01 // Abstract
This line asks how robust the R.01 channel is when its latent wire changes in transit. The tests use the sealed two-bit channel of R.01 (Qwen3-1.7B to SmolLM2-360M-Instruct). An attacker with full access to the models adds an optimized change to the 8 slots of the wire, within a budget, and tries to force the next symbol as the answer. The budget is a fraction of the slot norm. Each test compares the attack with five random changes of the same norm, because that difference does not depend on the chance level.
Below a budget of 0.12, the attack does not flip the answer: target success is 0.000 at 0.06 and 0.023 at 0.12, and the receiver keeps the true answer. At 0.25, the attack flips 53.1% of the answers, and attack minus random is +0.258 (95% CI 0.192 to 0.323). A second, balanced set gives the same pattern: 0.047 at 0.12, and 0.531 at 0.25 with attack minus random +0.275 (95% CI 0.238 to 0.325). Thus the threshold is between 12% and 25% of the size of the signal itself. Random changes and a replayed valid wire do not force the target; only the optimized direction does.
Limits: these results are exploratory. The preregistered sanity gate G1 failed, because the best attack reached 0.812, not 99%. Most threshold runs used 60 steps and 1 restart on 128 rows, not the preregistered 200 steps and 10 restarts. The first set is unbalanced, so its chance is 0.281 (majority) or 0.255 (by frequency), not 0.25. The second set is balanced (chance 0.25), but its run files do not record its seed. Three planned controls (zero wire, receiver only and sign flip) give a constant answer, so they are not controls. The confirmatory run is preregistered and is not run yet.
02 // Gates
| Gate | Preregistered condition | Result | Record |
|---|---|---|---|
| G0 Reproduce the channel | The sealed channel on the laptop GPU (MPS): correct at least 99%, controls within 24–27%. Measured: correct 1.00, controls about 0.25. | [ PASSED ] | 3610cc681fdf… |
| G1 Attack sanity | An unconstrained attack on all 8 slots reaches at least 99% target success. If not, the attack loop is not trusted. Best measured: 0.812, at budget 2.0 with 150 steps and 4 restarts. | [ FAILED ] | 07aa0df63692… |
| Confirmatory run | Preregistered on 2026-10-19: a new nonce seed, 96 balanced rows, 200 steps and 10 restarts. Threshold claim: attack minus random at budget 0.25 has a 95% CI above 0. Not run yet. | [ OPEN ] | 44def7041899… |
03 // Destructive controls
| Arm | What it removes | Result |
|---|---|---|
| Correct wire | The attack. The wire goes to the receiver unchanged. | Target 0.00; the receiver gives the true answer |
| Random change ×5 | The optimized direction. Five random changes with the same norm as the attack. | 0.25–0.29 at budget 0.25 (attack 0.53) |
| Benign replay | The optimization. A valid wire from another input goes in its place. | 0.23 and 0.34 at budget 0.25 (attack 0.53) |
04 // Record and artifacts
The preregistration, raw outputs, audit and code of this entry are not public yet. They will be linked here.