01 // Abstract
A trained bridge of the R.01 design (Qwen3-1.7B to SmolLM2-360M-Instruct, 569,857 parameters) passes every registered gate on familiar IDs. On eight whole IDs that are absent from training, it scores 0/32, while a text handoff scores 32/32. This series of sealed tests asks why. Supervised oracle codes through the bridge’s own frozen output path reach 32/32 on the novel IDs, and no two of these IDs share a code at any stage of the bridge. Thus the reader can bind when it gets a suitable code, and identity is not erased.
The learned codes are specific to each training ID. An additive letter-plus-digit model of them has 30.5% error in sample and 51.3% when one ID is left out, and codes built from parts score 32/224 against 15/224 for a shifted-label null. More training fits all 224 training rows, but it lowers the novel IDs from 4/32 to 0/32. This matches emergent-communication results: learned codes can solve a task without being compositional (Kottur et al., 2017), and generalization pressure alone does not produce compositionality (Chaabouni et al., 2020). The later R.01 test E3 uses a factored code, and it passes.
Limits: each design has one seed, one learning rate and one bridge size. The eight novel IDs had earlier evaluation history, so they are not a pristine test. The internal path of the failure for most IDs is still open: removing at least 99.1% of one probe’s readable signal changed no outcome on 27 failed IDs.
02 // Gates
| Gate | Preregistered condition | Result | Record |
|---|---|---|---|
| Familiar IDs | Five source orders and fresh contexts, 32 queries each: native original, shuffled-donor and changed counts at least 31/31/28 (reversal 28/28/28), with all controls and paired bounds. Measured: all six passed. | [ PASSED ] | d46612aa92eb… |
| Novel whole IDs | Eight whole IDs absent from the 56 training IDs, 32 queries: the same 31/31/28 minima. Measured: 0/0/0. Text handoff: 32 original, 29 changed. | [ FAILED ] | f155c4075a79… |
| Oracle code, 40 steps | Supervised 64-value codes through the frozen output path, with labels given: novel matched and swapped-insert counts at least 28/32 in original and changed contexts; wrong and random intended counts at most 8. Measured: 26/32. | [ FAILED ] | 44daaabc1e19… |
| Oracle code, 200 steps | The same rule, in one extension registered after the 40-step run failed. Optimizer, data and thresholds unchanged. Measured: 32/32 in all three pools. | [ PASSED ] | c0282ae48521… |
| Code composition | Codes built from letter and digit parts, one ID left out at a time, 224 training rows: letters ratio at least 0.90, joint at least 3 times a shifted-label null, null letters ratio at most 0.10. Measured: 0.406, 2.13 times, 0.424. | [ FAILED ] | de4ea4594879… |
| Longer training | Continue the trained bridge from step 200 to step 600 on the same 56 IDs, with the gates of the novel whole-ID test. Measured: training rows 224/224; novel IDs 4/32 at step 200, 0/32 at step 600. | [ FAILED ] | 2549a4c8343f… |
| Held-out training IDs | Four folds; each training ID is held out once (224 rows). The selected step must beat step 0 on joint and sequential counts. Measured: step 0 32 and 32; step 100 22 and 25; step 200 17 and 20. | [ FAILED ] | 91e8458e0d08… |
03 // Why it failed
The trained bridge learns a separate code for each training ID, not a code built from letter and digit parts, and more training lowers the novel IDs further.
04 // Destructive controls
| Arm | What it removes | Result |
|---|---|---|
| Wrong code | The matched code. The oracle code of another ID goes in its place. | Follows the inserted ID 32/32; intended ID 0/32 |
| Random code | The code content. Noise with the norm of each learned value. | Intended ID 0/32 |
| Shifted-label null | The link between labels and codes in the composition test. | 15/224 joint (composed 32/224) |
| Text handoff | The latent. The sender writes the ID as text for the receiver. | 32/32 and 29/32 (latent 0/32) |
| Orthogonal erasure | An equal-size subspace outside the probe-readable directions. | No correctness change on 27 IDs, same as treatment |
| Shifted probe labels | The pairing of IDs and labels in the probe fit. | 0/56 whole IDs |
05 // Record and artifacts
The preregistration, raw outputs, audit and code of this entry are not public yet. They will be linked here.