bind2_1 — a binding-architecture research model (results-only card)

Results-only card. No downloadable artifact. No weights, no code, no configuration, no training corpus, and no generator are published for bind2_1. This card reports what the model is, what was tested, and what the results mean. The numbers below are complete as reported; what is withheld is tooling and weights, not results. See Why no weights for the reason.

bind2_1 is one member of the bind-evolution line — a small-model architecture program for compositional state tracking at BabyLM 2026 Strict-Small scale. For the full falsification timeline and how the members relate, see the narrative hub SecludedCorner/bind-evolution.

---

Value at a glance

bind2_1 was put through a pre-registered campaign — thresholds and the judgment script frozen before any data existed — to separate two questions that are usually conflated: is the mechanism causally real at depth? and does its advantage grow with depth? The campaign returned a rare and useful combination:

what was asked	frozen result
Does the system discriminate on the hard, state-requiring items?	Yes. Pooled deep accuracy 85.83 vs every learned control at chance (17); a **69-point** margin.
Is that behavior caused by the mechanism?	Yes. Lesioning the state pathway removes 99.3% of the advantage; interchange patching flips 96.4% of answers to the donor's holder.
Does the advantage hold or grow with depth (the pre-registered criterion)?	No — frozen verdict NULL. The margin shrinks across depth; the NULL stands permanently.
What did the NULL actually reveal?	Limited effective depth — a concrete, measurable architectural target, not a rhetorical one.
Does more scale dissolve the effect?	No. Across a ~6× parameter ladder, the standard controls never leave their floor — the form change is principle-grade, not a small-model artifact.

The value of this member is not a leaderboard number. It is a cleanly separated causal result and a frozen NULL, published at the same resolution, so that a reader can trust exactly what was and was not established.

---

What bind2_1 is

A from-scratch binding architecture for compositional state tracking (tracking "who holds what" as a narrative proceeds), built at BabyLM 2026 Strict-Small scale (~24M parameters). It is the successor mechanism to bind2_0 (SecludedCorner/bind2_0, weights + code published there). This card describes results only; the mechanism's internals are not disclosed here.

Unlike a general language-model entry, bind2_1 was not evaluated as a single trained checkpoint. It was evaluated as a campaign: a frozen synthetic diagnostic exam (a narrative ladder of increasing binding depth), a teacher-forced readout, and a fixed roster of learned control and ablation arms trained under identical conditions so that any advantage can be attributed to the architecture rather than to the data or the training budget.

---

What was tested

A 10-arm campaign, 5 seeds per arm, with the success thresholds and the entire judgment script frozen before any data existed. The frozen gate was an AND over five criteria. The exam is a synthetic diagnostic ladder whose deep rungs (labelled R3–R6 below, corresponding to operation depth 3 through 6) require genuine state tracking; the shallow rungs are solvable by shortcut and are excluded from the claim by design.

Pre-registered decision parameters (frozen):

• Primary endpoint = pooled deep accuracy, micro-averaged over depth rungs R3–R6.
• Minimum effect of interest (SESOI) = 5.0 percentage points on the primary endpoint.
• Anchors = a scrambled negative control (destroys the input structure), an oracle upper bound (has perfect state access), chance, and an audited shortcut ceiling (the best score any state-blind strategy could reach).

---

Results

1. Discrimination — PASS

The full system scored 85.83 pooled deep accuracy. Every one of the seven learned control arms sat at chance (≈16.5–17.5; chance = 16.98) — a ~69-point margin over each control, on all 5 seeds, against the pre-registered 5-point minimum effect. The scrambled negative control scored below chance (12.67); the oracle upper bound scored 99.92. The audited shortcut ceiling was ≤19.69 pooled; the system exceeds it by ~66 points. The advantage is not reachable by any state-blind shortcut.

2. Causality — PASS

The discrimination is caused by the mechanism, established two independent ways:

• State-lesion (necessity): removing the mechanism's state pathway removes 99.3% of the deep-rung advantage, while non-query language-modeling perplexity stays flat (the lesion is targeted, not a general degradation).
• Interchange intervention (sufficiency): patching the state from a donor context flips 96.4% of answers to the donor context's holder, with residual specificity 1.0.

Both replicate on all 5 seeds. A bypassed or spurious module cannot carry 99.3% of an effect; the deep-rung behavior is carried by the mechanism, necessarily and sufficiently.

3. The pre-registered depth criterion — FAILED → frozen verdict NULL

One pre-registered criterion required the depth-interaction margin to hold or grow with depth. The observed per-rung margins instead shrink across the four deep rungs:

R3 73.3 → R4 74.5 → R5 66.5 → R6 60.3 points.

The pre-registration had labelled a shrinking margin a "bypass" signal, so the AND-gate returns NULL. This frozen verdict stands permanently and is never re-labeled PASS. Publishing it at full resolution — alongside the passing causal results — is the point.

4. The autopsy (recorded, not used to overturn the NULL)

Adversarial post-hoc analysis found the failed criterion was ceiling-confounded: because chance itself falls as depth increases, only a near-lossless system (like the oracle, which passes it) could possibly satisfy it; a mechanism already starting near 96% at shallow depth has no room to grow its margin. This is a design-time specification error, recorded honestly as exactly that — it is not used to modify the frozen verdict. Separately, the "bypass" interpretation of the NULL is refuted by the causal results: the deepest rung (72.8) sits ~57 points above the audited shortcut ceiling (≤15.5), which a bypassed module could not produce.

5. The real finding — limited effective depth

The mechanism's engagement decays gracefully with depth (≈95% → 69%) while the oracle stays ≈100%. In plain terms: depth-robust binding is achievable on this task, and the learned mechanism does not fully achieve it. That gap is a concrete, measurable architectural target — the most useful thing the NULL bought.

---

The dual-verdict structure (stated explicitly)

Two verdicts exist for bind2_1. They answer different questions and are never merged:

1. The original pre-registered verdict is NULL, permanently. Criteria and judgment were frozen before data; the depth criterion failed; the AND-gate returns NULL. Re-labeling it PASS after seeing the data would be criterion-shopping, which pre-registration exists to make impossible. The autopsy above is recorded alongside this verdict; it does not modify it.

2. A corrected-criterion confirmatory run is a separate, second question. A corrected criterion (per-rung margin above the pre-registered minimum effect, and deepest-rung accuracy far above the audited shortcut ceiling) was frozen on 2026-07-15, before touching five held-back seeds — those seeds played no role in designing it. The confirmatory rerun on those untouched seeds landed the same day: CONFIRMATORY PASS. Per-rung margins on the fresh seeds — 73.3 / 74.6 / 66.9 / 59.4 points — replicate the original seeds almost exactly; the deepest rung scores 72.7 against a 31.0 shortcut bar; and the causal lesion battery replicates on all five unseen-seed checkpoints (99.30% of the deep advantage removed by the state lesion). This answers "does the corrected criterion hold on fresh seeds?" (yes). It cannot retroactively change verdict #1.

One procedural note, disclosed in full: the judgment script as originally frozen expected control arms the confirmatory design never scheduled and exited without scoring; that output is preserved untouched, and the confirmatory verdict comes from a plumbing-fixed variant whose criteria are byte-identical to the frozen ones, adversarially reviewed before unblinding.

---

Does parameter scale alone dissolve the gap? (No)

The most compute-starved test in the program asked whether more parameters simply wash the form change out. The standard control architectures were re-trained across a parameter ladder spanning roughly 6× (from ~24M up to a matched ~145M configuration). On every claim-bearing deep rung, at every scale, the controls stayed on their floor — none crossed the discrimination bar, even when given up to 10× the campaign's training budget at the smallest and largest scales (3.3× at the two intermediate scales).

The controls neither climb toward the threshold nor decay toward chance as they grow; they sit on a fixed, scale-invariant plateau. Reading: the form change observed at small scale is principle-grade — it is not a small-model sample-efficiency artifact that more parameters would dissolve. The claim stays bounded to the tested budget × scale box, with no extrapolation to unlimited parameters (the largest level was measured at a matched ~145M configuration; the span is stated as ~6×).

---

Numbers

Mechanism axis (frozen ladder, synthetic diagnostic exam, teacher-forced readout; accuracy %, mean ± SD, n = 5 seeds per arm)

The seven learned control/ablation arms are anonymized here (A–G); their internal identities and configurations are part of the withheld tooling. The negative-control, upper-bound, chance, and ceiling rows are as frozen. Depth rungs R3–R6 are the claim-eligible deep tiers; shallow rungs R0–R2 are shortcut-reachable by design and intentionally not tabled.

arm	R3	R4	R5	R6	pooled deep (R3–R6)
the full system	96.00 ± 0.00	92.90 ± 0.06	81.58 ± 0.14	72.82 ± 0.21	85.83 ± 0.05
learned control A	21.38 ± 0.32	17.28 ± 1.22	15.10 ± 1.40	13.75 ± 1.17	16.88 ± 0.14
learned control B	22.30 ± 2.06	18.85 ± 1.52	14.53 ± 1.48	12.53 ± 0.54	17.05 ± 0.72
learned control C	21.93 ± 2.01	16.48 ± 0.99	15.68 ± 1.41	12.03 ± 1.14	16.53 ± 1.11
learned control D	22.05 ± 0.69	18.32 ± 1.38	14.35 ± 0.90	12.17 ± 0.26	16.73 ± 0.72
learned control E	23.43 ± 1.37	19.37 ± 1.86	14.67 ± 1.25	12.35 ± 1.13	17.46 ± 0.57
learned control F	22.65 ± 1.46	18.40 ± 1.22	15.07 ± 0.75	12.47 ± 1.11	17.15 ± 0.50
learned control G	21.68 ± 0.89	18.60 ± 1.78	14.40 ± 0.78	12.90 ± 1.38	16.89 ± 0.67
scrambled negative control	16.05 ± 0.07	14.45 ± 0.07	11.25 ± 0.15	8.95 ± 0.14	12.67 ± 0.05
oracle upper bound	99.95 ± 0.07	100.00 ± 0.00	99.90 ± 0.10	99.85 ± 0.14	99.92 ± 0.04
chance	22.50	18.33	14.58	12.50	16.98
audited zero-state shortcut ceiling	≤25.38	≤20.50	≤17.38	≤15.50	≤19.69

Causal results (frozen gate): state-lesion kill = 99.3% of the deep-rung advantage with non-query perplexity flat; interchange flip = 96.4%; residual specificity = 1.0; all 5 seeds pass.

Depth-interaction margins (the failed pre-registered criterion): R3 73.3 / R4 74.5 / R5 66.5 / R6 60.3 points.

---

Why there are no weights

The headline number is a teacher-forced readout on a synthetic diagnostic corpus (not BabyLM data), and a standard free-running export does not reproduce it. Publishing weights that cannot reproduce their own headline would be misleading — so this stage ships as numbers and narrative only. The frozen pre-registration, the full decomposition, and the per-item results are on record; the numbers reported here are complete. What is withheld is tooling and weights, not results.

---

What this is — and is not

• A causal-mechanism result plus a frozen NULL, both pre-registered and published at the same resolution. It is not a general language-model entry and not a natural-language transfer claim.
• The frozen NULL is permanent. The confirmatory PASS is a separate, second question and does not — and cannot — overturn it.
• Diagnostic, teacher-forced, synthetic exam. The 85.83 is a depth-diagnostic capability score, not an official-benchmark score.
• Architecture attribution is clean by design — the advantage rests on seven learned controls trained under identical conditions, plus the ~6× parameter-scale scan above; it is not a data or budget artifact.
• Single-hardware program. Every result was produced on one 8 GB laptop-class GPU; single seeds were used where five were wanted, and budgets were smaller than a lab's. This is disclosed as a constraint on iteration rate, not on the frozen verdicts.

The causally-verified mechanism was subsequently carried to the official entity-tracking question form as the extension bind2_1e; that separate result, its scope, and its caveats are documented in the bind-evolution hub — not here.

This is part of a research program that we hope to develop toward a serious entry in a future BabyLM cycle; we aim to place in 2027. No ranking or superiority claim is made here.

---

How to cite

Cite this member card, pinned to the commit SHA of the revision you are citing (the authoritative SHA is stamped on this card as a dated addendum at publish time). Do not cite the bind-evolution hub as a result source — it is a narrative and navigation hub. Naming: the member id is bind2_1; a bare bind2 is never a repository or artifact name — the family is referred to in prose only.

Related: SecludedCorner/bind-evolution (the falsification timeline) · SecludedCorner/bind2_0 (the predecessor, weights + code).