SecludedCorner/bind-evolution

bind evolution — a falsification timeline

Research in progress. This repo is a narrative and navigation hub — not a leaderboard showcase, and not a citation target for any single result. To cite a result, cite the member repo that carries it, pinned to a commit SHA (see How to cite).

This is the research log of a small-model architecture line ("bind") for compositional state tracking at BabyLM 2026 Strict-Small scale (~24M params, 10M-word training corpus), told the only way we trust: as a sequence of pre-registered questions, frozen verdicts, and honest refutations — including refutations of our own published results. Every generation below either falsified its predecessor's headline claim or was itself falsified by a frozen criterion. The value of the line is not any single score; it is that each verdict — positive or null — forced a sharper next question.

Ground rules that hold throughout:

• Frozen verdicts stand. A pre-registered NULL is never re-labeled PASS, even when a post-hoc autopsy finds the criterion itself was flawed.
• We audit our own wins adversarially. The strongest refutation in this log was performed by us, on our own already-public result.
• Negative results are published at the same resolution as positive ones.

What each test asked, why it was run, and what it buys — on one 8 GB laptop GPU

Every experiment in this program ran on a single RTX 4060 Laptop GPU (8 GB). That constraint shaped everything: single seeds where five were wanted, 30M-token budgets where the baselines got 150M, a serial queue where a lab would fan out overnight, and one scale scan that costs ~48 wall-clock hours here versus a fraction of that on a modern training node. Here is what each test asked, why, and what its answer unlocks.

1. Does the headline result replicate? A single-seed result is an anecdote. The frozen procedure was re-run on two untouched seeds: 93.96 / 94.12 / 94.22 — a 0.26-point spread across three seeds. The transfer replicates; "provisional" was removed. Unlocks: the result is load-bearing enough to build on.

2. Is the advantage the architecture's — or just the training data's? Any model can look good on the diet built for it, so a standard attention-only transformer was trained on the identical synthetic diet and scored on the same items. On the subset where state tracking is actually required it floors at chance (23.2%) while the binding model scores 92.6% there. Unlocks: the advantage is an architectural asset, worth carrying to scale — not a data artifact.

3. Does the mechanism survive real language in the diet? A mechanism that only works on a pure synthetic diet is a lab curiosity. Diluted to 35% of a mixed diet (65% natural text, same total budget), the mechanism held — 94.54%, slightly above the pure-diet run — and the model's general grammar score landed 0.08 points below a frozen "no-tax" line while training on 5× less data than the baselines. Unlocks: full-budget mixed training is the single most obvious next run, and it is purely compute-bound.

4. Does it already transfer to natural-language reference tracking? The endgame is real language, so the gap was measured instead of assumed: pre-registered NULL, and precisely diagnostic — the mechanism's readout literally never fires on natural text. The gap is a missing interface, not a disproven mechanism. Unlocks: the bridge to natural text is now a concrete, bounded workstream rather than a hope.

5. Can the training scaffold be removed — or replaced by signals raw text can provide? (answered 2026-07-17 — in halves.) The recipe leans on training supervision the real world doesn't hand out, so two pre-registered arms asked whether each supervision channel can be dropped; frozen bands, NULL as the default prediction — and neither arm landed on the default. One channel: retained. Dropping that supervision entirely holds the result — pooled 94.12% against a frozen retention line of 88.96, statistically indistinguishable from the original (93.96), with the shortcut-unsolvable subset at 92.6%, empty-container items at 97.4%, and a routing gap of 0.000. The other channel: partial. Removing it costs measurably — pooled 81.50%, well above the frozen collapse line (59.4, the audited shortcut ceiling) but below retention; the mechanism's gain survives removal (77.6% on the shortcut-unsolvable items), routing stays intact, and it degrades gracefully rather than collapsing the result. Unlocks: the supervision question splits cleanly — one channel is already corpus-derivable, and the other is now a measured cost curve instead of an unknown.

6. Will the mechanism emerge on its own when the data demands it? (answered 2026-07-17.) The deepest question in the line: the mechanism was forced by construction — would training pressure alone produce it? A standard architecture was trained on a corpus engineered to strip the shortcut reward (so the only way to lower the loss is to grow the real parse), and on a matched corpus that leaves the shortcut in. Frozen verdict: neither learned it — both floor on the mechanism-requiring subset (~0.14, below chance) while the forced mechanism scores 0.92 on the identical test. Removing the shortcut-solvable pressure, under a generous budget, does not rescue learning: pressure is refuted as a sufficient cause — the mechanism can be forced by construction, but training pressure alone does not induce it. This is a publishable negative and a precise one: it bounds where emergence does and does not happen under a generous budget, rather than leaving "emergence" an open hope.

7. Does parameter scale alone dissolve the gap? (resolved 2026-07-21 — the most compute-starved experiment here.) Re-trained across a ~6× parameter ladder (24M → 145M), the standard control architectures stayed on their floor on every deep rung at every level: none of the claim-bearing deep cells crossed the bar (up to 10× the campaign training budget at the smallest and largest levels, 3.3× at the two middle levels). Scale alone does not buy this capability — the form change is principle-grade, not small-model sample efficiency. The controls neither climb toward the threshold nor decay toward chance as they grow; they sit on a fixed, scale-invariant plateau. 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, disclosed; the span is stated as ~6×.)

The ask

The methodology is the guarantee: criteria freeze before data exists, NULLs are published at the same resolution as wins, and the strongest refutation in this log was performed by us on our own already-public result. Compute given to this program is not spent chasing a leaderboard number — it is spent buying frozen answers, at a far higher iteration rate than this hardware allows. H100/H200-class hardware would turn multi-day scans like this one into afternoons, single seeds into five-seed ladders everywhere, the 30M-token stress tests into full-budget runs, and the bridge program (question 4) plus the emergence program (question 6) into a serious entry for higher-budget tracks and the 2027 cycle, where we aim to place.

To say it plainly: everything in this log — every frozen verdict, every replication, every control — was asked and answered on one RTX 4060 laptop GPU. We are proud of what that card has managed to answer, and we sincerely hope that one day the right connection brings better equipment within reach, so that more of these questions — and harder ones — can be attempted properly. If this log reads to you like a program worth equipping, we would be glad to hear from you.

The timeline

Stage 1 — bind1: the win we refuted ourselves

Our official BabyLM 2026 strict-small entry (~24M params, an iterative role-binding loop) showed what looked like the paper's key result: entity tracking 39.55 vs 27.82 for the matched monolith (+11.7 points, single seed), alongside BLiMP 65.5 vs the GPT-2 baseline 65.1. We flagged it at submission time as "the key result to replicate, not a settled fact" — and then we replicated adversarially instead of celebrating.

The audit finding: the entity-tracking lead was not tracking. In the benchmark's item pool, "nothing." never appears as a distractor (0 of 9,483 items) — whenever it is among the options it is the answer — and completion scoring structurally favors it. Counterfactual probes showed that no model in an 11-model, multi-seed grid could actually distinguish empty from non-empty containers (discrimination ≈ coin-flip, AUC 0.47–0.58); on non-"nothing" items every model, every seed, sat at chance. The entire ablation gradient that looked like a mechanism story was the gradient of a state-blind "nothing." completion prior. Under the leaderboard's corrected scoring standard (nothing-gold items removed), the lead disappears.

Both scorings are published side-by-side on the public bind1 model card (repo map below). This refutation-of-our-own-result is the credibility opener of the whole line: it is why you can trust the verdicts that follow.

Stage 2 — the curriculum duel: a null that indicted the exam, not the architecture

A pre-registered curriculum duel asked whether the binding architecture out-learns a matched monolith on an in-context binding curriculum. Frozen verdict: NULL — but of a specific, diagnostic kind: both arms stayed flat near floor across all 12 checkpoints through 50M tokens (endpoint 0.1875 vs 0.1842). Under the frozen decision grid this lands in the "exam/scale problem" cell: an exam neither arm can learn discriminates nothing about architecture. We recorded it as an exam-design artifact and drew the obvious lesson — before asking who learns faster, first build an exam that is demonstrably learnable. That lesson directly shaped the next two generations.

Stage 3 — bind2_0: the mechanism works; the transfer doesn't ("no tax, no win")

bind2_0 combines delta-rule fast-weight memory with a forced bottleneck: attention is chunk-local, so cross-chunk information can only flow through a recurrent state. Three results, all kept:

1. Direct-task training works. On a purpose-built synthetic swap-tracking task (n=800 per eval, 5-way, chance 0.20), bind2_0 reaches 0.9988 accuracy while its matched controls (monolith, bind1-style loop, no-binding control) sit at 0.2125 / 0.1938 / 0.1938 — with a sharp grokking transition between 5M and 10M training tokens (0.179 → 0.969 → 0.996). Learnable exam: achieved.
2. It does not emerge for free. Trained as a plain LM on the real BabyLM strict-small corpus, the architecture showed no emergent zero-shot state-tracking advantage (on a 60-probe test with chance 0.50, no model — ours or baseline — beat chance).
3. No general-language tax. On the official zero-shot evaluation it is statistically tied with the matched baselines, slightly above the GPT-2 baseline on BLiMP (66.11 at 23.9M params vs 65.08).

Honest summary: "no tax, no win." Mechanism capability and benchmark transfer are separate questions, and conflating them is how fields fool themselves. What this stage forced next: split the confound — first prove the mechanism is causally real at depth under a pre-registered gate, separately from transfer.

Weights and code for this stage: SecludedCorner/bind2_0 (main = 23.9M build; branch 27m = 27M build).

Stage 4 — bind2_1: a NULL that stands, a causal result that survives, and a new finding

The successor mechanism (weights/code not released — see below) was tested the hard way: a 10-arm campaign, 5 seeds per arm, with thresholds and the judgment script frozen before any data existed. The frozen gate was an AND over five criteria. What happened is the most instructive verdict in this log:

• Discrimination passed. The full system scored 85.83 (pooled deep-rung accuracy) against every control arm sitting at chance (≈16.5–17.5, chance 16.98) — a ~69-point margin over each of seven learned controls, on all 5 seeds, against a pre-registered minimum effect of 5 points. A scrambled negative control scored below chance (12.67), and an oracle upper bound scored 99.92. An audited shortcut ceiling (the best any state-blind strategy could reach) was ≤19.7 pooled; the system exceeds it by ~66 points.
• Causality passed. Lesioning the mechanism's state pathway removes 99.3% of the deep-rung advantage while leaving non-query language modeling flat; interchange patching flips 96.4% of answers to the donor context's holder, with residual specificity 1.0. The deep-rung behavior is carried by the mechanism — necessarily and sufficiently.
• Criterion C failed → verdict NULL. C required the depth-interaction margin to hold or grow with depth; the observed margins shrink (73.3 → 74.5 → 66.5 → 60.3 across the four deep rungs), which the pre-registration had labeled a "bypass" signal. The frozen verdict is NULL and it stands permanently.
• The autopsy — recorded, not used to overturn. Adversarial post-hoc analysis showed criterion C was ceiling-confounded: because chance itself falls with depth, only a near-lossless mechanism (like the oracle, which passes C) could pass; a system starting at 96% has nowhere to grow its shallow margin. This is a design-time specification error and we record it as exactly that. The "bypass" interpretation of the NULL is, separately, refuted by the causal results: a bypassed module cannot carry 99.3% of the effect, and the deepest rung (72.8) sits ~57 points above the audited shortcut ceiling (≤15.5).
• The real finding: limited effective depth. The mechanism's engagement decays gracefully with depth (≈95% → 69%) while the oracle stays ≈100% — so depth-robust binding is achievable on this task and the learned mechanism does not fully achieve it. That gap is a concrete architectural target, not a rhetorical one.

Why there is no bind2_1 repo: 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. What is withheld is tooling and weights, not results — the numbers reported here are complete.

Because the corrected depth criterion (C′) was formulated after seeing the data, it cannot be scored on that data as anything but exploratory. So C′ was frozen on 2026-07-15, before touching the five held-back seeds, and 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). One procedural note, disclosed in full: the judgment script as originally frozen demanded control arms the confirmatory design never scheduled and exited without scoring; that output is preserved untouched, and the verdict above comes from a plumbing-fixed variant whose criteria are byte-identical to the frozen ones, adversarially reviewed before unblinding, with the full evidence chain on record internally. The discrimination criterion was inherited from the original seeds, not re-measured. The seeds-0-4 NULL stands unchanged.

Stage 5 — bind2_1e: the mechanism passes the official exam (single seed, provisional)

The transfer question bind2_0 failed ("no tax, no win") could now be asked properly: take the causally-verified binding mechanism, train it on a synthetic box-tracking corpus whose answer statistics are distribution-matched to the official entity-tracking benchmark, and score it on the official items with fully learned routing — no oracle assistance at test time.

Everything was pre-registered and frozen before the run: the corpus ruling, the success/null bands, the shortcut-decomposition gates, the abort rules. Default prediction: NULL.

Result (2026-07-15), one shot, first read final: 93.96% pooled on the two in-scope official subsets (6,259 items; chance 20%; every earlier model in this program — and the published baselines — sits at ≈19–21% on this benchmark). The regular subset scores 99.05% with zero decay across operation depth (99.4% at the deepest tier); the contents-move subset scores 88.80% with graceful depth decay. The routing fear died completely: the learned router matches oracle-hinted routing to the third decimal (transfer gap 0.000).

The credibility core is the pre-registered decomposition. The benchmark's strongest audited shortcut ("the last-touched box is the answer") can solve ~59% of items; on the 2,550 items that shortcut cannot solve, the model scores 92.4% — and 96/92/92/92/81% across operation depths 1–5. Empty-box golds (the classic prior-abuse trap from Stage 1) score 97.6%, with a perfect 100% on the regular subset. The verdict label, per the frozen bands: MECHANISM-TRANSFER — provisional, single seed.

What this is not, stated plainly:

• Single seed means exactly that: provisional until replicated.
• Subset scope: the score covers the regular and contents-move subsets only — 6,259 of the benchmark's 9,483 items; the ambiguous-reference subset was excluded by the frozen pre-registration as outside mechanism scope.
• This is a mechanism-transfer demonstration, not a general language model. The model was trained only on synthetic box-tracking text; every other suite in the official evaluation is expected to sit at chance by design, and no claim is made there.
• The natural-diet baselines are diet-confounded as an architecture comparison. Their chance-level scores come from natural-text training; the matched-diet architecture attribution rests on the bind2_1 campaign's seven matched controls (Stage 4) plus the dedicated matched-diet control reported just below.

Nothing from this stage is downloadable at this point: the weights, code, configuration, and the corpus/generator tooling are all withheld at this stage. The frozen pre-registration, the decomposition, and the per-item results are on record internally; the numbers reported here are complete.

Update (2026-07-17) — three pre-registered follow-ups, all resolved in the mechanism's favour. Each froze its bands before its run; the results:

• Seed replication (now n = 3). The frozen procedure re-ran on two untouched seeds: pooled 94.12% and 94.22%, against the original 93.96% — a spread of 0.26 points across three seeds, with the shortcut-unsolvable subsets and empty-box golds replicating in lockstep. The "provisional — single seed" qualifier is removed: the transfer replicates.
• Matched-diet control (the diet-confound, closed). A standard attention-only transformer trained on the identical box-tracking diet (same token budget, single seed) scores pooled 50.6% — but on the pre-registered decisive subset, the items the shortcut cannot solve, it scores 23.2% ≈ chance (20%), versus this model's 92.6% on the same subset. Its overall half-score is carried entirely by the empty-box prior and the last-touch shortcut; on the items that require state tracking it floors. A standard architecture on the same diet cannot do it — the advantage is architectural, not a diet artifact. (Single seed = provisional.)
• Mixed-diet stress test. Retrained on 35% box-tracking + 65% natural text (same 30M-token budget), the mechanism holds: pooled 94.54% — if anything above the boxes-only run (report-only, single seed), all decomposition gates passed by wide margins. On this mixed diet the model also produces real (non-chance) general-language scores; its BLiMP came in at 59.92 against a pre-registered no-tax gate of 60.0 — 0.08 below the line, recorded under the frozen label TAX-OR-BUDGET: attribution left open between an architecture cost and the 5×-smaller token budget of this run versus the 150M-token baselines, neither claimed. A separate zero-training probe for transfer to natural-language reference tracking returned the pre-registered NULL on both primary subjects — expected, since the mechanism's readout is inert on those inputs; recorded as no evidence, not as a mechanism failure.

Stage 6 — where the line is now

Current focus: consolidating the replicated transfer result and its controls.

The dual-verdict structure, stated explicitly

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

1. The original pre-registered verdict is NULL, permanently. Criteria and judgment were frozen before data; criterion C failed; the AND-gate returns NULL. Re-labeling it PASS after seeing the data would be criterion-shopping, and the entire point of pre-registration is to make that impossible. The autopsy that found criterion C ceiling-confounded is recorded alongside the verdict; it does not modify it.
2. The corrected-criterion confirmatory run is a separate, second question. C′ (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 the five backup seeds were touched; those seeds played no role in designing C′. Its result (2026-07-15): CONFIRMATORY PASS — per-rung margins 73.3/74.6/66.9/59.4 points on the fresh seeds, deepest rung 72.7 vs the 31.0 shortcut bar, causal battery replicated on all five unseen-seed checkpoints. It answers "does the corrected criterion hold on fresh seeds?" (yes) — it does not, and cannot, retroactively change verdict #1.

Numbers

Mechanism axis (bind2_1 campaign; 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.

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

Notes (from the frozen source): pooled deep = micro-average over R3–R6, the frozen primary endpoint; pre-registered minimum effect (SESOI) = 5.0 pp on pooled deep; shallow rungs R0–R2 are shortcut-reachable by design (claim-ineligible) and intentionally not tabled.

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

Depth margins vs learned control F, the pre-registered reference control (criterion C, failed): R3 73.3 / R4 74.5 / R5 66.5 / R6 60.3.

Official axis (BabyLM 2026 evaluation; "pending" = not yet measured)

model	params	scoring	BLiMP	BLiMP-supp	EWoK	Entity	COMPS	GlobalPIQA	GLUE	Reading	AoA	Overall
mono (paper baseline)	27.4M	pre-filter Entity standard [c]	64.35	58.55	pending	27.82	51.00	pending	pending	pending	pending	pending
mono	23.9M	local zero-shot, single seed [b]	65.35	58.17	51.32	21.16	51.55	pending	pending	pending	pending	pending
mono	27.4M	local zero-shot, single seed [b]	64.35	58.55	50.70	19.24	51.00	pending	pending	pending	pending	pending
bind1 (leaderboard, server-scored)	24.0M	official server [a]	65.83	54.18	51.08	18.92	51.19	34.21	61.22	9.59	0.00	38.12
bind1	23.9M	local zero-shot, single seed [b]	65.50	58.35	51.57	19.22	51.11	pending	pending	pending	pending	pending
bind1	27M	local zero-shot, single seed [b]	66.68	60.90	51.90	20.00	51.36	pending	pending	pending	pending	pending
bind2_0	23.9M	local zero-shot, single seed [b]	66.11	58.11	51.95	19.02	51.49	pending	pending	pending	pending	pending
bind2_0	27M	local zero-shot, single seed [b]	65.14	60.81	51.16	20.53	50.89	pending	pending	pending	pending	pending
bind2_1	—	official eval pending	pending	pending	pending	pending	pending	pending	pending	pending	pending	pending
bind2_1e	27.8M	boxes-diet transfer probe, single seed [d]	n/a [d]	n/a [d]	n/a [d]	93.96 [d]	n/a [d]	n/a [d]	n/a [d]	n/a [d]	n/a [d]	n/a [d]
bind2_1e-mixed (35/65)	27.8M	mixed diet, 30M, single seed [e]	59.92 [e]	57.96	50.24	69.95 [e]	50.01	pending	pending	pending	pending	pending
mono (matched-diet control)	27.4M	boxes diet, single seed [f]	n/a	n/a	n/a	50.57 [f]	n/a	n/a	n/a	n/a	n/a	n/a

• [a] Server-verified 2026-07-10; the leaderboard "Reading" aggregate is 6.46 (self-paced 3.34 / eye-tracking 9.59); the 9.59 cell above is the eye-tracking component as tabled in the source.
• [b] Local run of the official strict-small zero-shot pipeline, single seed, not submitted to the leaderboard; tasks the local pipeline cannot produce remain pending.
• [c] The paper-baseline Entity 27.82 uses the old pre-filter entity standard; the same checkpoint re-scored under the post-filter standard measures 19.24 — the two Entity columns are not directly comparable across scoring standards. (This scoring change is exactly the bind1 refutation story in Stage 1.)
• Reference point: GPT-2 strict-small baseline BLiMP = 65.08.
• [d] bind2_1e is a mechanism-transfer probe, not a general LM entry: trained solely on a synthetic distribution-matched box-tracking corpus (30M tokens, single seed, free routing at test). Its Entity cell covers the regular+move_contents subsets only (6,259/9,483 items; the ambiguous-reference subset is excluded by the frozen pre-registration — outside mechanism scope); other suites are expected ≈ chance by design and are not measured or claimed ("n/a"). Pre-registered decomposition: shortcut-unsolvable subset 92.39, empty-box gold 97.55, learned-routing gap 0.000 (the decomposition and per-item results are on record internally; withheld at this stage). Same-scorer baseline reruns (same subsets, same script) will be tabled as they land.
• [e] bind2_1e-mixed = the same architecture retrained on 35% box-tracking + 65% natural text (30M tokens, single seed), so its BLiMP/supp/EWoK/COMPS cells are real measurements. BLiMP 59.92 sits 0.08 below the frozen no-tax gate of 60.0 → recorded TAX-OR-BUDGET, attribution open (the natural-text baselines above trained on 150M tokens — 5× this run's budget); neither "tax" nor "no tax" is claimed. The Entity cell 94.54 is the mechanism-caliber one-shot score (regular+move_contents subsets, per-item records); under the official-pipeline caliber — the official 2026-07-12 exam, all three subsets including ambiguous-reference — the same run measures 69.95 on the official exam, versus the natural-text baseline's 19.24 (≈ chance). The two calibers are materially different and are never interchangeable. Decomposition: shortcut-unsolvable 92.55/89.92 (two frozen conventions), empty-box gold 98.94, routing gap 0.000.
• [f] mono matched-diet control = a standard attention-only transformer trained on the identical box-tracking diet (same budget, single seed). Pooled 50.57, but on the shortcut-unsolvable subset it scores 23.22 ≈ chance (20) versus bind2_1e's 92.55 on the same subset; its overall half-score is carried by the empty-box prior (nothing-gold 99.89) and the last-touch shortcut. This closes the architecture-vs-diet confound: a standard architecture on the same diet floors on the mechanism-requiring items. Single seed = provisional.
• Under the current (filtered) Entity standard, entity tracking is ~chance for all natural-diet grid models — the spread across architectures on the comparable local-zero-shot rows is within single-seed noise; no architecture wins the official exam from natural-text training alone, which is exactly the Stage-3 "no tax, no win" point. The bind2_1e rows are a different kind of entry — a diet-matched transfer probe (see [d]/[e] and Stage 5) — and do not overturn that point.

Repo map

repo	what it is
SecludedCorner/bind1-babylm2026-strict-small	bind1 entry weights + growth-checkpoint branches (cited by our workshop paper; unchanged). Dual scoring (pre/post-filter) published on its card
SecludedCorner/bind1-babylm2026-strict-small-r2	bind1 retrain/resubmission line, re-collated under the revised entity-tracking standard
SecludedCorner/bind1-babylm2026-ablations	bind1 ablation checkpoints
SecludedCorner/bind1-babylm2026-eval-artifacts (dataset)	bind1 evaluation artifacts, incl. the falsification-analysis data
SecludedCorner/bind2_0	bind2_0 weights + code; main = 23.9M, branch 27m = 27M
SecludedCorner/bind-evolution (this repo)	the narrative hub; navigation and story only

Not released: bind2_1 weights/code/config (see Stage 4 for why); bind2_1e weights, code, configuration, and its synthetic training corpus and generator (withheld at this stage — see Stage 5); the synthetic diagnostic corpora and their generators; and the probe harness. The numbers above are complete as reported; what is withheld is tooling and weights, not results.

How to cite

Cite member repos, not this hub, and always pin to a 40-character commit SHA (revision="<full-sha>" in from_pretrained, or the /tree/<sha> URL form). The authoritative SHA for each published artifact is stamped on that repo's card as a dated addendum at publish time. This hub's prose may be updated as research progresses (updates are dated); member-repo cards are frozen at publish, which is why they — at a pinned SHA — are the citation targets.