746 at 75.6%**, and the vanilla logit lens is actually better there (**1803**), so there is no J-lens-specific evidence that the model represented the Griffith criterion. **crack** improves over the logit baseline (**4136 at 87.8% vs. 12680**) but remains low-ranked and has no sustained onset. The late readout also drifts toward the wrong comparative: “higher” becomes dominant from about **85–100% depth**, while “lower” appears only below it, consistent with the failed clean-answer control rather than with correct inverse-square-root flaw-size scaling.\n\nThe exploratory top-1 list should be treated only as a heuristic, not a metric. **elasticity** is plausibly meaningful but mostly prompt-adjacent, reflecting “elastic plates” or modulus rather than an unspoken inference. **larger** and **compared** are also plausible scaffolding concepts: they track the comparison between the longer/larger flaw and the other plate, but they do not show the Griffith mechanism or the inverse-square-root scaling result. **thickness** and **density** look like generic materials-property intrusions; they are scientific-sounding but irrelevant to the stated criterion and likely small-model/topic-noise artifacts. **was**, **whereas**, **which**, **because**, and **will** are discourse or grammar continuations rather than substantive reasoning states. Overall, the lens shows a localized fracture-domain activation and comparative framing, but not an ordered emergence of “Griffith → lower stress”; the named mechanism is effectively absent in the J-lens, and the final disposition appears wrong."
+    },
+    {
+      "slug": "paper-v1-pileup-focus-key",
+      "title": "Pile-up focus replication: brass key",
+      "shape": "MODULATION",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "Within the registered workspace band (38–92% depth), this is only a weak directed-modulation signal. The best J-lens ranks are good for **pile** and **stress**—**pile** reaches rank 14 at 73.2% depth, and **stress** reaches rank 18 at 87.8%—but neither has a sustained onset, and neither beats the matched vanilla logit lens: **pile** is tied with the logit lens at rank 14, while **stress** is worse than the logit lens rank 11. **Dislocation** is not cleanly represented by the J-lens here: its best J rank is only 1563 at 78.0%, and the logit lens is stronger at rank 596. So the chart is consistent with some late availability of the prompted focus concepts, with **pile** preceding **stress**, but it is not a clean J-lens-specific ordered-emergence result, and the key term **dislocation** is essentially absent as a strong verbalizable concept.\n\nThe layer-position grid qualitatively shows metallurgy/focus-related fragments around the instruction region—e.g. **dislocat**, **pile**, **stress**, **tension**, **effects**, **concentr**—but this should not be upgraded into the registered metric. It is useful as a visual sanity check that the model has some local access to the focus instruction while producing the unrelated carrier sentence, but the scored readout does not show a strong advantage over logit lens. Since this is a directed-modulation item, it also should be interpreted only relative to matched suppress and neutral records; those comparisons are not present here, so no modulation claim can be made from this record alone.\n\nFor the exploratory candidates, most are not meaningful evidence of unspoken dislocation-pile-up reasoning. **because**, **but**, **also/Also**, and **was** look like generic discourse or continuation artifacts. **describing** is likely meta-instructional writing/sentence context. **amidst** is a literary carrier-style substitution for spatial wording, not the target concept. **shimmering** may reflect a material/image association with **brass**, but not the pile-up focus; it may also be a small-model/tokenization-style artifact. The only plausibly meaningful heuristic entries are **effects** and perhaps **implications**, because they fit the stress-concentration/consequence frame and appear near the stress-related region, but they remain qualitative and unregistered. Overall: some late, weak focus leakage for **pile** and **stress**, no strong **dislocation** representation, and no clean J-lens-specific success."
+    },
+    {
+      "slug": "paper-v1-pileup-suppress-key",
+      "title": "Pile-up suppression replication: brass key",
+      "shape": "MODULATION",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "Within the registered fixed workspace band (38–92%), this included directed-modulation item shows a real but incomplete suppressed-concept signal. The completion cleanly omits the forbidden material-science terms, but they are in the prompt, so this is not an input-absence/clean discovery case. At the readout position, **pile** becomes a strong transient J-lens concept, reaching rank **92 at 73.2% depth**, better than the matched logit-lens rank **212**. **stress** surfaces later, reaching rank **248 at 90.2%**, much stronger than logit-lens rank **1480**. **dislocation**, however, is not recovered: its best J rank is only **6331 at 46.3%**, with the logit lens slightly better at **6069**. None has a sustained onset. So the lens suggests the model is carrying pieces of the “pile-up / stress concentration” suppression content while writing the harmless sentence, but not the full metallurgical concept; in particular, “pile” is only the single-token fragment of “pile-up,” and “dislocation” is effectively absent.\n\nThe apparent ordering, **pile** before **stress**, is qualitatively consistent with the forbidden phrase structure and with a prompted avoidance state, but I would not treat it as a clean multi-step reasoning trace. The layer-position grid also qualitatively shows suppression-clause material such as pile/stress-like tokens in the band, but that all-position pattern is not the registered metric. The exploratory candidate list is mostly heuristic noise or task-frame signal: **also/Also**, **but**, and **because** are generic discourse-token artifacts, with capitalization duplication; **was** is likely grammar/tense carryover; **writing** plausibly reflects the active “write the sentence” task; **whatsoever** plausibly reflects the “do not write anything else” constraint; **implications** may be semantically related to consequences/stress concentration but is not registered evidence; **weird** and **mysterious** may reflect the odd “try not to think” framing, but they are weak qualitative artifacts rather than substantive hidden reasoning. Overall: a genuine transient J-lens signal for unspoken **pile** and **stress**, but no clean representation of **dislocation pile-up** as a full concept."
+    },
+    {
+      "slug": "paper-v1-pileup-control-key",
+      "title": "Pile-up neutral control: brass key",
+      "shape": "MODULATION",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "Within the registered workspace band (38–92% depth), this neutral control does **not** show a clean pile-up/materials reasoning signal. The tracked concept **pile** briefly becomes fairly strong in the J-lens, peaking at rank **53** around **70.7% depth**, better than the matched vanilla logit-lens rank **225**. But there is **no sustained onset**, and the top-token slice at that layer is dominated by unrelated multilingual/copying artifacts rather than a coherent “pile-up” representation. **Stress** is also somewhat J-lens-favored, rank **362** at **90.2%** versus logit rank **2069**, but this is late and unsustained. **Dislocation** is not supported: J-lens rank **6685** is worse than the logit-lens rank **3410**. Overall, the lens mostly reflects instruction following and exact copying of “The brass key rested under the folded map,” not unspoken mechanical reasoning.\n\nThe exploratory candidate list should be treated as qualitative only. **“verbatim”** and **“whatsoever”** look meaningful as instruction-following paraphrases of “Do not write anything else,” especially because they appear near the negative-instruction positions; they are not evidence for the registered pile-up concepts. **“tucked”** may be a genuine semantic paraphrase of “rested under the folded map,” but it is a descriptive/story-word association, not a pile-up or stress computation. **“was”** is a generic narrative-tense alternative and only weakly meaningful. **“also/Also,” “and,” “because,” “other,”** and especially **“masterpiece”** look like discourse/default-continuation or tokenizer/model-noise artifacts in this context. Thus the neutral control behaves as it should: some copy/instruction and sentence-semantics structure is visible, but no clean registered emergence of pile-up reasoning."
+    },
+    {
+      "slug": "paper-v1-pileup-focus-cup",
+      "title": "Pile-up focus replication: white cup",
+      "shape": "MODULATION",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "Within the registered workspace band, this record gives a selective signal for **pile**, but not a clean three-concept “dislocation pile-up / stress” workspace trace. **Pile** reaches J-lens rank **13 at 78.0% depth**, substantially stronger than the matched vanilla logit-lens rank **92**, so that is the main genuine J-lens finding here. By contrast, **dislocation** peaks only at J rank **388 at 85.4%** while the logit lens is better at **124**, and **stress** peaks at J rank **136 at 82.9%** while the logit lens is also better at **36**. None has a sustained onset. The chart’s late/final-layer improvement should not be promoted into the metric outside the fixed 38–92% band. Since these concepts are present in the instruction but absent from the output sentence, the lens is best read as showing some retained, unspoken task-conditioning around **pile-up**, not as a clean ordered reasoning sequence or a full material-mechanics representation. The directed-modulation claim still needs the matched suppress/neutral records; from this focus item alone we should not infer a replication effect.\n\nThe exploratory candidate list is useful only as a heuristic and mostly looks like carrier/style/top-token noise. **but**, **also/Also**, **because**, and **was** are generic discourse or syntactic continuations, not meaningful hidden concepts. **grammat** is a tokenizer fragment, though plausibly tied to the “write the sentence” control state rather than the materials rationale. **masterpiece** looks like literary-style noise. **amidst** may reflect scenic paraphrasing of “beside,” but not a registered concept. **padlock** near “closed” and **motionless** near “stood” are semantically interpretable local associations from the visible sentence, not evidence of dislocation reasoning. Overall, the record contains one notable J-lens-specific concept, **pile**, while the other tracked concepts and most exploratory tokens are either better captured by the logit lens or are local lexical/style artifacts."
+    },
+    {
+      "slug": "paper-v1-pileup-suppress-cup",
+      "title": "Pile-up suppression replication: white cup",
+      "shape": "MODULATION",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "This is a registered included directed-modulation item, and the completion used for output-absence checking contains only “A white cup stood beside the closed window.” So the target materials concepts are unspoken in the output, but they are explicitly present in the prompt; this is not a clean absent-concept discovery, only a suppression/negation readout. Within the fixed 38–92% workspace band, the J-lens does show the forbidden content being represented despite the instruction not to think about it. The strongest signal is **pile**, reaching J-rank **119** at **53.7%** depth versus matched logit-lens rank **405**. **Dislocation** is weaker but still J-favored, J-rank **2289** at **53.7%** versus logit **5113**. **Stress** is also weak-to-moderate, J-rank **1210** at **90.2%** versus logit **2148**. None has a sustained onset, so I would call this a transient suppression/mention trace rather than a stable workspace representation. Also, “pile” is a single-token piece of the hyphenated **pile-up** phrase, so the clearest signal is partly a tokenizer-friendly fragment of the forbidden phrase, not necessarily a robust representation of the full dislocation-pile-up mechanism.\n\nThe slice grid is consistent with this: around the suppress clause there are mid-depth cells containing **dislocat**, **pile**, **stress**, **concentr**, **physics**, and related fragments, while later layers increasingly return to the literal writing/copying task and then next-token behavior. There is no ordered multi-step reasoning to infer here; the model is not computing a materials-science chain, it is parsing a negative instruction that names the forbidden concepts and then producing the neutral sentence. The J-lens is nonetheless more sensitive than the vanilla logit lens for all three tracked terms, especially **pile**, so the result is a genuine but limited modulation signal.\n\nThe exploratory candidate list should be treated only as a heuristic. **writing** and **grammat** mostly reflect the task frame (“Write…”, “write the sentence”) and in the case of **grammat** likely a subword/stem artifact rather than a clean concept. **but**, **also/Also**, and **because** are generic discourse-control tokens and not meaningful evidence of hidden reasoning. **stillness** and **padlock** may be semantic associations to the carrier sentence, especially “closed window,” but they are surface/scene associations, not target materials concepts. **implications** is plausibly related to the “stress it concentrates” clause, but it is diffuse and unregistered. **masterpiece** looks like stylistic/noise spillover. None of these exploratory entries should be counted as a protocol success."
+    },
+    {
+      "slug": "paper-v1-pileup-control-cup",
+      "title": "Pile-up neutral control: white cup",
+      "shape": "MODULATION",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "This neutral control does **not** show clean unspoken pile-up/material reasoning. Within the registered workspace band, the best signal is a narrow **pile** blip at **53.7% depth**, where the J-lens reaches **rank 87** versus **logit-lens rank 207**. That is genuinely stronger than the matched logit baseline, but it has **no sustained onset** and does not form an ordered or persistent trajectory; in a neutral “white cup” copy task, I would treat it as a weak single-token association/artifact rather than evidence that the model computed “pile-up.” **Dislocation** and **stress** are weaker: **dislocation** reaches J rank **1730** at 53.7% versus logit **2180**, and **stress** reaches J rank **1590** at 85.4% versus logit **2664**. These are modest J-over-logit improvements but still low-ranked and unsustained. The late layers mostly move toward instruction/output-control tokens such as “Write,” “Period,” and “COPY,” consistent with literal next-token/task representation rather than hidden physical reasoning.\n\nThe exploratory candidate list is best read as a qualitative slice, not a metric. Some entries are plausibly meaningful local associations: **crockery** near “cup” is a reasonable semantic associate; **motionless** and **stillness** relate to “stood”; **padlock** relates to “closed”; **verbatim** and **whatsoever** likely reflect the instruction “Do not write anything else.” These show that the workspace contains paraphrastic or scene-related content, but not pile-up mechanics. By contrast, **also**, **and**, and **because** are mostly discourse/function-token noise, and **masterpiece** looks like a generic writing/literary continuation artifact rather than a controlled concept. Overall, the lens reveals ordinary copying/instruction and scene-association content, with one isolated **pile** token blip that should not be counted as a clean protocol success."
+    }
+  ],
+  "synthesis": "Across this exploratory materials run, the clearest workspace-like phenomena appeared in **ASSOCIATION** items, but they were partial rather than full replications of the paper’s strongest claims. The ductile-fracture vignette was the strongest included example: **coalescence** surfaced in the workspace band at J-rank **13 at 41.5% depth** versus logit-rank **3048**, with weaker but real support for **ductile** and **nucleation**. The cyclic-fracture vignette also had a meaningful J-specific signal for **propagation** at rank **337 vs. 2228**, and the boundary-attack vignette showed suggestive improvements for **boundary** and **sensitization**, though both remained around rank 10³ and were unsustained. These results are consistent with the paper’s claim that unspoken, verbalizable intermediate concepts can appear in the J-lens before the model’s answer, but only in a limited sense here: the signals were transient, often broad, and did not show clean sustained workspace plateaus. The low-temperature brittle-fracture item had a plausible **brittle** signal, but it was excluded from registered interpretation because **fracture** was already in the input.\n\nThe **MULTIHOP** and **RECOGNITION** items did not provide registered support for ordered internal reasoning. All listed multihop items were excluded by clean-answer failures, and the tracked mechanisms were weak or worse than the vanilla logit lens: **Hall** was much better in the logit lens than the J-lens, **Schmid** likewise, and **Griffith** was effectively absent as a J-specific concept. Some adjacent materials terms appeared, such as **fracture** in the flaw-size item and late **dislocation** in the shear item, but these did not form ordered chains such as “Hall–Petch → strengthening → higher,” “Schmid → slip,” or “Griffith → lower.” The diffraction recognition item was also excluded because **cubic** was in the input and the model did not produce the expected **FCC** answer; qualitatively, the lens showed a confused cubic/BCC/FCC answer space rather than a clean derivation of FCC from reflection parity. Thus the paper’s ordered-emergence claim was not demonstrated in this run.\n\nThe **MODULATION** records showed some qualitative suppression/focus leakage, especially for the single token **pile**, and sometimes for **stress**, but not a clean registered modulation result. In several focus or suppress records, **pile** reached strong J-lens ranks, for example rank **13** in the white-cup focus item and rank **119** in the white-cup suppress item, often beating the logit baseline. **Stress** was sometimes J-favored in suppress records, but in focus records it was often equal to or worse than the logit lens, and **dislocation** was generally weak or better captured by the logit lens. Neutral controls also produced isolated **pile** blips, so the modulation evidence should be read cautiously. Because these groups are explicitly sample-insufficient, and because the concepts were named in the prompt rather than absent from both input and output, these are best described as transient traces of prompted or forbidden content, not clean demonstrations of hidden mechanistic thought. No **REPORT_SWAP** analyses were provided in this run.\n\nThe run strongly illustrates the paper’s stated **single-token and model-capability limitations**. Important materials concepts were dropped or weakened because they are not clean single vocabulary tokens: **intergranular**, **transgranular**, **striations**, and **Petch** could not be registered, while **pile** is only a tokenizer-friendly fragment of **pile-up**. Some failures also look like genuine model limitations rather than lens failures: this target model did not reliably solve Hall–Petch, Griffith scaling, resolved-shear slip, or FCC recognition, and the J-lens mostly reported that absence or confusion. That is scientifically useful, but it means google/gemma-4-E4B-it is better suited here for exploratory demonstrations of vocabulary-level associations than for strong interpretability claims about multi-step materials reasoning. A larger or more capable model would be the more appropriate target for testing the paper’s global-workspace claims in this domain.\n\nThe exploratory surfaced candidates should be treated only as a heuristic, not as a paper metric. Plausible but broad signals include **metallurgical** as a domain-frame token and **indicating** as explanatory/diagnostic framing; **whatsoever** is meaningful mainly in modulation prompts as an instruction-following or “do not write anything else” trace. By contrast, the most frequent candidates—**because**, **also/Also**, **was**, **but**, **which**, **This**, and **and**—are mostly discourse, syntax, or continuation artifacts. **masterpiece** looks like stylistic or small-model noise rather than a materials concept. These candidates can help guide qualitative inspection, but they should not be counted as registered workspace successes or used to infer rates from this small exploratory run."
+}

jlens_materials/runs/gemma4-materials-qual_analysis.json

@@ -0,0 +1,81 @@
+{
+  "model": "google/gemma-4-E4B-it",
+  "provider": "openai",
+  "analysis_model": "gpt-5.5",
+  "offline": false,
+  "per_prompt": [
+    {
+      "slug": "materials-hall-petch",
+      "title": "Grain refinement to yield strength",
+      "shape": "MULTIHOP",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "This is **not a registered protocol success**: the clean-answer control failed because the model’s greedy first token was **“significantly”**, not the expected **“higher”**. Qualitatively, however, the readout looks like the model is moving toward the phrase **“significantly higher”**: in the readout-position top tokens, the mid/late band first carries materials-property content around depths **39–51%** (“resistance,” “metallurgical,” “toughness,” “tensile,” “hardening”), then causal/transition tokens around **53–61%** (“will,” “because,” “due,” “accordingly”), then comparative-answer tokens from about **66–88%** (“more,” “stronger,” “greater,” and **“higher”**, with “higher” top-ranked at **80.5–87.8%**). The last layers then switch to the literal next token **“significantly”** from about **90% onward**, matching the actual greedy output. So there is a real qualitative answer-emergence pattern, but it is excluded from the clean metric.\n\nFor the named law, the evidence is weak. **Hall** does not cleanly surface: its best recorded J-lens rank in the fixed band is only **9005 at 56.1%**, versus a matched logit-lens rank of **11294**. That is technically better than the logit baseline, but still extremely low and not sustained. **Petch** was dropped as multi-token/untrackable, which is exactly the single-token limitation of this protocol. **Strengthening** is the one more meaningful tracked signal: it reaches J-lens rank **479 at 48.8%**, materially better than the matched logit rank **2726**, and it occurs in the same mid-band region where the top tokens are metallurgical/material-property words. But there is no sustained onset, and “strengthening” is a generic concept rather than the specific **Hall–Petch** law. The lens therefore suggests the model has some generic grain-size/material-strength association, not a clean unspoken representation of the named law.\n\nThe exploratory candidate list is useful only heuristically. **“microstructure”** near “crystall-,” **“differing”** near “have,” **“sizes”** near “ite,” and perhaps **“respectively”** near “and” look semantically meaningful as reconstruction of the sample comparison, though “sizes” is partly a tokenizer/local-context artifact. **“because”** and **“Consequently”** reflect causal-discourse structure, not the scientific mechanism. **“toughness”** is a plausible materials-domain activation but not the correct target concept and may reflect a small-model/property-neighbor artifact. **“was,” “are,” and “which”** are basically syntactic/noise candidates. The grid also shows some distracting transport/electrical vocabulary around “carrier motion,” so the model may be mixing “carriers” with conductivity-style concepts rather than cleanly representing dislocation/grain-boundary Hall–Petch reasoning."
+    },
+    {
+      "slug": "materials-orowan-spacing",
+      "title": "Obstacle spacing to bypass stress",
+      "shape": "MULTIHOP",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "This is a clean included item for the answer direction: the model greedily says “increase.” However, the J-lens does **not** cleanly reveal the intended specific bypass mechanism, because **“Orowan” was dropped as multi-token and is therefore not trackable** in this protocol. Among tracked substitutes, **strengthening** is the only reasonably strong concept, reaching J-rank **208 at 82.9% depth**, but the matched vanilla logit lens is similar (**281**), so this is not a distinctive J-lens discovery. **Bowing** is weak in the J-lens, rank **3844 at 78.0%**, and the logit lens is actually better (**749**). **Precipitation** improves over the logit baseline, J-rank **5291 vs. logit 53852**, but it is still not a strong or sustained surfaced concept. Thus the main registered signal is generic “strengthening/increase,” not a clean latent “Orowan bypass” mechanism.\n\nThe layer grid has a plausible qualitative materials-mechanics texture in the workspace band: tokens such as “metallur/metallurgical,” “density,” “spacing,” “curvatur,” “bends,” “stiffer,” “pressure,” and “resistan” appear around the prompt positions describing particles, spacing, bowing, and required stress. But much of this mirrors or paraphrases the input rather than showing an unspoken intermediate. At the readout position, the top tokens transition from vague reasoning/discourse tokens to “therefore,” and then to the literal answer **“increase”** by about **80.5% depth**, so the late layers look more like answer commitment than an ordered emergence of a hidden mechanism.\n\nThe exploratory candidate list should be treated only as a heuristic. **“metallurgical”** is probably meaningful as a domain label; **“density,” “percentage,” and “distance”** are partly meaningful paraphrases of volume fraction/spacing but are close to input-derived restatements rather than new reasoning. **“because”** may reflect explanatory framing, not a scientific concept. **“was,” “this,” “and,” “they,” and “became”** are mostly function-word/coreference/discourse artifacts. Overall, this item shows the model knows the answer direction and some generic strengthening context, but it does **not** provide a clean J-lens success for the specific Orowan-bypass intermediate."
+    },
+    {
+      "slug": "materials-martensite-quench",
+      "title": "Rapid quench to martensite",
+      "shape": "MULTIHOP",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "This item is **excluded as a registered protocol success** because the clean model did not answer “tetragonal”; its greedy token was **“lower”**. Qualitatively, the J-lens does show that the model has some relevant crystallographic content before the final token: at the readout position, **“tetragonal”** enters the top tokens late in the workspace, rank 6 at 80.5%, rank 2 at 82.9–85.4%, rank 3 at 87.8%, and rank 1 at 90.2%, before the final layers pivot toward **“lower”**. It also strongly represents deformation language such as **“distorted”**, and earlier workspace tokens include **anisotropic/isotropy/crystallography/morphology**. So the lens suggests a partially correct internal description of an elongated, distorted cell, but not a clean answer pathway.\n\nThe registered mechanism concepts do **not** provide a clean ordered multihop signal. **Shear** reaches only J-rank 512 at 82.9%, essentially the same as the matched logit-lens rank 538, with no sustained onset. **Bain** is weaker in the J-lens: J-rank 3516 at 87.8%, while the vanilla logit lens is better at rank 829. Thus there is no convincing J-lens advantage for the intended transformation-mechanism intermediates, and no ordered emergence of “shear” then “Bain” before the product/symmetry. **Martensite** is correctly dropped from the registered tracking because it is multi-token in this tokenizer; the visible **“martens”** fragment is at most a suggestive subword/phase-prefix artifact, not evidence that the single concept “martensite” was cleanly tracked.\n\nThe exploratory candidate list should be treated only as a heuristic. **Metallurgical**, **microstructure**, **structure**, and **ratio** look meaningfully related to the prompt context; **toughness** is domain-relevant but likely cued by “hard” rather than an unspoken reasoning step. **Martens** is interesting but tokenizer-limited and not a valid registered concept. **Because**, **was**, **which**, and **This** are mostly syntactic/discourse tokens and should not be interpreted as scientific intermediates. Overall, this is a qualitative small-model pattern: the J-lens exposes relevant crystallography and a late transient **tetragonal** representation, but it does not validate the intended mechanism-chain hypothesis for this prompt."
+    },
+    {
+      "slug": "materials-schmid-slip",
+      "title": "Resolved stress to slip",
+      "shape": "MULTIHOP",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "This item is **excluded as a registered success** because the clean-answer control failed: the model’s greedy next token was **“yield”**, not the expected **“slip.”** Qualitatively, the J-lens still shows a real materials concept: **dislocation** becomes strong in the workspace band, with sustained onset around **41.5% depth** and **rank 1 at 43.9%**, substantially beating the matched vanilla logit lens (**rank 17**). The layer slice agrees: around 41.5–51.2%, the readout position is dominated by “dislocation”/“Dislocations”/“dislocations,” before later layers move toward deformation-event tokens such as **undergo**, **deform**, **deformation**, **slip**, and finally **yield**. This is a plausible unspoken physical-carrier representation, but because the model ultimately answers “yield,” it should be treated as a qualitative pattern, not a clean protocol success.\n\nBy contrast, **Schmid** does **not** surface as a law-level intermediate in the J-lens. Its best reported J-lens rank is only **808 at 87.8% depth**, with no sustained onset, and the vanilla logit lens is actually better (**rank 145**). That is the opposite of the strongest J-lens evidence: the named criterion is not clearly represented/verbalizable here, even though the prompt’s calculation is essentially a Schmid-factor/resolved-shear-stress setup. The late trajectory mostly reflects convergence toward answer-like deformation tokens: “deform” dominates around 78–90%, “slip” reaches top ranks around the band edge/just after it, and the final layers switch to the literal model output **“yield.”**\n\nThe exploratory candidate list should be read only as a heuristic. **“threshold”** is plausibly meaningful, since the prompt compares 120 MPa to a critical value of 80 MPa. **“metallurgical”** is a broad domain tag and consistent with the slice, but not a specific reasoning step. **“equals,” “zero,” “because,”** and **“for”** look mostly like arithmetic/connective or local-token continuations rather than concepts. **“hydrostatic”** is likely a domain-association error: the prompt is uniaxial/resolved shear, not hydrostatic stress. **“which,” “where,”** and **“was”** are generic positional/top-1 artifacts and should not be interpreted as reasoning content."
+    },
+    {
+      "slug": "materials-nabarro-herring",
+      "title": "High-temperature deformation mechanism",
+      "shape": "MULTIHOP",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "This is an included, clean item: the model’s greedy answer is the expected single token **“diffusion”**, and the tracked unspoken mechanism tokens are absent from the generated completion. The J-lens does show a real materials-domain concept, **creep**, but it is broad rather than mechanism-specific. In the fixed workspace band, **creep** rises sharply only late, peaking at **J-rank 2 at 80.5% depth**, while the answer token **diffusion** is already top-ranked from about **75.6% depth** onward at the readout position. This is therefore not a clean “intermediate before final answer” ordering: the transport-process answer has already crystallized before or alongside the generic creep concept. Also, the matched vanilla logit lens is actually stronger for **creep** (**rank 1**), so this is not a distinctive J-lens advantage. The specific **Nabarro-Herring** mechanism is not recovered: **Nabarro** is dropped as multi-token/not trackable, and **Herring** remains weak (**J-rank 1104 at 87.8%**, with the logit lens better at **490**). So the lens suggests the model has a broad “diffusion/creep” interpretation, but not a clearly verbalized single-token representation of **Nabarro-Herring creep**.\n\nThe exploratory candidates should be treated only as qualitative hints. **microstructure** is plausibly meaningful as a materials-domain abstraction from “fine-grained polycrystal,” but it is a broad context cue, not the target reasoning step. **deformation** is also meaningful, compressing “slow shape change,” and **metallurgical** is a plausible domain cue. **mechanism** may reflect the task frame or “dominant process” framing, but it is not evidence for the named mechanism. **migratory** is mostly a lexical echo of “migrate,” not an independent concept. **analysis** looks generic/noisy, and the function words **and**, **because**, **was**, and **but** are sentence-continuation artifacts rather than scientific intermediates. Overall, the run shows a genuine but generic creep/diffusion workspace signal, with no clean registered evidence that the model represented the specific Nabarro-Herring mechanism as an unspoken intermediate."
+    },
+    {
+      "slug": "materials-screw-character",
+      "title": "Line-defect character",
+      "shape": "MULTIHOP",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "This item is excluded: the model did not give the clean answer “screw,” instead greedily continuing with “determined.” Within the fixed workspace band, the J-lens does not show a clean registered recovery of the intended materials concepts. At the readout position, “Burgers” only reaches J-rank 101 at 87.8% depth, and “dislocation” only reaches J-rank 181 at 85.4%, with no sustained onset for either. Both are also worse than the matched vanilla logit lens baselines: “Burgers” logit-rank 41 and “dislocation” logit-rank 133. The top readout tokens are mostly generic continuation/classification language—“classification,” “characterized,” “classified,” “determined,” “defined”—ending in the literal wrong next-token basin “determined.” So the lens mainly reports a model settling on “the character is determined/characterized/classified,” not an unspoken computation of “Burgers vector parallel to line direction ⇒ screw dislocation.”\n\nThere is some qualitative domain activation in the position-slice grid, but it should not be promoted to a protocol success. The exploratory “dislocation” candidate is the most meaningful: it appears as a top-1 token at some prompt positions around the mid-band, consistent with recognizing the object class from “line defect”/“crystal.” However, this does not translate into a strong final readout-position signal. “Burgers” is not comparably surfaced, and there is no ordered emergence of “Burgers” before “dislocation”; if anything, both improve only late and weakly. “Perpendicular” is also semantically notable but suspicious: it is the wrong geometric relation for the stated screw case and may reflect confusion with the edge-dislocation contrast or generic line-defect associations. “microstructure,” “metallurgical,” and “morphology” are plausible materials-domain context words, but they look diffuse and generic rather than evidence of the specific reasoning chain. “because,” “was,” “which,” “often,” and much of “represents” are continuation/scaffolding tokens or syntactic noise, not meaningful hidden concepts. Overall, this is best read as a weak-model/failed-answer case where the J-lens faithfully shows absence of the intended single-token intermediate reasoning, aside from scattered qualitative domain vocabulary."
+    },
+    {
+      "slug": "materials-fcc-extinctions",
+      "title": "Diffraction selection-rule recognition",
+      "shape": "RECOGNITION",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "This item is **excluded as a registered success** because the clean model did not answer with the expected first token (`FCC` or `face`); it greedily produced `a`. Qualitatively, the lens does show a late move toward the broad category **cubic**: `cubic` becomes top-1 at 85.4% depth and is sustained from about 82.9% through the late workspace / pre-final layers. But this is not a J-lens-specific discovery: the matched vanilla logit lens also ranks `cubic` at 1, so the signal is real but not stronger than the baseline. More importantly, **cubic is not the registered answer**; the correct recognition is specifically face-centered cubic / FCC. I do not see clean evidence that the model has the more precise **FCC / face-centered** concept here. The tracked intermediate **extinction** is only weakly present: J-lens improves over logit lens substantially at its best point, rank 6,194 vs logit rank 26,140 at 48.8%, but that is still far outside an interpretable top-rank signal and has no sustained onset. The slice grid mostly shows prompt reconstruction and parity-rule vocabulary—`indices`, `integers`, `all`, `odd`, `even`, `mixed`, `parity`, `absent`—followed late by “lattice is cubic.” That is a plausible partial recognition, not a clean unspoken reasoning chain from extinction rule to FCC.\n\nThe exploratory candidate list should be treated only as a heuristic. Some entries are plausibly meaningful local paraphrases of the rule: **combinations** fits “mixed-parity combinations,” **integers** fits Miller-index triples `h,k,l`, and **simultaneously** loosely captures the all-odd/all-even constraint. **because** may reflect causal explanation framing, but not a substantive crystallographic concept. **was**, **information**, and **for** are generic syntactic/filler tokens. **spectroscopic** looks like domain-neighbor noise from diffraction/spectra rather than the relevant concept. **geomet** is a subword/tokenizer fragment of “geometric” and should not be counted as a clean concept. **perpendicular** is at best a loose geometry association around indices/planes, not evidence for the extinction-to-FCC recognition. Overall, the lens reveals that the model becomes disposed to say **cubic**, but not that it cleanly computed **face-centered cubic / FCC**; the registered protocol therefore remains a failure/exclusion despite a qualitative late cubic pattern."
+    },
+    {
+      "slug": "materials-ductile-fractography",
+      "title": "Unnamed ductile-fracture vignette",
+      "shape": "ASSOCIATION",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "Within the registered workspace band, this is a real but transient J-lens signal for the ductile-fracture interpretation. The broad concept **ductile** reaches J-rank 193 at 41.5% depth versus logit-lens rank 38,170, and the more specific mechanism terms **coalescence** and **nucleation** reach J-ranks 383 at 39.0% and 401 at 43.9%, versus logit ranks 46,648 and 25,643. Those are substantially stronger than the matched vanilla logit lens, and the generated completion was only `'ound'`, so these are not simply next-token output leakage. However, none has a sustained onset, and the ordering is not a clean physical reasoning sequence: **coalescence** appears before **nucleation**, while **void** is weak in the J-lens, only rank 7,186 at 53.7%, and is actually better in the logit lens at rank 3,381. Since **microvoid** was dropped as multi-token, the lens cannot directly test the canonical “microvoid nucleation and coalescence” phrase. So the honest read is: the model has an unspoken association to ductile fracture and void-coalescence terminology, but this is not a clean ordered mechanistic trace.\n\nThe slice grid supports a qualitative materials/fractography interpretation, with mid-band tokens like **metallurgical**, **fracture**, **rupture**, **phase**, **particle**, **cluster**, **saucer**, **cone**, **morpholo-**, and **character-** appearing near relevant prompt positions. But much of this is local semantic echoing or subword morphology, not a registered metric. In the exploratory candidate list, **metallurgical** and **fracture** look meaningful as domain/failure-concept readouts; **saucer** is plausibly a morphology association to “cup”/cup-and-cone but surface-level; **clusters** may reflect joined pits/particles but is coarse. **indicating**, **occurred**, and perhaps **because** are generic explanatory scaffolding rather than mechanism concepts. **which**, **was**, and **during** are mostly syntactic/discourse artifacts. The late layers turning toward **ound** and similar fragments are consistent with the final-token regime, not workspace concept tracking."
+    },
+    {
+      "slug": "materials-intergranular-attack",
+      "title": "Unnamed boundary-localized degradation",
+      "shape": "ASSOCIATION",
+      "domain": "materials",
+      "model_used": "openai:gpt-5.5",
+      "analysis": "Within the registered workspace band, the clearest signal is the unspoken concept **corrosion**: the J-lens brings it to rank **41 at 41.5% depth**, versus matched vanilla logit-lens rank **8955**, so this is a real J-lens gain over baseline. **Boundary** is weaker but still meaningfully above logit baseline, reaching rank **1613 at 39.0%** versus **9629**. **Sensitization** is only a faint signal: rank **5974 at 41.5%** versus **18127** is an improvement, but not a clean retrieval of the mechanism. None of the tracked concepts has a sustained onset, and the exact target phrase **intergranular corrosion** cannot be scored because **intergranular** was dropped as multi-token. Qualitatively, the grid shows metallurgy/corrosion/boundary/interior/unaffected vocabulary at relevant positions, but the per-readout top tokens also include off-domain degradation words such as **necrosis/pathogen/pathology**, so this is not a crisp top-token verbalization of the named degradation mode. The fair read is: the model has a strong generic **corrosion** association and some **boundary-localized** structure, but only weak evidence for the more specific **sensitization / intergranular** diagnosis.\n\nThe exploratory candidates are useful only as a heuristic. **metallurgical** looks meaningful as a domain scaffold around “austenitic stainless steel,” but it is not the target degradation concept. **corrosion** is the most meaningful candidate and agrees with the registered rank signal. **unaffected** and **remained** plausibly reflect the “grain interiors remain largely intact” contrast, though **remained** is close to an input morphological variant rather than a new concept. **environment** is a reasonable abstraction of “aggressive solution,” but again generic. **during** is mostly temporal/syntactic, tied to “held,” not a degradation concept. **and**, **but**, **these** are function-word/discourse artifacts; **because** may weakly reflect causal framing, but should not be counted as conceptual evidence. Overall, this item is an included protocol signal for unspoken **corrosion**, with partial boundary localization, but not a clean registered success for the full specialized concept **intergranular corrosion due to sensitization**."
+    }
+  ],
+  "synthesis": "Across these materials prompts, the clearest workspace-like phenomena appeared in the **ASSOCIATION** items, but only as item-level qualitative/registered-band signals, not as group-level evidence because the run is exploratory and the ASSOCIATION group is sample-insufficient. The strongest J-lens-over-logit effects were for unspoken domain concepts in the unnamed vignettes: in the ductile-fracture item, **ductile** reached J-rank 193 versus logit-rank 38,170, **coalescence** reached 383 versus 46,648, and **nucleation** reached 401 versus 25,643, though the ordering was not mechanistic and **void** was weak. In the boundary-localized degradation item, **corrosion** was the cleanest signal, reaching J-rank 41 at 41.5% depth versus logit-rank 8,955, with weaker but meaningful **boundary** evidence at rank 1,613 versus 9,629. These are the best examples in this run of unspoken scientific concepts surfacing in the fixed workspace band before being literally output, but they remain broad associations rather than clean reconstructions of full named mechanisms such as microvoid nucleation/coalescence or intergranular corrosion by sensitization.\n\nThe **MULTIHOP** items showed more mixed, mostly qualitative behavior. Several prompts were excluded from registered success because the model did not give the expected clean answer: grain refinement produced **“significantly”** rather than “higher,” rapid quench produced **“lower”** rather than “tetragonal,” resolved stress produced **“yield”** rather than “slip,” and line-defect character produced **“determined”** rather than “screw.” Some of these excluded items still had real-looking workspace content: grain refinement moved from materials-property words around 39–51% depth to causal tokens around 53–61% and comparative-answer tokens around 66–88%, with **higher** top-ranked at 80.5–87.8% before the final-layer pivot to **significantly**; rapid quench showed late **tetragonal** ranks 6→2→1 around 80.5��90.2% before the wrong final token; and resolved stress had a strong excluded **dislocation** signal, rank 1 at 43.9% versus logit rank 17. But the registered multihop mechanism claims were not cleanly supported. **Schmid**, **Bain**, **Burgers**, **bowing**, and **Herring** were weak or better in the vanilla logit lens, and in the included high-temperature deformation case **diffusion** was already top-ranked before or alongside **creep**, while **creep** itself was stronger in the logit lens. Thus the run gives some evidence for generic materials concepts and answer commitment in the workspace band, but not for reliable ordered emergence of specific multistep scientific intermediates.\n\nThe **RECOGNITION** item was also excluded: the model answered **“a”** rather than `FCC` or `face`. The lens did reveal a late disposition to say **cubic**, with `cubic` top-ranked around 85.4% and sustained late, but the vanilla logit lens also ranked it first, and **cubic** is not the registered target concept. The specific face-centered-cubic / FCC recognition was not cleanly present, and **extinction** remained far outside a strong interpretable rank despite improving over the logit baseline. There were no supplied **MODULATION** or **REPORT_SWAP** items in this materials run, so this synthesis cannot say whether this target model exhibits those phenomena. Relative to the paper’s claims, the run partially matches the broad observation that the mid-to-late band contains verbalizable content and that the last layers pivot to the literal next token, but it does not reproduce the stronger multihop claim that intermediates appear cleanly in the order computed.\n\nThe single-token limitation was a major constraint here. Many of the most scientifically precise targets were dropped or degraded because they are multi-token or subword concepts in this tokenizer: **Petch**, **Orowan**, **martensite**, **Nabarro**, **microvoid**, and **intergranular** could not be cleanly scored, and visible fragments such as **martens** should not be treated as registered concept recovery. The model-capability limitation also matters: this small target often failed the clean answer controls and often surfaced only generic neighbors such as **strengthening**, **creep**, **cubic**, **metallurgical**, or **microstructure**, while the named law or mechanism was absent or no better than the logit lens. That is consistent with the paper’s caution that the lens can only reveal concepts the model actually has; here it often faithfully reports partial domain association rather than expert mechanism-level reasoning.\n\nThe exploratory surfaced candidates should be read only as a heuristic, not as a paper metric. Among the frequent candidates, **metallurgical** (×7) and **microstructure** (×5) are plausible broad materials-domain signals, and **threshold** (×1) was plausibly meaningful in the resolved-stress prompt. **respectively** (×1) may reflect sample-comparison structure in the grain-refinement case, but it is not a scientific mechanism. **perpendicular** (×2) is semantically related to geometry but suspicious because it was sometimes the wrong contrast, especially for the screw-dislocation case. **during** (×2) is mostly temporal scaffolding. The very frequent **because** (×9), **was** (×8), **which** (×6), **and** (×4), **represents** (×2), and **where** (×1) are best treated as discourse, syntax, or local-continuation artifacts unless supported by a separate tracked concept. Overall, this model is useful for exploratory demonstrations of broad workspace readouts in scientific text, but for scientific interpretability of named mechanisms and multistep reasoning, a larger or more capable model—and preferably phrase-level tracking beyond single-token concepts—would be a more appropriate target."
+}