README.md

---

pretty_name: "bind2_1e — mechanism transfer on the official entity-tracking surface (results-only)"
language:
- en
tags:
- babylm
- babylm-2026
- strict-small
- state-tracking
- entity-tracking
- research-log
- results-only
---


# bind2_1e — the binding mechanism on the official entity-tracking surface



**Results-only card. Nothing here is downloadable** — no weights, no code, no configuration, no training corpus, no generator. This card publishes *numbers, what was tested, and why it matters*, and withholds the tooling that would let the result be rebuilt. It is a member of the **bind-evolution** line — the entity-tracking evolution of the `bind2_1` binding architecture (the version suffix says so); for the full falsification timeline and how the members relate, see the bind-evolution hub.

## What bind2_1e is



bind2_1e (~27.8M parameters) is the member that carries the **causally-verified binding mechanism** of `bind2_1` onto the **official BabyLM 2026 entity-tracking surface**. It is a letter-suffix engineering extension of `bind2_1` — the same binding primitive, with a readout shaped to the official entity/box-tracking question form. At the level this line publishes, the mechanism is described only as *a binding mechanism with learned routing*; the "how" is withheld.

The mechanism's pedigree comes from the `bind2_1` campaign, which was pre-registered and frozen before any data existed. There the binding pathway was shown to be **causally load-bearing**: lesioning its state pathway removes ~99.3% of the deep-tracking advantage, and interchange patching flips ~96.4% of answers to a donor context's holder — the full system scoring roughly 69 points above every matched control. That campaign's frozen pre-registered verdict was **NULL** (a depth-interaction criterion that a later adversarial autopsy found ceiling-confounded), and its honest headline finding was **limited effective depth**. A separate parameter/scale scan across a ~6× parameter ladder (24M → 145M) later showed that the small-model form-change is **not** dissolved by adding parameters. bind2_1e is the follow-on question those results earned: **does the mechanism actually transfer to the official exam when the interface fits?**



## What was tested, and why



Earlier in this line, `bind2_0` established **"no tax, no win"**: the binding architecture, trained as a plain LM on the real strict-small corpus, showed no zero-shot state-tracking advantage on the official exam — mechanism capability and benchmark transfer are *separate questions*, and conflating them is how a field fools itself. bind2_1e asks the transfer question the right way: train the causally-verified mechanism on a training surface it can actually use, then score it on the **official entity-tracking items with fully learned routing** — no test-time oracle assistance — and read the first checkpoint, once.



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



## Results



Two runs were scored. All entity figures are accuracy (%), chance ≈ 20%; every earlier model in this line and the published baselines sit at ≈19–21% on this benchmark.



### The headline (official exam, version-consistent)



On the **official 07-12 evaluation tree** — the same exam version scored for both models, so the comparison is apples-to-apples — the mixed-diet run scores:



> **entity_tracking = 69.95** vs the monolingual baseline's **19.24**



That is the ~50-point lift the mechanism buys on the official surface, measured under the official scorer across all three official subsets. It is the version-consistent headline: both numbers come from the same benchmark version, so the gap is not an artifact of comparing across scoring standards.



### Mechanism-caliber transfer



Restricted to the two in-scope subsets the mechanism targets (regular + contents-move, 6,259 of 9,483 items; per-item scorer):



| run | training diet | mechanism-caliber entity | note |

|---|---|---|---|

| boxes-only | synthetic box-tracking only | **93.96** | replicated to n=3 seeds: 93.96 / 94.12 / 94.22 (spread 0.26 pt) |

| mixed | 35% box-tracking / 65% natural text, 30M tokens | **94.54** | + real general-language scores (below) |



**Why 69.95 and ~94 are both true, and not interchangeable.** The **official caliber (69.95)** spans all three official subsets — including the **ambiguous-reference** subset, which the frozen pre-registration placed *outside* mechanism scope — and uses the official scorer. The **mechanism caliber (~94)** restricts to the two in-scope subsets. The two calibers are materially different and are never swapped for one another; both are reported here so neither over- nor under-states what was shown.



### The result is not a shortcut artifact



The pre-registered decomposition is the credibility core. On the items the benchmark's strongest audited shortcut heuristic **cannot** solve, the model still scores **~92.4–92.6%**; the empty-container "gold = nothing" trap — the exact state-blind prior that produced this line's earlier self-refutation — scores **97.6% (boxes-only) / 98.94% (mixed)**; and the learned router matches oracle-hinted routing to the third decimal (**routing gap = 0.000**). The score is carried by state tracking, not by a positional or prior-abuse shortcut.



### The architecture, not the diet



A standard attention-only transformer trained on the **identical** box-tracking diet scores ~50.6% pooled — but only **23.2% ≈ chance** on the shortcut-unsolvable subset, versus this model's **~92.6%** on that same subset. On the same diet, a standard architecture floors on the items that require tracking. (Single seed — provisional.)



### General-language ability survives the mixed diet



The mixed-diet run trained on 65% natural text, so its general scores are real measurements, not "n/a":



| suite (official 07-12 tree, single seed) | baseline (27.4M) | bind2_1e-mixed |
|---|---|---|
| entity_tracking | 19.24 | **69.95** |

| BLiMP | 64.35 | 59.92 |

| BLiMP-supplement | 58.55 | 57.96 |

| COMPS | 51.00 | 50.01 |

| EWoK | 50.70 | 50.24 |



BLiMP 59.92 sits 0.08 below a frozen no-tax gate of 60.0 → recorded **TAX-OR-BUDGET**, with attribution left open between a possible architecture cost and this run's 5×-smaller token budget (30M vs the baselines' 150M); neither is claimed. COMPS and EWoK sit near chance for both arms (difference not meaningful). On the official **GLUE fine-tuning** protocol the mixed-diet run wins **5 of 7 tasks** over the monolingual baseline (MNLI, MRPC, MultiRC, QQP, RTE; baseline higher on BoolQ and WSC). So the mixed run is a competitive general model that *also* carries the tracking mechanism — not a single-trick specialist.



## Why this matters



`bind2_0` left the transfer question open with an honest negative ("no tax, no win"). bind2_1e closes the positive half of it: **the binding mechanism demonstrably learns the hard structured-tracking task once the interface fits** — ~94% at mechanism caliber, and a version-consistent ~50-point official-exam lift over the baseline (69.95 vs 19.24) — while the mixed-diet model keeps real, competitive general-language ability. The earlier gap was a **missing interface, not a disproven mechanism**. That converts an open hope into a bounded, measured result.



## What this is not (four mandatory caveats)



1. **Single seed is single seed — provisional until replicated.** The boxes-only mechanism-caliber headline replicated across three seeds (93.96 / 94.12 / 94.22); the mixed-diet run and the matched-diet control are each single-seed and remain provisional.

2. **Subset scope.** The mechanism-caliber score covers the **regular + 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.

3. **This is a mechanism-transfer demonstration, not a general language model.** The boxes-only run was trained solely on a synthetic box-tracking corpus; every other suite in the official evaluation is expected to sit at chance **by design**, and no claim is made there. (The mixed-diet run's general scores are real, but come from 5× less training data than the baselines.)

4. **The natural-diet baselines are diet-confounded as an architecture comparison.** Their chance-level entity scores come from natural-text training; the matched-diet architecture attribution rests on the `bind2_1` campaign's matched controls plus the dedicated matched-diet control reported above — not on the natural-diet rows.

## What is not released

At this stage, **everything buildable is withheld**: weights, code, configuration, and the synthetic training corpus and its generator. The single honest disclosure about that corpus: it is a **synthetic box-tracking corpus whose answer statistics are distribution-matched to the official entity-tracking benchmark**. The generation method is not published. The frozen pre-registration, the full decomposition, and the per-item results are on record internally; the numbers reported here are complete as reported — **what is withheld is tooling and weights, not results**.

## Citation and context

This is a member card in the **bind-evolution** line, whose value is the sequence of pre-registered questions and frozen verdicts, not any single score. For the full timeline — including this line's refutation of its own earlier published result — see the bind-evolution hub. Cite this member at a pinned commit revision as stamped on this card at publish time.