Fixed methods. New recordings.

What was actually measured

All 20 pinned recordings qualified, totaling 151.848631 source hours. The fixed first-four-hour candidate grid contains 9,600 thirty-second blocks. Quality rules passed 4,117; time-based selection retained 174. Source hours, selected support and independent people are different denominators.

The new-person cohort fails the frozen minimum of three complete participants. Its selected waveforms remain in the descriptive event ledger, but no model effects or primary tests are manufactured for that cohort. New-session evidence concerns new nights from development participants; it is not new-person replication.

The complete eight-test family

Geometry and change effects compare the original-minus-independent-phase event-to-encoder correlations. Fixed within-night medians are combined across both required nights, then tested at participant level. Missing results remain in the eight-test correction family.

With at most six complete people, the smallest possible two-sided exact p is 2/64 = 0.03125; multiplying by eight gives 0.25. This design cannot establish adjusted significance even with maximally consistent effects. Effect direction, magnitude, exclusions and replication limits are the useful record; thousands of windows do not create thousands of independent people.

Captured evidence

The ledger contains 174 completed blocks and 4,002 event partitions. The pinned model runs are codebrain: 650 forwards (130 blocks × 5 controls); cbramod: 650 forwards (130 blocks × 5 controls), on the eligible new-session cohort. Selected inputs, masks, pinned model output hashes, partition membership and cohort gates remain linked to the frozen source.

Full results and exclusions as JSON. The recorded-ledger replay checks every partition and recomputes comparisons and the scientific summary. It does not regenerate detectors or repeat neural inference.

A separate 12-contact technical check

EESM19 uses only the first 30 seconds of nights 005 and 006 from one source participant, all 12 true ear contacts and the documented validity masks. This check is descriptive and outside the eight primary endpoints. Four-channel encoders are not mapped onto these 12 contacts. Physiological calibration, reference equivalence and transfer to physical Neurable hardware remain unknown.

Preserved failures and limits

Qualification initially exceeded the unchanged 2 GiB memory cap. The first cleanup correction also exceeded it. An independently reviewed sequential-process correction retained the same decoder and scientific rules; all 20 sources subsequently qualified. The failed attempts and technical exposure history are preserved. There was no post-access threshold tuning or source replacement.

Waveform events can reflect reference, sensor and artifact structure. Encoder pretraining overlap is unknown. Continuous numerical embeddings are not discrete semantic tokens. Native provider monetary cost remains UNKNOWN; no paid API call was made per window.

Reproduce

Frozen protocol · Exact reproduction instructions · v0.9.0 release · Earlier Experiment 012

The data release includes the recorded ledger and source. Full raw recordings and pinned checkpoint files remain upstream; the latter are declared hash-checked verifier dependencies even for replay.