Two days of ring-logic experiments, including everything that failed

Working sessions on a laptop CPU, no GPU, starting from a handoff of earlier runs. Each step below has its finding filed separately.

What was tried, in order

1. Reproduce the handoff. The fast build matched the earlier result files byte for byte; the slow reference build agreed on every noise-robust count.
2. Swap in the real write instruction. The earlier runs used a ±1 write; the machine's own instruction set only adds. Coverage fell from 256 to 232 functions and every late-selector multiplexer disappeared.
3. Try to win them back with lenses. About 29,000 one-, two- and three-lens configurations; the best reached 21 of 32 noisy trials per step, at isolated settings.
4. Two register rings and an output ring. Site-addressed copy gates appeared, then failed when run end to end: one winding triple hides several attractors.
5. Read a phasor instead of a winding number. 8 of 8 multiplexers instead of 4, but no better under noise, because the slip-based operations were the failure.
6. Make the inputs continuous too. Frequency drives in one basin: the multiplexer in one settle, 100% under the same noise, a few percent lost to continuous noise.

Retractions

• 256 of 256 functions for one ring: true only with a subtracting write the instruction set does not have.
• A perfect three-lens multiplexer: an artefact of the fast build's early lock test, which declared a drifting state settled. It ran at 62.5% with no noise.
• Fast-build entries for writes from |q| = 3: the coarse step folds the ring back to q = ±1; the reference build shows no valid state.

Decisions

• Every program a table predicts is confirmed by running it kick by kick before it is reported. The failures in steps 3 and 4 were invisible in the tables and obvious in the runs.
• Lock means velocity lock whenever a ring carries a lens or a coupling; the early-exit test stays for pristine rings only, where it is exact.
• Stop treating the winding number as the state. The inference engine never used winding numbers; once rings are coupled they are not even a complete description. What carries the computation is the continuous phase configuration.

Open

A machine that keeps state between steps; continuous-time noise on a physical reference phase; and whether stronger coupling or more rings move the slip edge away from the drive strengths the multiplexer needs.

Everything behind these notes — instruments, raw outputs, the full results — is in research/forks, with a README mapping each note to its files.