01 // Abstract
Can a frozen receiver read the mid-layer states of a model from another family, with no gradient training? Closed-form ridge maps, fit on states from training cards, send Phi-4-mini layer-16 states into the input embeddings of Qwen3-1.7B, with one map for each sub-token position. On 200 test cards whose names never occur in training, the receiver scores 98.0%, against 99.9% for the text card and 12.0% for shuffled states. Across five pairs from three families (Phi, Qwen, Granite), cards score 90.0% to 98.5%; on SQuAD, three pairs keep 78% to 87% of the text F1 gain, and two keep 36% and 12%.
One shared map gives 89.8%, because the two tokenizers split numbers differently. A separate map for each sub-token position reads each digit out of the sender’s multi-digit token, and codes rise from 67.0% to 100%. Averaging two adjacent vectors drops the score to 31.8%, so the receiver needs one vector per token. At this width the channel works as a re-tokenization of the text, not as compression.
The closest prior work found is StateBridge (arXiv 2608.13317), which aligns final-layer states into the input space of the same model; it lists different models as future work. The repo’s literature check (2026-10-11) found no closed-form map from one family’s states into another family’s input embeddings.
Limits: the record is a working log with no preregistration file and no independent audit script. The decision rule was written before the run, but the set of pairs changed. The main result uses synthetic cards.
02 // Gates
| Gate | Preregistered condition | Result | Record |
|---|---|---|---|
| Channel check | Text card near the ceiling and no card near chance (12.5% for 8 choices), so that any gain must come through the channel. Measured: text 99.9%, no card 11.6%. | [ PASSED ] | c3c5ecfd74a8… |
| Cards, five pairs | First part of the rule written before the run: at least 3 of 4 pairs reach 90% on cards. Measured: all 5 pairs, 90.0% to 98.5%; shuffled 10.0% to 12.6%. | [ PASSED ] | c3c5ecfd74a8… |
| SQuAD, five pairs | Second part of the same rule: at least 3 of 4 pairs keep 75% of the text F1 gain. Measured: 82%, 87% and 78% in three pairs; 36% and 12% in two. The planned SmolLM3 pair did not run, and two Granite pairs ran in its place, so the rule cannot be applied as written. | [ OPEN ] | c3c5ecfd74a8… |
03 // Destructive controls
| Arm | What it removes | Result |
|---|---|---|
| Shuffled | The card. The receiver gets the mapped states of another card. | 12.0% (text 99.9%) |
| No card | The whole message. | 11.6% (chance 12.5%) |
| One shared map | The split by sub-token position. One map serves every receiver token. | 89.8% (codes 67.0%) |
| Mean pooling of 2 | One vector per token. Two adjacent vectors are averaged. | 31.8% |
04 // Record and artifacts
The preregistration, raw outputs, audit and code of this entry are not public yet. They will be linked here.