01 // Abstract
Can a small trained bridge carry a message from a frozen sender to a frozen receiver of another family? The channel sends Qwen3-1.7B states through a bridge of 569,857 parameters into SmolLM2-360M-Instruct. On a synthetic two-bit task it passed Gates 0–4, including a held-out cross-family test on 1,024 new rows: correct 100%, shuffled 26.37%, random 25.49% and 25.68%, zero 25.00%. A text handoff also scores 100%, so the channel shows no accuracy advantage over text.
Two earlier attempts failed: the first architecture (channel v1) failed Gate 1, and a Llama-3.2-1B receiver failed the tiny gate. Binding is harder. Binding gate V1 failed (best 14/32, against a tiny gate of at least 95%); R.06 records why. Test E3, with a factored letter, digit and color code, passed on 56 never-seen IDs (93.8%), and its free-form arm failed (14.3%). The receiver itself reads the table as text at only 25–30%, so in E3 the trained reader and bridge do the binding, not the receiver.
02 // Gates
| Gate | Preregistered condition | Result | Record |
|---|---|---|---|
| Channel v1, Gate 1 | First architecture, one seed, 300 steps. Tiny set (32 examples): correct at least 31 of 32. Measured: at most 25.00%. | [ FAILED ] | 8d500bfebdf6… |
| Llama receiver, tiny gate | Qwen3-1.7B to Llama-3.2-1B-Instruct, 32-example overfit gate. Measured: failed. The receiver also failed the oracle text-handoff check, so the cause is not isolated. | [ FAILED ] | a93cbe3a2f7e… |
| Gate 0 | Mechanics: hook-disabled and zero-memory logits equal stock logits bit for bit; a nonzero memory changes the output; bridge gradients are finite and nonzero; endpoint gradients are absent; a repeat run is equal. | [ PASSED ] | 8be13af87f99… |
| Gate 1 | Tiny set (32 examples): correct accuracy ≥ 95% (at least 31 of 32). | [ PASSED ] | 8be13af87f99… |
| Gate 2 | Tiny set: shuffled and each random arm ≤ 50%; zero = receiver-only = 25%. | [ PASSED ] | 8be13af87f99… |
| Gate 3 | Held-out set, same family (Qwen3-1.7B to Qwen3-0.6B): correct ≥ 90%, each class ≥ 85%, receiver-only 25%, shuffled and random ≤ 35%, correct minus each control above +0.50 (99% lower bound). | [ PASSED ] | b268b1a8472e… |
| Gate 4 | Held-out set, cross family (Qwen3-1.7B to SmolLM2-360M, 1,024 new rows): the Gate 3 rule, unchanged. | [ PASSED ] | 18e4a0f5df4b… |
| Binding gate V1 | Same bridge, queried ID-color bindings. Tiny set (32 queries): correct ≥ 95%, shuffled and random ≤ 25%, donor following ≥ 95%. Then ≥ 90% on known IDs and on novel IDs. | [ FAILED ] | 5a6f3408a767… |
| Binding E3 | Factored letter, digit and color code; 56 never-seen IDs (224 rows): joint accuracy ≥ 75%, changed-answer following ≥ 75%, each control ≤ 10%. | [ PASSED ] | 987dc2f8d726… |
03 // Destructive controls
| Arm | What it removes | Result |
|---|---|---|
| Shuffled | The message of the row. Each row gets the latent of a different row (label-blind derangement, seed 61001). | 26.37% |
| Random | The content. Gaussian memories with the same per-slot norm (seeds 71001 and 71002). | 25.49% / 25.68% |
| Zero | The whole message. The memory is all zeros. | 25.00% |
| Text-handoff | The latent. The symbol of the sender goes to the receiver as compact text. | 100% (latent 100%) |
04 // Record and artifacts
The preregistration, raw outputs, audit and code of this entry are not public yet. They will be linked here.