manuscript_source,source_line,kind,label,caption,scope,groups,artifact_prefixes,supporting_file_count,analysis_commands,publication_assets,manifest
paper/graph_generalization_audit.tex,18,table,,Mechanism orientation in the 72-prompt audit. The physical label is generated exactly by $y=sx$.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]",[],FILE_INVENTORY.csv
paper/graph_generalization_audit.tex,44,table,,"Whole-mechanism-held-out graph learning. The numerical-direction column is an input-readback positive control because that feature was supplied to the absolute-label network. Even with direction supplied, absolute physical polarity does not generalize; the separate polarity-free relation network also fails. ROC--AUC 0.5 denotes no ordering information.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]",[],FILE_INVENTORY.csv
paper/graph_generalization_audit.tex,70,table,,Community agreement as progressively more graded similarity information is retained. The smallest corrected $q$-value over the 12 secondary representation-by-density tests is 0.092.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]",[],FILE_INVENTORY.csv
paper/graph_generalization_audit.tex,92,table,,Exhaustive partition results. Values are six-family mean ARI; $p$ is the exact blockwise-null probability.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]",[],FILE_INVENTORY.csv
paper/graph_primary_edge_inventory.tex,6,longtable,,,reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]",[],FILE_INVENTORY.csv
paper/graph_primary_edge_inventory.tex,41,longtable,,,reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]",[],FILE_INVENTORY.csv
paper/graph_primary_edge_inventory.tex,76,longtable,,,reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]",[],FILE_INVENTORY.csv
paper/graph_primary_edge_inventory.tex,111,longtable,,,reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]",[],FILE_INVENTORY.csv
paper/graph_primary_edge_inventory.tex,146,longtable,,,reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]",[],FILE_INVENTORY.csv
paper/graph_primary_edge_inventory.tex,181,longtable,,,reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens.tex,134,figure,fig:design,"\textbf{Overview of the study.} A materials-science prompt is evaluated with a frozen open-weight language model while its intermediate states are recorded. We study those states in three complementary ways: vocabulary readout tests which scientific concepts are readable; matched counterfactual comparison tests whether state changes preserve constitutive orientation; and causal intervention tests whether changing an internal representation changes the model's engineering decision. The right-hand panels illustrate the corresponding evidence types (decoded terms, relational trends, and intervention-induced answer shifts). In the causal card, arrows show the raw change in ``higher''-minus-``lower'' answer preference: refinement shifts toward ``higher'' and coarsening toward ``lower'', so the physically correct effects have opposite signs. Full experiments and statistical tests are reported in the following sections. \hlyellow{The three branches represent complementary evidence types; the diagram does not imply that one shared feature is followed across all experiments.}",study-wide summary/provenance; consult named rows,"[""readout"", ""target_free"", ""geometry"", ""physical_equivalence"", ""graphs"", ""relational"", ""steering"", ""patching"", ""causal_coordinate"", ""multitoken"", ""31b_streams""]","[""jlens_materials/experiments/answer-code-binding-2026-07-17/"", ""jlens_materials/experiments/answer-scaffold-audit-2026-07-17/"", ""jlens_materials/experiments/candidate-activation-patching-2026-07-16/"", ""jlens_materials/experiments/candidate-followups-2026-07-16/"", ""jlens_materials/experiments/candidate-nonsteering-2026-07-16/"", ""jlens_materials/experiments/causal-relational-comment3-audit-2026-09-06/"", ""jlens_materials/experiments/causal-relational-knockout-2026-09-02/"", ""jlens_materials/experiments/cross-mechanism-activation-patching-2026-07-18/"", ""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-semantic-streams-2026-08-26/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/lexical-adversarial-representation-2026-07-17/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/experiments/materials-heldout-v1_latent"", ""jlens_materials/experiments/multitoken-sequence-robustness-2026-07-18/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/"", ""jlens_materials/experiments/relational-grain-steering-v4"", ""jlens_materials/experiments/semantic-steering-v3"", ""jlens_materials/runs/""]",641,"[""python scripts/analyze_answer_code_binding.py"", ""python scripts/analyze_answer_scaffold_audit.py"", ""python scripts/analyze_candidate_followups.py"", ""python scripts/analyze_candidate_nonsteering.py"", ""python scripts/analyze_causal_relational_knockout.py"", ""python scripts/analyze_counterfactual_activation_patching.py"", ""python scripts/analyze_cross_mechanism_activation_patching.py"", ""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/analyze_late_physics_replication.py"", ""python scripts/analyze_lexical_adversarial_representation.py"", ""python scripts/analyze_materials_heldout_v1.py"", ""python scripts/analyze_materials_latent_geometry.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py"", ""python scripts/analyze_relational_grain_steering_v4.py"", ""python scripts/analyze_semantic_steering_v3.py"", ""python scripts/build_final_paper_figures.py"", ""python scripts/build_gemma31b_semantic_streams.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_multitoken_sequence_robustness.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_causal_relational_knockout.py"", ""python scripts/validate_graph_generalization_audit.py""]","[{""publication_path"": ""paper/figures/figure1-study-summary-tikz.tex"", ""sha256"": ""d67e7620934fb1b0d02903d50edbe440ca0e332bbebf37497327a4bd4560b2b8"", ""matching_archived_files"": []}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens.tex,141,table,tab:study-map,"\textbf{Study map and evidential status.} \hlyellow{Here ``frozen'' means that the listed endpoint and success rule were recorded as fixed before the corresponding output or calculation in our internal chronology; it does not establish externally timestamped preregistration or imply that every underlying state array or motivating cohort was unseen.} ``Jacobian-specific'' requires a matched advantage over direct unembedding or a direct-direction control, rather than merely a positive Jacobian result.",study-wide summary/provenance; consult named rows,"[""readout"", ""target_free"", ""geometry"", ""physical_equivalence"", ""graphs"", ""relational"", ""steering"", ""patching"", ""causal_coordinate"", ""multitoken"", ""31b_streams""]","[""jlens_materials/experiments/answer-code-binding-2026-07-17/"", ""jlens_materials/experiments/answer-scaffold-audit-2026-07-17/"", ""jlens_materials/experiments/candidate-activation-patching-2026-07-16/"", ""jlens_materials/experiments/candidate-followups-2026-07-16/"", ""jlens_materials/experiments/candidate-nonsteering-2026-07-16/"", ""jlens_materials/experiments/causal-relational-comment3-audit-2026-09-06/"", ""jlens_materials/experiments/causal-relational-knockout-2026-09-02/"", ""jlens_materials/experiments/cross-mechanism-activation-patching-2026-07-18/"", ""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-semantic-streams-2026-08-26/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/lexical-adversarial-representation-2026-07-17/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/experiments/materials-heldout-v1_latent"", ""jlens_materials/experiments/multitoken-sequence-robustness-2026-07-18/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/"", ""jlens_materials/experiments/relational-grain-steering-v4"", ""jlens_materials/experiments/semantic-steering-v3"", ""jlens_materials/runs/""]",641,"[""python scripts/analyze_answer_code_binding.py"", ""python scripts/analyze_answer_scaffold_audit.py"", ""python scripts/analyze_candidate_followups.py"", ""python scripts/analyze_candidate_nonsteering.py"", ""python scripts/analyze_causal_relational_knockout.py"", ""python scripts/analyze_counterfactual_activation_patching.py"", ""python scripts/analyze_cross_mechanism_activation_patching.py"", ""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/analyze_late_physics_replication.py"", ""python scripts/analyze_lexical_adversarial_representation.py"", ""python scripts/analyze_materials_heldout_v1.py"", ""python scripts/analyze_materials_latent_geometry.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py"", ""python scripts/analyze_relational_grain_steering_v4.py"", ""python scripts/analyze_semantic_steering_v3.py"", ""python scripts/build_final_paper_figures.py"", ""python scripts/build_gemma31b_semantic_streams.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_multitoken_sequence_robustness.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_causal_relational_knockout.py"", ""python scripts/validate_graph_generalization_audit.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens.tex,166,table,tab:families,\textbf{Complete structure of the held-out dataset.} Every row contains five independent descriptions of the listed physical evidence. The right column gives all terms declared before execution and omitted from every corresponding input. Daggers mark documented multi-token terms excluded from the one-token rank endpoint. The Supplementary Information prints all 50 exact prompts and their terms.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""readout""]","[""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/runs/""]",81,"[""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_materials_heldout_v1.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens.tex,213,figure,fig:controlled,"\textbf{Can either readout recover an engineering term that was deliberately omitted from the description?} Before running the 50 prompts, we declared 150 scientifically appropriate one-token terms and then asked where each term ranked among Gemma's 262,144 vocabulary entries. (A) The cutoff $k$ is relaxed from rank 1 to rank 100; a higher curve means more declared terms have become strongly readable. Thin teal curves are the three independent Jacobian fits, the thick teal curve is their mean, and gray is direct unembedding of the same states. At $k=100$, the fractions are 7.2\% and 3.3\%, respectively. (B) Each point is one prompt's recovery area under the curve (AUC); a point above the diagonal favors Jacobian transport. Eleven prompts improve, 36 tie, and three decline. (C) The same differences are grouped at the scientifically appropriate level: gray circles are the five phrasings and diamonds are the ten family means. Positive values favor the Jacobian lens. Boundary attack and notch resistance produce the largest gains, whereas cyclic damage and particle strengthening favor direct unembedding. Across families, the mean difference is $+0.0129$, but the hierarchical 95\% interval spans $-0.0187$ to $+0.0486$ and the one-sided family sign-flip test gives $p=0.234$. Thus the experiment finds selective, mechanism-dependent gains rather than a universal Jacobian advantage.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""readout""]","[""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/runs/""]",81,"[""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_materials_heldout_v1.py""]","[{""publication_path"": ""paper/figures/heldout-controlled-recovery.pdf"", ""sha256"": ""22627c15c08d9a34414403b533ec6cfa5bbfb9ab675b6641ac89096832233729"", ""matching_archived_files"": [""jlens_materials/figures/materials-heldout-v1/heldout-controlled-recovery.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens.tex,236,figure,fig:openvocab,"\textbf{Target-free family vocabularies.} Each row is one physical family and each cell is a word selected without the declared concept list. Support is the number of independent phrasings, out of five, in which the word appeared under the three-fit consensus rule; darker blue-green cells indicate stronger background-corrected support. Asterisks mark exact declared-word overlap added only after ranking. Read across a row to ask whether the words form a recognizable materials neighborhood. Boundary attack, surface oxidation, and rapid transformation show coherent mechanism or microstructure terms; high-temperature deformation and cleavage show why stable decoding can still be generic. The figure therefore measures what assembles naturally, not whether the algorithm can find a supplied answer.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""readout""]","[""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/runs/""]",81,"[""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_materials_heldout_v1.py""]","[{""publication_path"": ""paper/figures/heldout-open-vocabulary.pdf"", ""sha256"": ""2c81592b3b5d6e032186859a0531570a58636d31628440af6331de77570dc5a5"", ""matching_archived_files"": []}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens.tex,243,figure,fig:targetfree-classification,"\textbf{Do the freely discovered words identify the materials mechanism without a supplied answer list?} Each bar is leave-one-phrasing-out accuracy across the 50 prompts and ten balanced mechanism families. A classifier learned four phrasings per family and predicted the held-out fifth using only target-free candidate words from the corresponding readout. ``Exact prompt words removed'' deletes literal input copies; ``morphology-aware removal'' additionally deletes conservative prefix-related variants. The dotted line is 10\% chance. The dashed line is a TF-IDF classifier that instead sees the prompt words themselves. The target-free Jacobian and direct lists remain well above chance after the strictest removal, at 58\% and 64\%, but the prompt baseline reaches 56\% and direct does not differ significantly from Jacobian. Thus the discovered vocabulary contains cross-phrasing mechanism information, while the comparison rules out a lens-specific superiority claim. The complete confusion matrix and per-prompt predictions are in the Supplementary Information.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""target_free""]","[""jlens_materials/experiments/candidate-nonsteering-2026-07-16/""]",21,"[""python scripts/analyze_candidate_nonsteering.py""]","[{""publication_path"": ""paper/figures/semantic-classification-summary.pdf"", ""sha256"": ""a81e1352f536565e0c031991c023f822b7e164ba13e79e623f33d8f179c4b10c"", ""matching_archived_files"": [""jlens_materials/experiments/candidate-nonsteering-2026-07-16/figures/semantic-classification-summary.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens.tex,273,figure,fig:streams,"\textbf{What does the stricter target-free filter remove, and what materials vocabulary remains?} Rows A-E correspond to the five exact prompts printed in the text. Each left panel shows the original-style unrestricted leading words from lens fit 0. Each right panel removes input and output words, the frozen 214-word function list, and the globally common readout scaffold, then requires the word to occur in all three independent fits at that layer. No declared concept was supplied to either selection. The gray region marks the fixed 38-92\% analysis band. The comparison shows that much of the visually continuous unfiltered stream is generic discourse vocabulary, while the stricter view retains localized physical neighborhoods: \code{corrosion}/\code{damage} (A), \code{resilience}/\code{protection} (B), \code{irreversible}/\code{mechanism} (C), and \code{collapse}/\code{disintegration} (D). The generic cleavage stream (E) is a useful negative case. Gaps mean that none of the displayed words passed the strict rule at that layer; they are not evidence of absent computation. Ribbon thickness is a rank-derived visualization score, not token probability, causal influence, or a literal chain of thought.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""readout"", ""geometry""]","[""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/experiments/materials-heldout-v1_latent"", ""jlens_materials/runs/""]",81,"[""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_materials_heldout_v1.py"", ""python scripts/analyze_materials_latent_geometry.py"", ""python scripts/build_final_paper_figures.py""]","[{""publication_path"": ""paper/figures/figure4-semantic-streams.pdf"", ""sha256"": ""8fb79e9f897187beef29d19fc9ade92549ca6ec93a6384443fd755c80c244ea5"", ""matching_archived_files"": []}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens.tex,294,figure,fig:geometry,"\textbf{Do differently worded descriptions of the same mechanism remain close in the model's full state space?} (A) For each of five folds, four phrasings per family define ten centroids and the fifth phrasing is classified; the procedure is repeated at every registered layer and for every baseline representation. (B) Full-dimensional Jacobian classification peaks at 62\% at 43.9\% depth: 31 of 50 descriptions are correct, above the 13-of-50 threshold obtained after shuffling family labels and retaining the best of 25 layers ($p=0.0002$). (C) UMAP is only a display of those 50 best-layer states. Its axes have no physical units and no test uses them; color and two-letter code denote family, marker shape denotes phrasing, and stars are family centroids. (D) The full-dimensional confusion matrix reveals the physical specificity: boundary attack (BA), line-defect motion (LD), and notch resistance (NR) are separated across all five phrasings, while ductile failure (DU), high-temperature deformation (HT), and surface oxidation (OX) overlap. CL, CY, PS, and RT denote cleavage, cyclic damage, particle strengthening, and rapid transformation. (E) The same classifier reaches 54\% on raw hidden states and 76\% on a simple average of the prompt's own word embeddings. The result therefore supports statistically reliable family structure in the transported states, but it does not show that Gemma created that structure independently of the engineering vocabulary in the prompt.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""geometry""]","[""jlens_materials/experiments/materials-heldout-v1_latent"", ""jlens_materials/runs/""]",57,"[""python scripts/analyze_materials_latent_geometry.py"", ""python scripts/build_final_paper_figures.py""]","[{""publication_path"": ""paper/figures/figure5-latent-geometry.pdf"", ""sha256"": ""d2c4a08c5c59797cb6ac174455da7c40072dbd0e160f81fb35e5704cf6d0e3cd"", ""matching_archived_files"": []}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens.tex,319,figure,fig:physical-equivalence,"\textbf{Does physical equivalence overcome deliberately stronger wording similarity?} Every triplet compares an anchor with a differently worded prompt that has the same physical answer and a near-verbatim prompt whose physical relation is reversed. Positive margin means that the same-physics pair is closer; negative margin means that wording dominates. (A) In the disjoint replication cohort, both Jacobian and direct decoder-basis states begin strongly wording-aligned and rise sharply after roughly 60\% depth. The shaded region is the prospectively frozen 80-96\% late window. The frozen late-minus-middle Jacobian change is $+1.402$ ($0.951$-$1.793$), positive in 24/24 triplets and 6/6 families. (B) The absolute late-window means are positive in both the discovery and disjoint cohorts, but their family-hierarchical intervals cross zero; the replication's registered primary endpoint therefore failed. (C) Four replication families end positive, obstacle spacing is approximately neutral, and porosity remains wording-aligned, demonstrating real mechanism heterogeneity. (D) A target-free comparison of decoded word sets also trends positive late, but its intervals cross zero. Teal and purple nearly coincide throughout, showing that the transition is not uniquely produced by Jacobian transport. This is a representational similarity test, not a causal intervention or a literal chain of thought.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""physical_equivalence""]","[""jlens_materials/experiments/answer-code-binding-2026-07-17/"", ""jlens_materials/experiments/answer-scaffold-audit-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/lexical-adversarial-representation-2026-07-17/""]",65,"[""python scripts/analyze_answer_code_binding.py"", ""python scripts/analyze_answer_scaffold_audit.py"", ""python scripts/analyze_late_physics_replication.py"", ""python scripts/analyze_lexical_adversarial_representation.py""]","[{""publication_path"": ""paper/figures/late-physics-representation-replication.pdf"", ""sha256"": ""f39e8a384b2575b2ce0972b8d3e6485d85bbd57b95941fda284f6661c68d084e"", ""matching_archived_files"": [""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/figures/late-physics-representation-replication.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens.tex,338,figure,fig:relation-graph,"\textbf{Where does option-free comparative graph structure appear, and where does it fail?} Every node is one complete materials question and every selected edge points to the most similar eligible different material case. AUC uses all eligible candidates rather than only the selected edge; 0.5 denotes no ordering information. (A) The same 72 scientific stems are read at three prompt boundaries. At the natural question end, no answer word or response scaffold is present and Jacobian AUC is 0.642. A checkpoint-suffixed variant shows no detectable signal, while the explicit answer-mapping condition reaches 0.816. Because suffix and token position differ across conditions, this comparison localizes prompt-boundary sensitivity rather than a unique causal effect of the added words. Direct unembedding and raw states follow the same pattern. (B) At the natural question end, solid bars show the fraction of 144 selected edges that preserve the registered binary relation and hatched bars show all-candidate AUC. The dashed line is the exact case-preserving null mean. Jacobian selected-edge precision is 0.674 ($p=0.02195$) and AUC is 0.642 ($p=0.01235$); word and character TF-IDF are at or below the null. (C) The positive result depends on supplying the governing mechanism. When candidates come from other mechanisms, AUC falls to 0.513; when raw numerical direction deliberately conflicts with physical outcome, it falls to 0.471. (D) The actual 72-node natural-question graph contains six supplied mechanism islands and 144 selected edges. Circles, squares, and triangles are anchor, paraphrase, and near-verbatim counterfactual prompts. Filled and open nodes are the two registered orientations; teal and coral edges preserve or reverse the registered binary relation. Ninety-seven edges are teal. This figure demonstrates an option-free comparative regularity; Figure~\ref{fig:graph-identifiability} determines what that regularity can identify.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]","[{""publication_path"": ""paper/figures/relation-graph-robustness.pdf"", ""sha256"": ""1f3042ebdbd328088543cc4cf0808350acd812d2986350577aaddb38ce16ee5f"", ""matching_archived_files"": [""jlens_materials/experiments/relation-graph-visualization-2026-07-18/figures/relation-graph-robustness.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens.tex,361,figure,fig:graph-identifiability,"\textbf{What can the option-free graph identify?} (A) A numerical increase raises the property for a direct law but lowers it for an inverse law. In this prompt set, physical polarity is exactly $y=sx$; within one mechanism, the same/different physical relation collapses algebraically to the same/different numerical relation. (B) A graph isomorphism network is trained on four mechanisms, selected on a fifth, and tested on an unseen sixth. The purple AUC of 1.0 is a positive-control input readback: numerical direction $x$ was deliberately supplied to the absolute-label network, so this curve does not show graph discovery. The actual test asks whether graph structure can combine that supplied $x$ with an unseen law to infer physical polarity $y$; it fails, as does a separate same/different relation network that receives neither $x$ nor surface-variant identity. Points are held-out mechanisms and connecting lines guide the eye. (C) Community agreement rises when graded similarities are retained instead of only the strongest edge, showing that sparsification discards structure. Direct and especially raw states are at least as strong as the Jacobian states, so this organization is not lens-specific. (D) Exhaustively evaluating every balanced partition shows that the registered physical split is not the uniquely preferred division, including when counterfactual phrasing is held out. Bars are means, dots are mechanisms, and intervals show the corresponding blockwise null distribution. Together, the panels support transferable comparative geometry while rejecting a universal, Jacobian-specific physical-sign graph.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]","[{""publication_path"": ""paper/figures/graph-identifiability-summary.pdf"", ""sha256"": ""2c67031f8d824a6371c2fcd78fe844f2d3783f1c11ee985ea525784675c72338"", ""matching_archived_files"": []}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens.tex,398,figure,fig:relational-physics,"\hlyellow{\textbf{Constitutive orientation persists without a supplied equation across Gemma scale.} Every law contributes eight matched up-minus-down state contrasts along a model-specific direction fitted only on the same earlier development laws. (A) The matched 4B prompt-format comparison. Open markers supplied the equation and prescribed sign composition; filled markers used the corresponding equation-free questions. Both use the original 4B neutral scale, and black segments are medians. (B) The complete 60-law distributions for every available condition. Each waterfall independently orders all laws from most inverse to most direct; horizontal bars begin at that condition's calibrated neutral zero. The 4B supplied and equation-free facets are the matched format pair. The 12B and 31B facets contain equation-free tests only. Each facet has its own horizontal robust-$z$ scale, so signs and category ordering are comparable across facets but absolute bar lengths are not common hidden-state units. Purple, gray, and teal denote inverse, neutral, and direct relations. (C) Direct-inverse AUC for the hidden-state direction, output-logit change, and lexical controls across the same four available conditions. Hidden and output separation survives scaffold removal and reaches 1.000 at 12B and 31B, whereas word and character controls remain near chance. (D) The three registered pairwise AUCs, neutral-cut directional accuracy, and exact generated-answer accuracy for the equation-free cohort. Both larger checkpoints orient 40 of 40 directional laws and reach ceiling on all three relational AUCs; only 31B converts that representation into nearly perfect generated answers. The comparison supports a scale-robust ordinal constitutive orientation, not an invariant law-by-law magnitude, unfamiliar-law generalization, or a literal reasoning trace.}",reviewed result group(s); prefixes include related endpoints and dependencies,"[""relational""]","[""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/""]",125,"[""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py""]","[{""publication_path"": ""paper/figures/equation-free-relational-multiscale.pdf"", ""sha256"": ""4189fbe4097f242a5b7f717c097cbc4d53f70fd355dae92fbaa6544383aaa131"", ""matching_archived_files"": [""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/figures/equation-free-relational-multiscale.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens.tex,417,figure,fig:blinded,"\textbf{Can an automated interpreter recognize a materials mechanism from discovered words alone?} (A) Twenty candidate sets (ten mechanism families from each of two readouts) were shuffled and stripped of the original prompt, method identity, and answer key. The automated interpreter selected one of ten family labels for each set in five order-randomized passes; it did not solve the original engineering question. (B) Accuracy stays far above the 10\% chance level on every pass. (C) Majority-vote correctness for every family and readout: both columns are correct for nine families and both miss cleavage. Thus both naturally assembled vocabularies carry recognizable materials meaning, with no Jacobian-specific advantage in this secondary test (paired difference 0, $p=1.0$). (D) Each point is one Jacobian family. The horizontal coordinate is predeclared exact-term recovery; the vertical coordinate is the fraction of automated passes assigning the target-free words correctly. Cleavage is weak on both measures, but most families are semantically identified even when exact targets are rarely top-ranked, producing only a weak relationship ($\rho=0.254$). Controlled recovery and target-free recognition therefore measure complementary properties. The five passes are repeated judgments from one automated system, not independent scientific expert (hence this figure is limited to reproducible machine-based semantic validation).",reviewed result group(s); prefixes include related endpoints and dependencies,"[""readout""]","[""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/runs/""]",81,"[""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_materials_heldout_v1.py""]","[{""publication_path"": ""paper/figures/automated-family-identification.pdf"", ""sha256"": ""418fcc81622d200d702f4dcb3d08ee69c835afbeea0489d111f2f4232aa0d6ee"", ""matching_archived_files"": []}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens.tex,434,figure,fig:cases,"\textbf{Two magnified examples of the controlled-rank measurement.} The panels print the exact prompts, predeclared absent words, and one-token continuations. (A-C) The notch-resistance example is an earlier development item excluded from all held-out inference. Panel B plots \code{toughness}'s full-vocabulary rank at every sampled layer for three independent Jacobian fits and direct unembedding; upward is stronger because rank 1 is best. The registered 38-92\% band is shaded. Panel C reduces each word to its best rank within that band, showing that the localized rank-2 event is specific to \code{toughness}. (D-F) The boundary-attack example is one of the 50 held-out prompts and was already counted in Figure~\ref{fig:controlled}; it is not a new test. All three lenses sustain \code{corrosion} at rank 1 over two sampled layers, while direct unembedding reaches rank 111. Panel F shows that \code{sensitization} and \code{boundary} do not receive the same advantage. Thus Figure~\ref{fig:controlled} asks how often declared concepts are recovered across the frozen population, whereas this figure shows the layer location, fit-to-fit agreement, duration, and word specificity that its aggregate statistic necessarily discards. Neither trajectory is a literal chain of thought.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""readout"", ""geometry""]","[""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/experiments/materials-heldout-v1_latent"", ""jlens_materials/runs/""]",81,"[""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_materials_heldout_v1.py"", ""python scripts/analyze_materials_latent_geometry.py"", ""python scripts/build_final_paper_figures.py""]","[{""publication_path"": ""paper/figures/figure7-layer-resolved-cases.pdf"", ""sha256"": ""d535c6ca713c1ef188db35dd4082d2d9308846564c27c6c5691cd9b968b4889c"", ""matching_archived_files"": [""jlens_materials/figures/materials-heldout-v1/figure7-layer-resolved-cases.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens.tex,455,figure,fig:steering-screen,"\textbf{Broad steering is mechanism dependent.} Each panel averages ten new physical conditions and plots how an intermediate-state intervention changes the log odds of the first scientific answer relative to the second: \code{grooves} versus \code{clean} for corrosion (A), \code{hard} versus \code{soft} for transformation (B), and \code{higher} versus \code{lower} for grain size (C). Zero is the unperturbed model; the horizontal coordinate is the added direction's size relative to the final-prompt residual norm. Points are condition-weighted means and shaded bands are one standard error across physical conditions. The teal curve applies the scientifically matched Jacobian direction. Purple applies the two unrelated materials directions at the identical layer, green constructs the matched direction without Jacobian transport, and gray averages ten random directions. Corrosion is the only family that passes every \hlyellow{internally prespecified gate}: its matched curve is strong, monotonic, and distinct from all controls. Transformation trends in the intended direction but lacks \hlyellow{the prespecified specificity criterion}. The grain curve averages refinement, for which \code{higher} is correct, with coarsening, for which \code{lower} is correct; its cancellation motivated the relation-aware experiment in Figure~\ref{fig:steering}. The figure therefore shows both the causal promise and the non-universality of mechanism steering.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""steering""]","[""jlens_materials/experiments/candidate-followups-2026-07-16/"", ""jlens_materials/experiments/relational-grain-steering-v4"", ""jlens_materials/experiments/semantic-steering-v3""]",64,"[""python scripts/analyze_candidate_followups.py"", ""python scripts/analyze_relational_grain_steering_v4.py"", ""python scripts/analyze_semantic_steering_v3.py""]","[{""publication_path"": ""paper/figures/family_dose_response.pdf"", ""sha256"": ""bd3ce94f4bfa2be41679874eedced9ffa5d1ff7a690dcb5fccba5c15cee811a8"", ""matching_archived_files"": [""jlens_materials/figures/semantic-steering-v3/family_dose_response.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens.tex,472,figure,fig:steering,"\textbf{Can one frozen grain-size mechanism direction change a scientific answer in the relation-appropriate direction?} Six new materials were each written as a matched pair: grain refinement and grain coarsening with alloy identity, the two grain sizes, controlled covariates, and answer words held fixed. At the final prompt token in layer 16, the same frozen Jacobian direction was swept from $-4\%$ to $+4\%$ of the state's norm; the full strings \code{higher} and \code{lower} were scored at each dose. (A) For every material, the raw change in \code{higher}-minus-\code{lower} log odds is positive for refinement and negative for coarsening. Each connecting line therefore crosses zero: the direction responds to the relation rather than always favoring one word. (B) Signs are aligned so that positive always means movement toward the physically correct answer, then refinement and coarsening are averaged within material. Every circle is one independent matched pair; diamonds and bars are the mean and pair-clustered 95\% interval. The matched Jacobian effect is $+0.852$ and exceeds random, unrelated-mechanism, and direct controls in all six pairs. (C) The mean dose curves show the reversal continuously rather than only at endpoints. The experiment prospectively confirms a localized context-dependent causal pathway; it does not reveal a prose chain of thought or establish general materials reasoning. Exact prompts, both answer orders, clean outputs, all 2,400 intervention rows, and the full protocol are printed or linked in the Supplementary Information.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""steering""]","[""jlens_materials/experiments/candidate-followups-2026-07-16/"", ""jlens_materials/experiments/relational-grain-steering-v4"", ""jlens_materials/experiments/semantic-steering-v3""]",64,"[""python scripts/analyze_candidate_followups.py"", ""python scripts/analyze_relational_grain_steering_v4.py"", ""python scripts/analyze_semantic_steering_v3.py""]","[{""publication_path"": ""paper/figures/relational_confirmation.pdf"", ""sha256"": ""37895d87ab6e4b8be70fa5a1d11856cf5ba81d599dc43cece3843994e2ddbd9c"", ""matching_archived_files"": [""jlens_materials/figures/relational-grain-steering-v4/relational_confirmation.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens.tex,497,figure,fig:patching,"\textbf{Can a hidden state from the opposite grain-size relation causally transfer its answer?} A receiver prompt and donor prompt use the same answer words but describe opposite refinement/coarsening relations. At one layer, the receiver's final-prompt-token state is replaced by the donor state; positive values mean movement toward the answer appropriate to the donor's reversed relation. If the state contains only wording or alloy identity, the reversed transplant should not consistently redirect the answer. (A) Read from left to right, same-material and cross-material reverse patches have little effect early but produce a sharp late-layer shift of roughly 11 log-odds units at the peak. Relation-preserving and answer-order-only directions are rescaled to the same receiver-donor distance but remain much smaller. The shaded vertical region is the fixed 38-92\% analysis band; the band-mean same-material reverse effect is $+6.249$ ($5.663$-$6.795$). (B) The same-material reverse effect appears in all six matched alloy pairs, excluding a result driven by one material. (C) Layers where refinement and coarsening are more separable by a vocabulary readout tend also to show larger patching effects. The Jacobian level correlation is $\rho=0.818$ and the direct correlation is $\rho=0.571$, but onset and first-difference checks support only a broad late-stage correspondence, not exact layer-by-layer localization. The experiment establishes causal sufficiency in this constrained decision: a naturally occurring reversed-relation state can redirect the answer and can transfer between alloys. It does not identify a unique one-dimensional mechanism, prove necessity, or reveal a prose chain of thought. It is called exploratory because it reused a grain-size cohort whose strong effect was already known, not because the intervention is noncausal.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""patching""]","[""jlens_materials/experiments/candidate-activation-patching-2026-07-16/"", ""jlens_materials/experiments/cross-mechanism-activation-patching-2026-07-18/""]",42,"[""python scripts/analyze_counterfactual_activation_patching.py"", ""python scripts/analyze_cross_mechanism_activation_patching.py""]","[{""publication_path"": ""paper/figures/counterfactual-activation-patching.pdf"", ""sha256"": ""04b3fed9c3fd6f8e9cb28a62494e10c0adbe855c8130395007fd17a98dd10060"", ""matching_archived_files"": []}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens.tex,520,figure,fig:causal-knockout,"\textbf{A development-fitted relational coordinate makes a selective, context-dependent causal contribution to equation-free physical decisions.} (A) Setting only the coordinate to its development midpoint reduces the correct-answer margin at all three scales; equal-length activation-matched controls remain near zero. Pale points are 24 law identities; large markers and bars are law means and 95\% law-bootstrap intervals. (B) Replacing only the coordinate with its value from the matched reversed-input prompt moves the decision toward the counterfactual answer. Full-state replacement shows the total transferable effect; matched controls remain near zero. (C) The coordinate-only share of full-state transfer decreases with scale, but restoring the original coordinate within the counterfactual state removes approximately one-half of the total effect at every scale. (D) Neutralization changes discrete accuracy chiefly at 4B. Full-state counterfactuals reverse almost all original decisions, while coordinate restoration recovers many of them. The experiment establishes selective causal involvement, not a unique or universally necessary representation.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""causal_coordinate"", ""relational""]","[""jlens_materials/experiments/causal-relational-comment3-audit-2026-09-06/"", ""jlens_materials/experiments/causal-relational-knockout-2026-09-02/"", ""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/""]",219,"[""python scripts/analyze_causal_relational_knockout.py"", ""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py"", ""python scripts/validate_causal_relational_knockout.py""]","[{""publication_path"": ""paper/figures/causal-relational-knockout.pdf"", ""sha256"": ""584f0d774905411e25fa227be967d6a78eae30709d93c831fd45e9a9d32e3dca"", ""matching_archived_files"": []}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,78,longtable,,,study-wide summary/provenance; consult named rows,"[""readout"", ""target_free"", ""geometry"", ""physical_equivalence"", ""graphs"", ""relational"", ""steering"", ""patching"", ""causal_coordinate"", ""multitoken"", ""31b_streams""]","[""jlens_materials/experiments/answer-code-binding-2026-07-17/"", ""jlens_materials/experiments/answer-scaffold-audit-2026-07-17/"", ""jlens_materials/experiments/candidate-activation-patching-2026-07-16/"", ""jlens_materials/experiments/candidate-followups-2026-07-16/"", ""jlens_materials/experiments/candidate-nonsteering-2026-07-16/"", ""jlens_materials/experiments/causal-relational-comment3-audit-2026-09-06/"", ""jlens_materials/experiments/causal-relational-knockout-2026-09-02/"", ""jlens_materials/experiments/cross-mechanism-activation-patching-2026-07-18/"", ""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-semantic-streams-2026-08-26/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/lexical-adversarial-representation-2026-07-17/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/experiments/materials-heldout-v1_latent"", ""jlens_materials/experiments/multitoken-sequence-robustness-2026-07-18/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/"", ""jlens_materials/experiments/relational-grain-steering-v4"", ""jlens_materials/experiments/semantic-steering-v3"", ""jlens_materials/runs/""]",641,"[""python scripts/analyze_answer_code_binding.py"", ""python scripts/analyze_answer_scaffold_audit.py"", ""python scripts/analyze_candidate_followups.py"", ""python scripts/analyze_candidate_nonsteering.py"", ""python scripts/analyze_causal_relational_knockout.py"", ""python scripts/analyze_counterfactual_activation_patching.py"", ""python scripts/analyze_cross_mechanism_activation_patching.py"", ""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/analyze_late_physics_replication.py"", ""python scripts/analyze_lexical_adversarial_representation.py"", ""python scripts/analyze_materials_heldout_v1.py"", ""python scripts/analyze_materials_latent_geometry.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py"", ""python scripts/analyze_relational_grain_steering_v4.py"", ""python scripts/analyze_semantic_steering_v3.py"", ""python scripts/build_final_paper_figures.py"", ""python scripts/build_gemma31b_semantic_streams.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_multitoken_sequence_robustness.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_causal_relational_knockout.py"", ""python scripts/validate_graph_generalization_audit.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,138,table,,Primary held-out controlled result. Lens-fit seeds are averaged within prompt; population uncertainty uses ten mechanism families.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""readout""]","[""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/runs/""]",81,"[""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_materials_heldout_v1.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,162,table,tab:s-multiscale-equation-free,"\hlyellow{Equation-free relational physics across Gemma scale. The same 128 development prompts and 960 test prompts are used at every checkpoint. A separate development-law direction is fitted in each model's raw-state coordinates at approximately 83\% normalized depth; no test state contributes to the fit. ``D/I,'' ``D/N,'' and ``N/I'' are direct--inverse, direct--validation-neutral, and validation-neutral--inverse AUC.}",reviewed result group(s); prefixes include related endpoints and dependencies,"[""relational""]","[""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/""]",125,"[""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,193,table,tab:s-causal-knockout,\hlyellow{Original 24-relation equation-free intervention across Gemma scale. The retrospective overlap-exclusion analysis is reported separately in Section S7A. Values in brackets are 95\% law-bootstrap intervals. Margin loss is clean correct-answer margin minus neutralized margin; counterfactual movement is toward the answer appropriate to the reversed numerical input. Controls average both signs of three activation-derived axes at the identical intervention length.},reviewed result group(s); prefixes include related endpoints and dependencies,"[""causal_coordinate"", ""relational""]","[""jlens_materials/experiments/causal-relational-comment3-audit-2026-09-06/"", ""jlens_materials/experiments/causal-relational-knockout-2026-09-02/"", ""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/""]",219,"[""python scripts/analyze_causal_relational_knockout.py"", ""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py"", ""python scripts/validate_causal_relational_knockout.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,216,table,,Post hoc cross-law continuous-state audit. The contrast is same-physical/opposite-numerical similarity minus same-numerical/opposite-physical similarity across nine direct-law--inverse-law mechanism pairs. Positive values would favor physical outcome.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,237,table,,Frozen 60-law neutral-anchored relational result. The hidden score is fitted on earlier laws and then read only through matched numerical reversals. Calibration-neutral laws define the center and scale; validation-neutral laws remain test cases.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""relational""]","[""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/""]",125,"[""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,259,table,,Retrospective target-free cross-phrasing classification. Each entry is accuracy across 50 prompts and ten balanced families. The prompt TF--IDF baseline is 56\%; balanced chance is 10\%.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""target_free""]","[""jlens_materials/experiments/candidate-nonsteering-2026-07-16/""]",21,"[""python scripts/analyze_candidate_nonsteering.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,278,table,,"Broad mechanism-steering results. Endpoints are positive-answer minus negative-answer log odds at $+4\%$ dose minus the same contrast at $-4\%$. The family verdict required every frozen gate, not merely a positive mean.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""steering""]","[""jlens_materials/experiments/candidate-followups-2026-07-16/"", ""jlens_materials/experiments/relational-grain-steering-v4"", ""jlens_materials/experiments/semantic-steering-v3""]",64,"[""python scripts/analyze_candidate_followups.py"", ""python scripts/analyze_relational_grain_steering_v4.py"", ""python scripts/analyze_semantic_steering_v3.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,299,table,,Prospective relational grain-size results. Positive values are aligned toward the frozen correct answer before refinement and coarsening are averaged within material.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""steering""]","[""jlens_materials/experiments/candidate-followups-2026-07-16/"", ""jlens_materials/experiments/relational-grain-steering-v4"", ""jlens_materials/experiments/semantic-steering-v3""]",64,"[""python scripts/analyze_candidate_followups.py"", ""python scripts/analyze_relational_grain_steering_v4.py"", ""python scripts/analyze_semantic_steering_v3.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,325,table,,Exploratory counterfactual activation patching. Positive values mean that a receiver moved toward the answer appropriate to the donor's reversed grain-size relation. Means and 95\% intervals average the fixed 38--92\% layer band with the six material pairs as independent units.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""patching""]","[""jlens_materials/experiments/candidate-activation-patching-2026-07-16/"", ""jlens_materials/experiments/cross-mechanism-activation-patching-2026-07-18/""]",42,"[""python scripts/analyze_counterfactual_activation_patching.py"", ""python scripts/analyze_cross_mechanism_activation_patching.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,347,table,,Exploratory geometry and secondary blinded results.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""geometry"", ""readout""]","[""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/experiments/materials-heldout-v1_latent"", ""jlens_materials/runs/""]",81,"[""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_materials_heldout_v1.py"", ""python scripts/analyze_materials_latent_geometry.py"", ""python scripts/build_final_paper_figures.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,367,table,,Retained negative secondary analyses. These tests constrain interpretation and are not evidence for the proposed structures.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""target_free""]","[""jlens_materials/experiments/candidate-nonsteering-2026-07-16/""]",21,"[""python scripts/analyze_candidate_nonsteering.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,381,table,,"Lexical-adversarial physical-equivalence results. Margin is centered cosine(anchor, physical paraphrase) minus cosine(anchor, near-verbatim physical counterfactual). Positive means that physics outranks wording. Intervals use the frozen two-stage family/triplet bootstrap.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""physical_equivalence""]","[""jlens_materials/experiments/answer-code-binding-2026-07-17/"", ""jlens_materials/experiments/answer-scaffold-audit-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/lexical-adversarial-representation-2026-07-17/""]",65,"[""python scripts/analyze_answer_code_binding.py"", ""python scripts/analyze_answer_scaffold_audit.py"", ""python scripts/analyze_late_physics_replication.py"", ""python scripts/analyze_lexical_adversarial_representation.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,404,table,,"Post hoc answer-scaffold audit and prospectively frozen arbitrary-code negative control. The scaffold comparison reuses stored states and was defined after both source experiments were inspected. Its first four rows are positive-minus-negative decoder-logit differences, equivalently pairwise log-odds units; they are not the centered-cosine margins reported in the preceding table.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""physical_equivalence""]","[""jlens_materials/experiments/answer-code-binding-2026-07-17/"", ""jlens_materials/experiments/answer-scaffold-audit-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/lexical-adversarial-representation-2026-07-17/""]",65,"[""python scripts/analyze_answer_code_binding.py"", ""python scripts/analyze_answer_scaffold_audit.py"", ""python scripts/analyze_late_physics_replication.py"", ""python scripts/analyze_lexical_adversarial_representation.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,423,table,,"Complete mechanism inventory for the two lexical-adversarial cohorts. Each row contains four material systems, each rendered as anchor, physical paraphrase, and lexical counterfactual. ``Positive/negative'' are the exact one-token scientific outcomes used only for clean-answer and answer-scaffold audits; the full-state similarity endpoint does not search for these words.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""physical_equivalence""]","[""jlens_materials/experiments/answer-code-binding-2026-07-17/"", ""jlens_materials/experiments/answer-scaffold-audit-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/lexical-adversarial-representation-2026-07-17/""]",65,"[""python scripts/analyze_answer_code_binding.py"", ""python scripts/analyze_answer_scaffold_audit.py"", ""python scripts/analyze_late_physics_replication.py"", ""python scripts/analyze_lexical_adversarial_representation.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,448,table,,"\textbf{Post hoc signed-relation graph across different material cases.} A source prompt selects one nearest eligible target in each other surface-variant group. Same material cases are prohibited. Precision is the fraction of 144 directed edges whose source and target share the registered binary relation. The mechanism family is supplied when candidates are defined. The later exact audit shows that this within-family physical relation is identical to same/different numerical direction, so these values measure comparative structure rather than independent identification of a constitutive-law sign.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,472,table,,\textbf{Graph-identifiability audit summary.} All endpoints reuse the same 72 natural questions and archived states; no new prompt or Gemma forward pass was added.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,491,table,,"\textbf{Post hoc cross-phrasing mechanism graph from the original 50-prompt suite.} Every prompt selects one nearest neighbor in each other phrasing fold; 10\% is balanced mechanism chance. This graph asks whether differently worded descriptions from the same family remain neighbors. It is complementary to, and less lexically controlled than, the signed-relation graph.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs"", ""readout""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/"", ""jlens_materials/runs/""]",225,"[""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/analyze_materials_heldout_v1.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,527,figure,fig:s1-reproducibility,"\textbf{Independent lens fits give nearly identical controlled ranks.} Each cell is the Spearman correlation across all 150 held-out prompt--concept pairs. The three Jacobian fits use independent 1,000-record WikiText samples yet agree at $\rho=0.9978$--$0.9987$. Direct unembedding is much less correlated with their ordering ($\rho\approx0.21$). This figure establishes repeatability of the fitted measurement, not that its decoded terms are causal or universally more accurate.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""readout""]","[""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/runs/""]",81,"[""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_materials_heldout_v1.py""]","[{""publication_path"": ""paper/figures/figure-s1-fit-reproducibility.pdf"", ""sha256"": ""81c47c89e887f68597ae7812858faca864cf327aa78b1e467f40ebdc60932f76"", ""matching_archived_files"": [""jlens_materials/figures/materials-heldout-v1/figure-s1-fit-reproducibility.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,534,figure,fig:s2-development-stream,"\textbf{Sample semantic stream.} The exact prompt was: ``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.'' The one-token continuation was ``The.'' The ribbons were selected from the unrestricted leading decoded tokens at each layer, so the plotting code was not given a word list. The scientifically meaningful term \texttt{coalescence} appears naturally among those tokens, alongside generic discourse words that reveal the readout's noise. However, \texttt{coalescence} had been declared separately as a tracked concept when this earlier development prompt was designed. This is therefore an illustrative one-lens development display, not an unanticipated held-out discovery and not a literal chain of thought.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""readout"", ""geometry""]","[""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/experiments/materials-heldout-v1_latent"", ""jlens_materials/runs/""]",81,"[""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_materials_heldout_v1.py"", ""python scripts/analyze_materials_latent_geometry.py"", ""python scripts/build_final_paper_figures.py""]","[{""publication_path"": ""paper/figures/figure-s2-development-semantic-stream.pdf"", ""sha256"": ""9f9c20c11b3eb6265efa5ca54f61d25c2f6f3c9615da64b3b808746dabaee821"", ""matching_archived_files"": [""jlens_materials/figures/materials-heldout-v1/figure-s2-development-semantic-stream.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,541,figure,fig:s-targetfree-classification,"\textbf{Complete target-free cross-phrasing classification.} (A) Accuracy for Jacobian and direct candidate lists before and after removing exact or morphologically related prompt words. The dotted line is balanced chance; the dashed line is a classifier that sees TF--IDF prompt words. The strict-filter result remains above chance for both readouts but does not favor the Jacobian lens. (B) Prompt-level confusion matrix for the strict Jacobian configuration. Each row contains five independently written prompts. Boundary attack and surface oxidation are perfectly identified, whereas cleavage, ductile failure, high-temperature deformation, and notch resistance share vocabulary with neighboring mechanisms. The complete 50-row predictions, family accuracies, macro-F1 scores, paired correctness, and permutation outputs are stored in \texttt{experiments/candidate-nonsteering-2026-07-16/}.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""target_free""]","[""jlens_materials/experiments/candidate-nonsteering-2026-07-16/""]",21,"[""python scripts/analyze_candidate_nonsteering.py""]","[{""publication_path"": ""paper/figures/semantic-classification.pdf"", ""sha256"": ""ef77c61162bf146cace69711ae5407556f392027cad6be51c5493a0745c3ef59"", ""matching_archived_files"": [""jlens_materials/experiments/candidate-nonsteering-2026-07-16/figures/semantic-classification.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,552,figure,fig:s-ontology-rsa,"\textbf{The frozen expert ontologies do not align significantly with Jacobian family geometry.} At every layer, pairwise distances among the ten family centroids are compared with (A) four broad response classes and (B) a multi-attribute materials ontology. Curves show Jacobian-transported and raw states; horizontal lines show target-layer and prompt-embedding baselines. The best multi-attribute Jacobian correlation is $\rho=0.148$ at 58.5\% depth, but the maximum-over-layers corrected value is $p=0.9695$. Prompt embeddings align more strongly at $\rho=0.453$. The result does not support a claim that the lens recovered this expert ontology.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""target_free""]","[""jlens_materials/experiments/candidate-nonsteering-2026-07-16/""]",21,"[""python scripts/analyze_candidate_nonsteering.py""]","[{""publication_path"": ""paper/figures/materials-ontology-rsa.pdf"", ""sha256"": ""673f763e195d6581d6401bd61ea49e6b9544f14a9930cf7a5e9a67dd636337b5"", ""matching_archived_files"": [""jlens_materials/experiments/candidate-nonsteering-2026-07-16/figures/materials-ontology-rsa.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,559,figure,fig:s-concept-roles,"\textbf{No universal entity--process--property depth ordering is detected.} Predeclared single-token concepts were assigned before calculation to physical entity/state, process/mechanism, or response/property roles. The panels show the distribution of each concept's strongest registered depth under Jacobian and direct readouts. The largest Jacobian separation among role medians is 8.2 percentage points and does not exceed a family-blocked role-label null ($p=0.1988$). Missing sustained events are retained separately rather than assigned an artificial late depth.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""target_free""]","[""jlens_materials/experiments/candidate-nonsteering-2026-07-16/""]",21,"[""python scripts/analyze_candidate_nonsteering.py""]","[{""publication_path"": ""paper/figures/concept-role-depths.pdf"", ""sha256"": ""a4168ba62f5feba46bf4c488ce2b1099a23769336a1bd5200feea427a042350f"", ""matching_archived_files"": [""jlens_materials/experiments/candidate-nonsteering-2026-07-16/figures/concept-role-depths.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,567,figure,fig:s-lexical-adversarial,"\textbf{The broad physical-equivalence hypothesis fails in the first frozen cohort.} Each triplet contains an anchor, a physically equivalent paraphrase whose wording and units change, and a near-verbatim counterfactual whose physical answer reverses. Positive margin means the anchor is closer to the physical paraphrase than to the lexical counterfactual. The registered 38--92\% band mean is negative under both readouts (panel A), and no family has a positive mean (panel B): wording dominates over the broad window. Panel C shows the unregistered late rise that motivated, but does not itself confirm, a new disjoint cohort with its late window fixed before execution. This retained negative result is why the manuscript claims a temporal shift rather than general physics-first geometry.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""physical_equivalence""]","[""jlens_materials/experiments/answer-code-binding-2026-07-17/"", ""jlens_materials/experiments/answer-scaffold-audit-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/lexical-adversarial-representation-2026-07-17/""]",65,"[""python scripts/analyze_answer_code_binding.py"", ""python scripts/analyze_answer_scaffold_audit.py"", ""python scripts/analyze_late_physics_replication.py"", ""python scripts/analyze_lexical_adversarial_representation.py""]","[{""publication_path"": ""paper/figures/lexical-adversarial-representation.pdf"", ""sha256"": ""251abac14605833c435f2448cbb9b5017ec9ebef8cded2ddfc88a7868184ecd6"", ""matching_archived_files"": [""jlens_materials/experiments/lexical-adversarial-representation-2026-07-17/figures/lexical-adversarial-representation.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,575,figure,fig:s-answer-scaffold,"\textbf{Explicit semantic answer scaffolding strongly amplifies the relation contrast.} The audit reuses stored states from the disjoint cohort and was specified only after both lexical-adversarial studies were inspected. (A) Before the answer choices, the late pre-choice state weakly separates the physically equivalent pair from the counterfactual under both readouts; after the prompt presents ordinary semantic choices such as \texttt{higher}/\texttt{lower}, the final-state contrast is much larger. Lines connect the same triplets and open symbols are cohort means. (B) The after-minus-before gain is positive in every mechanism family. (C) Ordinary semantic answer pairs are selected more reliably than the arbitrary A/B codes used in the separate falsification, both overall (87.5\% versus 62.5\%) and in most families; the dashed line is 50\% forced-pair chance. Because position, suffix, and answer-word exposure change together, this post hoc comparison does not identify which factor causes the gain. It does show why the late physical-equivalence transition cannot be interpreted as an option-free constitutive variable.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""physical_equivalence""]","[""jlens_materials/experiments/answer-code-binding-2026-07-17/"", ""jlens_materials/experiments/answer-scaffold-audit-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/lexical-adversarial-representation-2026-07-17/""]",65,"[""python scripts/analyze_answer_code_binding.py"", ""python scripts/analyze_answer_scaffold_audit.py"", ""python scripts/analyze_late_physics_replication.py"", ""python scripts/analyze_lexical_adversarial_representation.py""]","[{""publication_path"": ""paper/figures/answer-scaffold-audit.pdf"", ""sha256"": ""6e606c73c72ab695df6c9d7db1d7383cf5c63c8af1b7538fbead739adf83f046"", ""matching_archived_files"": [""jlens_materials/experiments/answer-scaffold-audit-2026-07-17/figures/answer-scaffold-audit.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,582,figure,fig:s-answer-code,"\textbf{The arbitrary A/B answer-code falsification is inconclusive because its manipulation check fails.} The prospective protocol replaced meaningful answer words with arbitrary \texttt{A}/\texttt{B} labels and reversed their mapping across matched prompt variants. (A) Before the mapping appears, the physical-equivalence separation remains small and heterogeneous. (B) After the mapping, the required A/B contrast is also unstable across layers; the orange line is after mapping and gray is the pre-mapping reference. Shading in A and B is the frozen late window. (C) Family means do not show a consistent transition from pre-code physics separation (circles) to post-mapping A/B separation (squares). (D) Clean forced-pair code accuracy is only 62.5\% overall, with 0\% for obstacle spacing; the dashed line is 50\% chance. Only 5 of 24 triplets formed a complete anchor--counterfactual--remapped pattern, and neither code token was the global top continuation in any of 72 prompts. The frozen geometric endpoint is therefore not interpretable as successful answer-independent binding. Retaining this failure prevents the semantic-scaffold result from being presented as if arbitrary output codes had already reproduced it.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""physical_equivalence""]","[""jlens_materials/experiments/answer-code-binding-2026-07-17/"", ""jlens_materials/experiments/answer-scaffold-audit-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/lexical-adversarial-representation-2026-07-17/""]",65,"[""python scripts/analyze_answer_code_binding.py"", ""python scripts/analyze_answer_scaffold_audit.py"", ""python scripts/analyze_late_physics_replication.py"", ""python scripts/analyze_lexical_adversarial_representation.py""]","[{""publication_path"": ""paper/figures/answer-code-binding.pdf"", ""sha256"": ""df36ec04638be489e87d95e7295c7926a471fa3c6c6288c519174588f213a848"", ""matching_archived_files"": [""jlens_materials/experiments/answer-code-binding-2026-07-17/figures/answer-code-binding.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,590,figure,fig:s-mechanism-graph,"\textbf{Complementary graph evidence from the 50 held-out mechanism descriptions.} Every prompt selects one neighbor in each other phrasing fold. A correct edge stays inside the same mechanism family. (A) Layerwise same-family edge fraction for Jacobian and raw states, with word and character TF--IDF shown as horizontal lexical references. The gray band is the frozen 38--92\% summary window and the dashed gray line is 10\% chance. (B) After pairwise Jacobian similarity is residualized against word TF--IDF, character TF--IDF, token overlap, prompt length difference, and phrasing fold, the frozen-band graph still reaches 19.0\% ($p=0.00026$), although raw states are similar. (C) For each family, filled symbols compare the most similar true-family target with the most lexically similar wrong-family target; positive values favor the physical family. Open symbols restrict the comparison to 126 lexical traps in which TF--IDF itself prefers the wrong family. RT, PS, NR, LD, OX, HT, DU, CY, CL, and BA denote the ten families defined in the main paper. (D) One frozen-band Jacobian adjacency matrix. Teal marks selected same-family edges and light gray marks cross-family edges; rows and columns are grouped only for display. This analysis shows statistically reliable cross-phrasing organization and resistance to selected lexical traps, but input-token and TF--IDF graphs remain competitive. It therefore does not establish a fully non-lexical materials ontology or a Jacobian-specific advantage.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs"", ""readout""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/"", ""jlens_materials/runs/""]",225,"[""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/analyze_materials_heldout_v1.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]","[{""publication_path"": ""paper/figures/mechanism-graph-evidence.pdf"", ""sha256"": ""f8dc7c4a233828d60a9404992d1fefacbf543963c8602063ef16e4a0e22ad31e"", ""matching_archived_files"": [""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/figures/mechanism-graph-evidence.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,598,figure,fig:s-relation-ranking,"\textbf{Exact-null and full-candidate checks for the 72-prompt signed-relation graph.} (A) The gray histogram contains all 46,656 balanced base-case assignments under the exact null. Each assignment preserves all three surface variants of every material case, including the counterfactual sign transform. Its mean same-direction edge fraction is 0.56; the observed Jacobian graph is 0.82 at the far right ($p=0.009623$). (B) Rather than evaluating only the selected edge, pairwise ROC--AUC uses every eligible same-direction and opposite-direction candidate. Jacobian, direct decoder-basis, and raw states reach AUC 0.82--0.83, while word TF--IDF, character TF--IDF, and answer order remain at or below 0.48. Error bars show the leave-one-mechanism-out range. The hatched numerical-direction oracle is a ceiling derived directly from the experimental numbers, not from Gemma. The later identifiability proof shows that, inside each supplied family, the registered physical pair label and this numerical-direction oracle are exactly the same target. The figure therefore establishes distributed comparative ranking, not an independent constitutive-law representation.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]","[{""publication_path"": ""paper/figures/relation-ranking-robustness.pdf"", ""sha256"": ""c9104aafa3600fcd6caafdb2199ce06d62032568097b32320a2743b77dd4b6ab"", ""matching_archived_files"": [""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/figures/relation-ranking-robustness.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,606,figure,fig:s-option-free-question,"\textbf{Frozen natural question-end robustness result.} The runner stops after the final token of the complete scientific question and adds no answer choice, answer word, A/B code, response instruction, or checkpoint marker. (A) Solid bars are selected-edge precision and hatched bars are full-candidate AUC in the 38--92\% band. The dashed line is the exact case-preserving null mean. (B) Layerwise precision rises late for Jacobian, direct, and raw states. (C) Five of six mechanism-family AUCs exceed 0.5, with substantial heterogeneity. (D) The same 72 stems at three positions show why prompt boundary must be controlled: the natural question end is significant, the checkpoint-suffixed condition is null, and the answer mapping is strongly amplified. Suffix and token position change together, so the comparison does not isolate a unique cause. This is a frozen positional robustness result on an inspected cohort, not an independent replication.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]","[{""publication_path"": ""paper/figures/option-free-question-end.pdf"", ""sha256"": ""d3d86f598e5900e6ccdae3bf71e48fa5bef3dc5da7170e554c3cbf9583e17aa1"", ""matching_archived_files"": []}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,613,figure,fig:s-cross-mechanism-graph,\textbf{The natural-question relation graph does not generalize across governing mechanisms.} Each source is ranked against cases from the other five mechanisms. The counter-numeric subset keeps only mechanism pairs for which numerical increase/decrease maps to physical outcome with opposite orientation. Jacobian overall AUC is 0.513 and counter-numeric AUC is 0.471; neither passes the exact 20-assignment orientation null. Direct and raw states behave similarly. Pair and layer panels retain all mechanism combinations and show that a few positive pairs do not form one transferable physical-outcome coordinate.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]","[{""publication_path"": ""paper/figures/cross-mechanism-outcome.pdf"", ""sha256"": ""db8670f52854f81597bb2fc95d82d5fa50c0c14b5c04a57b909f6d46c6c1e242"", ""matching_archived_files"": []}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,620,figure,fig:s-graph-model-tests,"\textbf{Held-out graph learning separates numerical direction from physical meaning.} The graph audit uses the same 72 natural questions and archived states; it creates no additional scientific prompts. Every panel holds out each entire mechanism in turn. (A) A graph isomorphism network (GIN) trained to identify absolute physical polarity fails (blue), as do the graphless and edge-shuffled controls. Numerical increase/decrease was explicitly included among the absolute-label node features, so its perfect purple readback is a positive-control check rather than graph discovery. The meaningful contrast is that graph structure cannot convert this supplied direction into physical polarity for an unseen law. Below-chance physical AUC can occur when a coherent two-way split receives the wrong unnamed orientation. (B) A separate polarity-free same/different relation GIN evaluates 72 ordered different-case, cross-variant pairs within each held-out 12-node mechanism graph. It excludes numerical direction and surface variant and does not consistently exceed its pair-MLP or shuffled-edge controls. (C) Without fitting a graph network, graded cosine similarity ranks same-direction pairs above opposite-direction pairs in all six mechanisms; Jacobian, direct, and raw states nearly coincide. Because same/different physical and numerical relation are the identical target inside each family, panel C demonstrates comparative structure rather than independent physics.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]","[{""publication_path"": ""paper/figures/graph-generalization-model-tests.pdf"", ""sha256"": ""27454398429d9d97a511b67d5370302b2fa3547db50e01712bf4adbcdcf75e0b"", ""matching_archived_files"": [""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/figures/graph-generalization-model-tests.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,627,figure,fig:s-graph-partition-tests,"\textbf{Community results depend on retained graph density and do not yield a unique transferable physical partition.} Spectral clustering is label-blind and scored with adjusted Rand index (ARI), which is unchanged when the two community names are swapped. (A) The blue line is the six-mechanism Jacobian mean; gray lines are individual mechanisms. Retaining all graded candidate similarities raises mean ARI from 0.267 to 0.674, showing that the top-edge visualization discards information and that mechanisms are heterogeneous. (B) In the weighted top-one graph, direct states match Jacobian while raw-state behavior varies by family; positive ARI is therefore not lens-specific. (C) Bars show six-mechanism mean ARI and open circles show mechanisms after exhaustively scoring all 462 balanced divisions. The registered split is not the uniquely preferred partition, either on the complete graph or when counterfactuals are held out. Black marks summarize the surface-block-preserving null. The panels distinguish a favorable pairwise or spectral projection from a stable global partition.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]","[{""publication_path"": ""paper/figures/graph-generalization-partition-tests.pdf"", ""sha256"": ""58c87377948c6936c4dce368886ed2210094fa65450b1d3e2ebe28b1e6d3862c"", ""matching_archived_files"": [""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/figures/graph-generalization-partition-tests.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,634,figure,fig:s-graph-network-examples,"\textbf{Representative weighted mechanism graphs show what the statistical audit evaluates.} Nodes are complete option-free materials questions, not words. Each mechanism contains four material cases rendered as anchor (circle), physical paraphrase (square), and lexical counterfactual (diamond); edge width encodes graded Jacobian-state similarity. The node numbers identify the twelve prompts and the two colors identify the registered binary directions. (A) Particle fraction has a clearly recovered split, (B) crosslink density is only partly separated, and (C) Orowan spacing is not recovered by the exhaustive partition endpoint. Direct-response and inverse-response mechanisms can nevertheless display similar two-sided form while assigning opposite physical meanings to the sides. That symmetry is why an unlabeled graph may transfer comparative structure without identifying which side means that the material property rises. These examples visualize archived measurements; they are not reasoning chains or additional statistical observations.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]","[{""publication_path"": ""paper/figures/graph-generalization-network-examples.pdf"", ""sha256"": ""6e23bece2bc73398b225861380b9fb6acf626eb3aec85c51d5bd12d7ffe7f7b8"", ""matching_archived_files"": [""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/figures/graph-generalization-network-examples.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,641,figure,fig:s-relation-atlas,"\textbf{All 25 actual option-free relation graphs.} Each panel contains the same 72 natural-question nodes in a fixed layout and all 144 selected Jacobian edges at one registered layer. The six compact islands, ordered Orowan, porosity, pearlite, dislocation, particles, and crosslinks, are supplied mechanism strata rather than discovered communities. Teal edges preserve the registered binary relation and coral edges do not. Circles, squares, and triangles denote anchor, physical paraphrase, and near-verbatim counterfactual; filled and open nodes denote the two registered orientations. Within each family this physical relation is exactly aliased with numerical direction, so the late transition is comparative rather than independently constitutive. Each panel prints the transformer layer, normalized depth, and relation-preserving edge percentage. No edge is interpreted as a reasoning step.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]","[{""publication_path"": ""paper/figures/relation-graph-layer-atlas.pdf"", ""sha256"": ""357d93ddbe210ca372a0caae8730635c15d7cc335b634d678fc586fc5dbcecc8"", ""matching_archived_files"": [""jlens_materials/experiments/relation-graph-visualization-2026-07-18/figures/relation-graph-layer-atlas.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,648,figure,fig:s-relation-persistence,"\textbf{The 25-layer persistence union contains 391 unique directed edges.} Every line is an edge selected at least once across the 25 registered layers; width and opacity encode the fraction of layers in which it recurs. Teal preserves the registered within-family relation and coral does not. The display reveals recurring mechanism-conditioned comparative subgraphs, but the six islands are imposed by the candidate rule and physical relation is exactly aliased with numerical direction in this cohort. It is therefore a large, data-supported visualization of measured topology rather than evidence for a discovered universal materials network. The complete 3,600 layer-edge rows and 391-row persistence table are machine readable.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]","[{""publication_path"": ""paper/figures/relation-graph-persistence.pdf"", ""sha256"": ""3f8d278a856207707d7b0b612de9c9146059d28b44576033b35e75a5a4c8a4c1"", ""matching_archived_files"": [""jlens_materials/experiments/relation-graph-visualization-2026-07-18/figures/relation-graph-persistence.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,655,figure,fig:s-multitoken,"\textbf{Complete-word scoring gives mixed multi-token evidence.} The exact two-piece contrast is \texttt{transgranular} versus \texttt{intergranular}; the exact three-piece contrast is \texttt{martensite} versus \texttt{bainite}. The first piece is scored from the intermediate readout and remaining pieces by teacher forcing through unchanged Gemma. Family means favor the target under all three Jacobian fits and exceed direct unembedding, but only 7 of 10 prompts are positive, failing the frozen 8-of-10 breadth gate. The fragment-level and full-sequence comparison shows why the rank of a first token piece is not a valid substitute for a technical word probability.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""multitoken""]","[""jlens_materials/experiments/multitoken-sequence-robustness-2026-07-18/""]",15,"[""python scripts/plot_multitoken_sequence_robustness.py""]","[{""publication_path"": ""paper/figures/multitoken-sequence-robustness.pdf"", ""sha256"": ""e222b5b975dd054177dc40e3d8047e54e43d6960a7557ffa6227ebd0fa06fdea"", ""matching_archived_files"": []}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,706,figure,fig:s-relational-layer-sweep,"\textbf{Descriptive depth profile for the disjoint 12-law confirmation.} At each layer, a new direction is fitted only on the 16 development laws and evaluated on the same 12 held-out direct and inverse laws. The purple point marks the development-selected layer 34 used for the frozen confirmation; the remaining depths are descriptive. The occasional high early values and small 12-law sample preclude a claim of first emergence, whereas the uninterrupted late plateau supports the narrower conclusion of late-stage stabilization. This sweep does not use the final 60-law cohort.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""relational""]","[""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/""]",125,"[""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py""]","[{""publication_path"": ""paper/figures/relational-contrast-layer-sweep.pdf"", ""sha256"": ""276df770f2a77441119fe550306bbfcc3f136624ec8cacb1d20d813b88ca197e"", ""matching_archived_files"": []}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,713,figure,fig:s-relational-physics,"\textbf{Complete neutral-anchored relational summary.} (A) All inverse, neutral, and direct laws relative to the calibration-neutral median; black segments are category medians. (B) Every law ordered by its neutral-scaled matched contrast. (C) Hidden-state and output-logit relational AUCs compared with word and character TF--IDF controls. (D) Equation-surface transfer, threshold-free separation, and calibrated-cut accuracy. The figure summarizes all 60 laws; the generated inventories below print every law-level value and every one of the 480 matched comparisons.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""relational""]","[""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/""]",125,"[""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py""]","[{""publication_path"": ""paper/figures/neutral-anchored-relational-physics.pdf"", ""sha256"": ""bd7f3039b31997f625f0195a386c4d0454dc0c9d840c3319d7cc6d9fa11d17a6"", ""matching_archived_files"": [""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/figures/neutral-anchored-relational-physics.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,725,longtable,tab:s-equation-free-examples,"\hlyellow{\textbf{Representative supplied-equation and equation-free prompt transformations.} The supplied-equation column gives the law-specific core added to the common instruction to infer monotonicity, determine numerical direction, and compose the two signs. The equation-free column prints the exact canonical surface-A prompt. Results summarize all 16 controlled variants of each law, not only the displayed string. The complete manifests retain all 960 exact prompts in each format.}",reviewed result group(s); prefixes include related endpoints and dependencies,"[""relational""]","[""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/""]",125,"[""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,774,table,tab:s-v3-family,Complete \hlyellow{internally prespecified} broad-screen family verdicts. Parentheses are 95\% condition-bootstrap intervals. ``Valid'' is the fraction of matched rows whose global top beginning is one of the two registered answers.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""steering""]","[""jlens_materials/experiments/candidate-followups-2026-07-16/"", ""jlens_materials/experiments/relational-grain-steering-v4"", ""jlens_materials/experiments/semantic-steering-v3""]",64,"[""python scripts/analyze_candidate_followups.py"", ""python scripts/analyze_relational_grain_steering_v4.py"", ""python scripts/analyze_semantic_steering_v3.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,792,figure,fig:s3-v3-specificity,"\textbf{Layer-matched mechanism specificity.} Rows identify the scientific answer being scored; columns identify the mechanism direction applied at that row's frozen layer. The diagonal is the scientifically matched direction. Corrosion is strongly selective ($+0.904$ versus $-0.721$ transformation and $+0.254$ grain directions). Transformation is $+0.282$ on its diagonal versus $-0.076$ and $-0.235$. Grain size is $+0.166$ versus $+0.024$ and $-0.153$. This comparison asks whether any technical direction changes any answer, or whether the sign is mechanism dependent.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""steering""]","[""jlens_materials/experiments/candidate-followups-2026-07-16/"", ""jlens_materials/experiments/relational-grain-steering-v4"", ""jlens_materials/experiments/semantic-steering-v3""]",64,"[""python scripts/analyze_candidate_followups.py"", ""python scripts/analyze_relational_grain_steering_v4.py"", ""python scripts/analyze_semantic_steering_v3.py""]","[{""publication_path"": ""paper/figures/cross_mechanism_specificity.pdf"", ""sha256"": ""c806f4c8b36b4efcecc297af626ddc666647d0fdff58bd7ef9b23aece822b357"", ""matching_archived_files"": [""jlens_materials/figures/semantic-steering-v3/cross_mechanism_specificity.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,803,figure,fig:s4-v3-aligned,"\textbf{Hypothesis-generating correct-answer alignment.} Signs are reoriented after execution so that positive denotes movement toward the frozen correct answer. Circles are determinate physical conditions; diamonds and bars are mean and 95\% intervals. Grain size shows the most consistent context dependence, motivating a new \hlyellow{internally prespecified} refinement/coarsening cohort. Because this analysis was chosen after the sign pattern was observed, it is not confirmatory evidence by itself.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""steering""]","[""jlens_materials/experiments/candidate-followups-2026-07-16/"", ""jlens_materials/experiments/relational-grain-steering-v4"", ""jlens_materials/experiments/semantic-steering-v3""]",64,"[""python scripts/analyze_candidate_followups.py"", ""python scripts/analyze_relational_grain_steering_v4.py"", ""python scripts/analyze_semantic_steering_v3.py""]","[{""publication_path"": ""paper/figures/expected_answer_alignment_secondary.pdf"", ""sha256"": ""8a6ffce0e2067ddd8a6632befd4cad863b73063e3296b410f3ab6193484840e9"", ""matching_archived_files"": [""jlens_materials/figures/semantic-steering-v3/expected_answer_alignment_secondary.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,828,figure,fig:s-steering-transfer,"\textbf{The frozen grain direction does not generalize to new answer words or the inverse Hall--Petch regime.} The left cohort uses six new conventional polycrystals and replaces \texttt{higher}/\texttt{lower} by \texttt{increase}/\texttt{decrease}; only 7 of 12 condition signs and 1 of 6 complete refinement/coarsening pairs are correct. The right cohort explicitly states an inverse Hall--Petch nanocrystalline regime; only 6 of 12 condition signs and 0 of 6 pairs are correct. Across both cohorts, 23 of 24 conditions move toward the designated positive answer word. The unchanged vector therefore behaves primarily as an answer-vocabulary direction outside its original confirmation format.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""steering""]","[""jlens_materials/experiments/candidate-followups-2026-07-16/"", ""jlens_materials/experiments/relational-grain-steering-v4"", ""jlens_materials/experiments/semantic-steering-v3""]",64,"[""python scripts/analyze_candidate_followups.py"", ""python scripts/analyze_relational_grain_steering_v4.py"", ""python scripts/analyze_semantic_steering_v3.py""]","[{""publication_path"": ""paper/figures/candidate-new-cohorts.pdf"", ""sha256"": ""5c435c12e8d8f386237690dc2fc105de692a6e3275d4722f8d43a404fabc3502"", ""matching_archived_files"": [""jlens_materials/experiments/candidate-followups-2026-07-16/figures/candidate-new-cohorts.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,862,figure,fig:s-activation-patching,"\textbf{Complete exploratory activation-patching result.} (A) Layerwise movement toward the reversed-relation answer. The matched same-material and transferable cross-material reverse states become strongly causal late in the network. Relation-preserving and answer-order-only directions are exactly distance matched and remain much smaller. The legend is placed below the panels so it does not obscure the late-layer curves. (B) Same-material reverse shifts for each alloy pair; all six show the same late-layer transition. (C) Comparison between causal patching and vocabulary-readable relation separation. The correlation supports broad stage agreement but, together with onset and first-difference diagnostics, does not identify the exact causal layer.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""patching""]","[""jlens_materials/experiments/candidate-activation-patching-2026-07-16/"", ""jlens_materials/experiments/cross-mechanism-activation-patching-2026-07-18/""]",42,"[""python scripts/analyze_counterfactual_activation_patching.py"", ""python scripts/analyze_cross_mechanism_activation_patching.py""]","[{""publication_path"": ""paper/figures/counterfactual-activation-patching.pdf"", ""sha256"": ""04b3fed9c3fd6f8e9cb28a62494e10c0adbe855c8130395007fd17a98dd10060"", ""matching_archived_files"": []}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,878,figure,fig:s-cross-mechanism-patching,"\textbf{Complete option-free cross-mechanism activation-patching result.} (A) Physical-outcome and numerical-direction contrasts for all 15 unordered mechanism pairs. (B) Pair-bootstrap intervals for all pairs and the three prespecified transfer subsets. (C) Both features emerge late, but numerical direction becomes stronger. (D) Donor-family effects reveal the failed breadth gate: four physical-outcome means are positive, while Orowan and porosity have the wrong physical sign. Because transfer survives different answer words, it is not simple token copying. Because numerical direction is stronger and family signs are heterogeneous, it is not a universal materials-relation state.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""patching""]","[""jlens_materials/experiments/candidate-activation-patching-2026-07-16/"", ""jlens_materials/experiments/cross-mechanism-activation-patching-2026-07-18/""]",42,"[""python scripts/analyze_counterfactual_activation_patching.py"", ""python scripts/analyze_cross_mechanism_activation_patching.py""]","[{""publication_path"": ""paper/figures/cross-mechanism-activation-patching.pdf"", ""sha256"": ""d6f8ca0f2a576248a7ac3d325f024c69b9782a2fe6dbe045dfde7e5d7b98d093"", ""matching_archived_files"": []}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,902,longtable,tab:s-causal-law-inventory,"Complete 24-law target inventory. ``Direct'' means that increasing the control increases the response; ``inverse'' means that it decreases the response. Exact systems, endpoint values, and all eight prompt variants per law are in the accompanying prompt files.",reviewed result group(s); prefixes include related endpoints and dependencies,"[""causal_coordinate""]","[""jlens_materials/experiments/causal-relational-comment3-audit-2026-09-06/"", ""jlens_materials/experiments/causal-relational-knockout-2026-09-02/""]",94,"[""python scripts/analyze_causal_relational_knockout.py"", ""python scripts/validate_causal_relational_knockout.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,961,table,tab:s-causal-overlap-sensitivity,\hlyellow{Retrospective causal sensitivity after excluding three repeated development relations. Each scale retains 21 relations and 168 prompts. Margin effects subtract the mean of the six equal-length activation-derived controls; brackets are 95\% relation-bootstrap intervals. The accuracy p-values use paired relation-level accuracy differences.},reviewed result group(s); prefixes include related endpoints and dependencies,"[""causal_coordinate"", ""relational""]","[""jlens_materials/experiments/causal-relational-comment3-audit-2026-09-06/"", ""jlens_materials/experiments/causal-relational-knockout-2026-09-02/"", ""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/""]",219,"[""python scripts/analyze_causal_relational_knockout.py"", ""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py"", ""python scripts/validate_causal_relational_knockout.py""]",[],FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,989,figure,fig:s-causal-law-audit,"\hlyellow{\textbf{The neutralization effect is broad across target relations.} Each line is one law and connects the mean effect of six equal-length activation-matched controls to the effect of neutralizing the frozen relational coordinate. The expected physical-minus-control sign occurs in 23/24 laws at 4B and 24/24 at 12B and 31B. The only 4B exception, creep-rupture life versus temperature, has seven negative margins and one tied margin in the unmodified 4B checkpoint.}",reviewed result group(s); prefixes include related endpoints and dependencies,"[""causal_coordinate""]","[""jlens_materials/experiments/causal-relational-comment3-audit-2026-09-06/"", ""jlens_materials/experiments/causal-relational-knockout-2026-09-02/""]",94,"[""python scripts/analyze_causal_relational_knockout.py"", ""python scripts/validate_causal_relational_knockout.py""]","[{""publication_path"": ""paper/figures/causal-relational-knockout-law-audit.pdf"", ""sha256"": ""61aae3a85f6f148e45e3c414cd17caf1e26385d967fc0d61df20278e26c280bc"", ""matching_archived_files"": [""jlens_materials/experiments/causal-relational-knockout-2026-09-02/figures/causal-relational-knockout-law-audit.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,996,figure,fig:s-causal-error-repair,\hlyellow{\textbf{Moving the coordinate toward the correct development centroid repairs a subset of natural errors.} Every line is a prompt with a strictly negative clean correct-answer margin; no case was selected by its post-intervention behavior. Teal lines cross the decision boundary: 4/8 errors at 4B and 2/8 at 12B are repaired. Gray lines remain on the wrong side. The 31B checkpoint made no baseline error in this cohort and is therefore omitted.},reviewed result group(s); prefixes include related endpoints and dependencies,"[""causal_coordinate""]","[""jlens_materials/experiments/causal-relational-comment3-audit-2026-09-06/"", ""jlens_materials/experiments/causal-relational-knockout-2026-09-02/""]",94,"[""python scripts/analyze_causal_relational_knockout.py"", ""python scripts/validate_causal_relational_knockout.py""]","[{""publication_path"": ""paper/figures/causal-relational-knockout-error-repair.pdf"", ""sha256"": ""0191a1323aaf88f3a8aa8f9b33af99b6d4baa3e5ec35912a835bd5c27c0e32c9"", ""matching_archived_files"": [""jlens_materials/experiments/causal-relational-knockout-2026-09-02/figures/causal-relational-knockout-error-repair.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,1013,figure,fig:s-equation-free-cross-layer,"\hlyellow{\textbf{Equation-free physical ordering is preserved while the fitted 31B readout axis changes across depth.} (A) The independently fitted layer-40 and layer-49 physical-outcome directions are shown exactly in their two-vector spanning plane; their cosine is 0.352 (69.4 degrees). This is a two-depth comparison, not an observed continuous rotation trajectory. (B) Despite the changed axis, raw matched contrasts for the same 60 held-out laws are nearly identical across layers (Pearson $r=0.996$; Spearman $\rho=0.949$). (C) Each line follows one law between the two measured depths. Direct, neutral, and inverse relations retain their ordering. A separate direction was fitted at each layer using the unchanged supplied-equation development cohort and then evaluated on the identical equation-free prompts. The analysis was defined after both depth runs existed and is descriptive.}",reviewed result group(s); prefixes include related endpoints and dependencies,"[""relational""]","[""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/""]",125,"[""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py""]","[{""publication_path"": ""paper/figures/equation-free-cross-layer-geometry.pdf"", ""sha256"": ""00c70381bbffe7f53e4a5252de8c5a806d69722546e3a221ab59d12d9f91aeb8"", ""matching_archived_files"": [""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/figures/equation-free-cross-layer-geometry.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,1025,figure,fig:s-vocabulary-scale,"\hlyellow{\textbf{Controlled materials-term readability across Gemma scale.} The same 50 held-out prompts and 150 predeclared one-token concepts are evaluated at every checkpoint. (A) Jacobian recovery as the vocabulary cutoff is relaxed from rank 1 to rank 100. (B) Matched direct-unembedding recovery. (C) Mean controlled-recovery AUC with 95\% family-hierarchical intervals; filled circles are Jacobian and open squares are direct. (D) Within-checkpoint Jacobian-minus-direct AUC with the same population intervals. The 4B and 12B Jacobian curves average three independent lens fits, whereas 31B has one fit. Larger checkpoints show higher absolute readability, but the Jacobian-minus-direct contrast changes sign and every interval crosses zero. This secondary result therefore does not establish a general scale-dependent advantage of Jacobian transport.}",reviewed result group(s); prefixes include related endpoints and dependencies,"[""readout""]","[""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/runs/""]",81,"[""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_materials_heldout_v1.py""]","[{""publication_path"": ""paper/figures/heldout-readability-across-gemma-scales.pdf"", ""sha256"": ""7f671c9f04ddcd767946dff8ef4cb6e095a1e2dbdc9942f49dcd2a1607b96758"", ""matching_archived_files"": [""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/figures/heldout-readability-across-gemma-scales.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,1050,figure,fig:s-31b-streams,"\hlyellow{\textbf{What vocabulary is readable across depth from the single 31B lens for the same five materials prompts used at 4B?} Rows A--E preserve the case order and paired format of the main-paper semantic-stream figure. Each left panel applies the identical unrestricted single-fit renderer. Each right panel removes exact input and continuation words, the frozen function list, low-frequency English strings, and lower-ranked prefix duplicates, then favors words that are specific across the complete 50-prompt 31B archive. The color keys in the empty early-depth region name the six displayed ribbons; moving the labels there is a layout operation and does not alter the data. The gray region marks the fixed 38--92\% analysis band. Chemical degradation (A), failure thresholds (B), permanent deformation (C), and joining (D) produce recognizable local neighborhoods, whereas the cleavage case (E) remains mixed. Gaps mean only that a displayed word was absent from the stored top-32 list at that depth. Ribbon thickness is a rank-derived visualization score, not probability, causal influence, or a literal chain of thought. Only one 31B lens was available, so this case series is descriptive and is not a fit-stability or model-scale comparison.}",reviewed result group(s); prefixes include related endpoints and dependencies,"[""31b_streams"", ""readout""]","[""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/gemma4-31b-semantic-streams-2026-08-26/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/runs/""]",89,"[""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_materials_heldout_v1.py"", ""python scripts/build_gemma31b_semantic_streams.py""]","[{""publication_path"": ""paper/figures/figure-s28-gemma31b-semantic-streams.pdf"", ""sha256"": ""2309efba07366b7e6213a9f68b60485ea17cc984cf1ecb1fb35f59dddf19a3cf"", ""matching_archived_files"": [""jlens_materials/experiments/gemma4-31b-semantic-streams-2026-08-26/figures/figure-s28-gemma31b-semantic-streams.pdf""]}]",FILE_INVENTORY.csv
paper/materials_jacobian_lens_supplement.tex,1067,longtable,,,study-wide summary/provenance; consult named rows,"[""readout"", ""target_free"", ""geometry"", ""physical_equivalence"", ""graphs"", ""relational"", ""steering"", ""patching"", ""causal_coordinate"", ""multitoken"", ""31b_streams""]","[""jlens_materials/experiments/answer-code-binding-2026-07-17/"", ""jlens_materials/experiments/answer-scaffold-audit-2026-07-17/"", ""jlens_materials/experiments/candidate-activation-patching-2026-07-16/"", ""jlens_materials/experiments/candidate-followups-2026-07-16/"", ""jlens_materials/experiments/candidate-nonsteering-2026-07-16/"", ""jlens_materials/experiments/causal-relational-comment3-audit-2026-09-06/"", ""jlens_materials/experiments/causal-relational-knockout-2026-09-02/"", ""jlens_materials/experiments/cross-mechanism-activation-patching-2026-07-18/"", ""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-semantic-streams-2026-08-26/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/lexical-adversarial-representation-2026-07-17/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/experiments/materials-heldout-v1_latent"", ""jlens_materials/experiments/multitoken-sequence-robustness-2026-07-18/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/"", ""jlens_materials/experiments/relational-grain-steering-v4"", ""jlens_materials/experiments/semantic-steering-v3"", ""jlens_materials/runs/""]",641,"[""python scripts/analyze_answer_code_binding.py"", ""python scripts/analyze_answer_scaffold_audit.py"", ""python scripts/analyze_candidate_followups.py"", ""python scripts/analyze_candidate_nonsteering.py"", ""python scripts/analyze_causal_relational_knockout.py"", ""python scripts/analyze_counterfactual_activation_patching.py"", ""python scripts/analyze_cross_mechanism_activation_patching.py"", ""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/analyze_late_physics_replication.py"", ""python scripts/analyze_lexical_adversarial_representation.py"", ""python scripts/analyze_materials_heldout_v1.py"", ""python scripts/analyze_materials_latent_geometry.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py"", ""python scripts/analyze_relational_grain_steering_v4.py"", ""python scripts/analyze_semantic_steering_v3.py"", ""python scripts/build_final_paper_figures.py"", ""python scripts/build_gemma31b_semantic_streams.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_multitoken_sequence_robustness.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_causal_relational_knockout.py"", ""python scripts/validate_graph_generalization_audit.py""]",[],FILE_INVENTORY.csv
paper/primary_prompt_inventory.tex,4,longtable,tab:all-primary-prompts,Complete primary association dataset: all 50 exact prompts and their predeclared candidate concepts. Each family contains five independent phrasings of the same physical mechanism. Asterisks mark declared multi-token labels retained in the manifest but dropped from direct single-token ranking.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""readout""]","[""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/runs/""]",81,"[""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_materials_heldout_v1.py""]",[],FILE_INVENTORY.csv
paper/protocol_chronology.tex,6,longtable,,,study-wide summary/provenance; consult named rows,"[""readout"", ""target_free"", ""geometry"", ""physical_equivalence"", ""graphs"", ""relational"", ""steering"", ""patching"", ""causal_coordinate"", ""multitoken"", ""31b_streams""]","[""jlens_materials/experiments/answer-code-binding-2026-07-17/"", ""jlens_materials/experiments/answer-scaffold-audit-2026-07-17/"", ""jlens_materials/experiments/candidate-activation-patching-2026-07-16/"", ""jlens_materials/experiments/candidate-followups-2026-07-16/"", ""jlens_materials/experiments/candidate-nonsteering-2026-07-16/"", ""jlens_materials/experiments/causal-relational-comment3-audit-2026-09-06/"", ""jlens_materials/experiments/causal-relational-knockout-2026-09-02/"", ""jlens_materials/experiments/cross-mechanism-activation-patching-2026-07-18/"", ""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma-multiscale-heldout-v1-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-semantic-streams-2026-08-26/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/lexical-adversarial-representation-2026-07-17/"", ""jlens_materials/experiments/materials-heldout-v1"", ""jlens_materials/experiments/materials-heldout-v1_latent"", ""jlens_materials/experiments/multitoken-sequence-robustness-2026-07-18/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/"", ""jlens_materials/experiments/relational-grain-steering-v4"", ""jlens_materials/experiments/semantic-steering-v3"", ""jlens_materials/runs/""]",641,"[""python scripts/analyze_answer_code_binding.py"", ""python scripts/analyze_answer_scaffold_audit.py"", ""python scripts/analyze_candidate_followups.py"", ""python scripts/analyze_candidate_nonsteering.py"", ""python scripts/analyze_causal_relational_knockout.py"", ""python scripts/analyze_counterfactual_activation_patching.py"", ""python scripts/analyze_cross_mechanism_activation_patching.py"", ""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_gemma_multiscale_heldout_v1.py"", ""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/analyze_late_physics_replication.py"", ""python scripts/analyze_lexical_adversarial_representation.py"", ""python scripts/analyze_materials_heldout_v1.py"", ""python scripts/analyze_materials_latent_geometry.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py"", ""python scripts/analyze_relational_grain_steering_v4.py"", ""python scripts/analyze_semantic_steering_v3.py"", ""python scripts/build_final_paper_figures.py"", ""python scripts/build_gemma31b_semantic_streams.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_multitoken_sequence_robustness.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_causal_relational_knockout.py"", ""python scripts/validate_graph_generalization_audit.py""]",[],FILE_INVENTORY.csv
paper/relational_physics_inventory.tex,13,table,,\textbf{Frozen neutral-anchored relational design.} The ten calibration-neutral laws alone define the empirical center and scale; the ten validation-neutral laws are untouched test cases.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""relational""]","[""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/""]",125,"[""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py""]",[],FILE_INVENTORY.csv
paper/relational_physics_inventory.tex,60,longtable,,,reviewed result group(s); prefixes include related endpoints and dependencies,"[""relational""]","[""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/""]",125,"[""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py""]",[],FILE_INVENTORY.csv
paper/relational_physics_inventory.tex,135,longtable,,,reviewed result group(s); prefixes include related endpoints and dependencies,"[""relational""]","[""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/""]",125,"[""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py""]",[],FILE_INVENTORY.csv
paper/relational_physics_inventory.tex,210,longtable,,,reviewed result group(s); prefixes include related endpoints and dependencies,"[""relational""]","[""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/""]",125,"[""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py""]",[],FILE_INVENTORY.csv
paper/relational_physics_inventory.tex,701,table,,\textbf{Complete frozen endpoint verdict.},reviewed result group(s); prefixes include related endpoints and dependencies,"[""relational""]","[""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/""]",125,"[""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py""]",[],FILE_INVENTORY.csv
paper/relational_physics_inventory.tex,720,table,,\textbf{Portable relational-physics SI files.} Hashes refer to the publication copies bundled with this Supplementary Information.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""relational""]","[""jlens_materials/experiments/elicited-physics-abstraction-2026-07-18/"", ""jlens_materials/experiments/equation-free-relational-physics-2026-08-21/"", ""jlens_materials/experiments/gemma-equation-free-multiscale-2026-08-25/"", ""jlens_materials/experiments/gemma4-12b-equation-free-relational-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-layer-comparison-2026-08-23/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-layer40-v1/"", ""jlens_materials/experiments/gemma4-31b-equation-free-relational-v1/"", ""jlens_materials/experiments/neutral-anchored-relational-physics-2026-07-18/"", ""jlens_materials/experiments/relational-contrast-confirmation-2026-07-18/""]",125,"[""python scripts/analyze_equation_free_relational_benchmark_amended.py"", ""python scripts/analyze_gemma31b_equation_free_layers.py"", ""python scripts/analyze_gemma_equation_free_multiscale.py"", ""python scripts/analyze_neutral_anchored_relational_benchmark.py""]",[],FILE_INVENTORY.csv
paper/review_robustness_prompt_inventory.tex,12,longtable,tab:s-option-free-prompts,Complete option-free natural question-end prompt inventory.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""graphs""]","[""jlens_materials/experiments/cross-mechanism-outcome-2026-07-18/"", ""jlens_materials/experiments/graph-isomorphism-generalization-2026-07-18/"", ""jlens_materials/experiments/graph-topology-rigorous-2026-07-17/"", ""jlens_materials/experiments/late-physics-representation-replication-2026-07-17/"", ""jlens_materials/experiments/option-free-question-end-2026-07-18/"", ""jlens_materials/experiments/relation-graph-visualization-2026-07-18/""]",144,"[""python scripts/analyze_graph_topology_rigorous.py"", ""python scripts/plot_graph_generalization_audit.py"", ""python scripts/plot_relation_graph_robustness.py"", ""python scripts/validate_graph_generalization_audit.py""]",[],FILE_INVENTORY.csv
paper/review_robustness_prompt_inventory.tex,107,longtable,tab:s-multitoken-prompts,Complete multi-token prompt and sequence-score inventory.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""multitoken""]","[""jlens_materials/experiments/multitoken-sequence-robustness-2026-07-18/""]",15,"[""python scripts/plot_multitoken_sequence_robustness.py""]",[],FILE_INVENTORY.csv
paper/review_robustness_prompt_inventory.tex,135,table,tab:s-cross-mechanism-pairs,All Jacobian counter-numeric mechanism-pair results.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""patching""]","[""jlens_materials/experiments/candidate-activation-patching-2026-07-16/"", ""jlens_materials/experiments/cross-mechanism-activation-patching-2026-07-18/""]",42,"[""python scripts/analyze_counterfactual_activation_patching.py"", ""python scripts/analyze_cross_mechanism_activation_patching.py""]",[],FILE_INVENTORY.csv
paper/review_robustness_prompt_inventory.tex,160,table,tab:s-option-free-patching-subsets,Frozen option-free activation-patching endpoints. Intervals resample unordered mechanism pairs. The pair-sign and structured columns are two-sided exact probabilities.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""patching""]","[""jlens_materials/experiments/candidate-activation-patching-2026-07-16/"", ""jlens_materials/experiments/cross-mechanism-activation-patching-2026-07-18/""]",42,"[""python scripts/analyze_counterfactual_activation_patching.py"", ""python scripts/analyze_cross_mechanism_activation_patching.py""]",[],FILE_INVENTORY.csv
paper/review_robustness_prompt_inventory.tex,177,table,tab:s-option-free-patching-families,Activation-patching results by donor mechanism. Positive physical transfer means that positive-outcome donors increased the receiver's positive-minus-negative answer margin relative to negative-outcome donors.,reviewed result group(s); prefixes include related endpoints and dependencies,"[""patching""]","[""jlens_materials/experiments/candidate-activation-patching-2026-07-16/"", ""jlens_materials/experiments/cross-mechanism-activation-patching-2026-07-18/""]",42,"[""python scripts/analyze_counterfactual_activation_patching.py"", ""python scripts/analyze_cross_mechanism_activation_patching.py""]",[],FILE_INVENTORY.csv
paper/steering_prompt_inventory.tex,31,longtable,,,reviewed result group(s); prefixes include related endpoints and dependencies,"[""steering""]","[""jlens_materials/experiments/candidate-followups-2026-07-16/"", ""jlens_materials/experiments/relational-grain-steering-v4"", ""jlens_materials/experiments/semantic-steering-v3""]",64,"[""python scripts/analyze_candidate_followups.py"", ""python scripts/analyze_relational_grain_steering_v4.py"", ""python scripts/analyze_semantic_steering_v3.py""]",[],FILE_INVENTORY.csv
paper/steering_prompt_inventory.tex,107,longtable,,,reviewed result group(s); prefixes include related endpoints and dependencies,"[""steering""]","[""jlens_materials/experiments/candidate-followups-2026-07-16/"", ""jlens_materials/experiments/relational-grain-steering-v4"", ""jlens_materials/experiments/semantic-steering-v3""]",64,"[""python scripts/analyze_candidate_followups.py"", ""python scripts/analyze_relational_grain_steering_v4.py"", ""python scripts/analyze_semantic_steering_v3.py""]",[],FILE_INVENTORY.csv
paper/steering_v3_result_table.tex,6,longtable,,,reviewed result group(s); prefixes include related endpoints and dependencies,"[""steering""]","[""jlens_materials/experiments/candidate-followups-2026-07-16/"", ""jlens_materials/experiments/relational-grain-steering-v4"", ""jlens_materials/experiments/semantic-steering-v3""]",64,"[""python scripts/analyze_candidate_followups.py"", ""python scripts/analyze_relational_grain_steering_v4.py"", ""python scripts/analyze_semantic_steering_v3.py""]",[],FILE_INVENTORY.csv
paper/steering_v4_result_table.tex,6,longtable,,,reviewed result group(s); prefixes include related endpoints and dependencies,"[""steering""]","[""jlens_materials/experiments/candidate-followups-2026-07-16/"", ""jlens_materials/experiments/relational-grain-steering-v4"", ""jlens_materials/experiments/semantic-steering-v3""]",64,"[""python scripts/analyze_candidate_followups.py"", ""python scripts/analyze_relational_grain_steering_v4.py"", ""python scripts/analyze_semantic_steering_v3.py""]",[],FILE_INVENTORY.csv