 ruv
|
192b4c6b4a
|
fix(ruview-gamma): safety becomes a property of construction (safety review, 2 critical + 3 major)
- safety stop is now control flow: run_calibration terminates the sweep on
any stop (partial calibration recorded)
- cross-session terminate-and-lock: persisted latched ParticipantSafetyState;
locked participants refuse sessions across governor instances until
unlock_with_acknowledgment writes an audit record; documented class->lock map
- SafetyEnvelope safe-by-construction: private fields, validated try_new/with_*
constructors, serde routed through the validator, compiled-in absolute bounds
(floor >=30Hz keeps the whole band above the 15-25Hz photosensitive zone;
caps <=0.6; <=30min) — an 18-22Hz config fails to deserialize
- dose governance: <=4 sittings/24h + >=60min cooldown from the persisted ledger
- per-tick latched SafetyMonitor wired into run_session (simulator emits ticks;
mid-session latch truncates delivered duration); clean-session witness
byte-identical (pinned proof unchanged)
- cohort warm-start k-floor (>=3 profiles)
ADR-250 + crate README updated. 125+13 tests green.
Co-Authored-By: claude-flow <ruv@ruv.net>
|
2026-06-11 15:26:17 -04:00 |
|
|
Claude
|
3408883f79
|
feat(ruview-gamma): implement ADR-250 Adaptive Gamma Entrainment crate
Governed, deterministic, safety-constrained personalization of 40 Hz-prior
light+sound stimulation (RuView sensing + RuVector response modeling + RuFlo
audit). 40 Hz is the starting prior, not the hard-coded answer.
11 modules: stimulus/SafetyEnvelope, safety (exclusion screen + latched
monitor), response (20-field PersonResponseVector + optional EEG), objective
(safe-entrainment score, safety as hard gate), simulator (deterministic
ChaCha20 frequency_response_curve), optimizer (calibration sweep + GP/EI +
closed-loop), bandit (LinUCB), session (reproducible SHA-256 witness), ruflo
(consent/exclusion/envelope/audit/trial-sham/clinician-export/claim-discipline),
proof (deterministic bundle witness), math.
Safety invariant asserted in tests: no emitted stimulus can ever leave the
SafetyEnvelope; non-finite inputs clamp to the conservative floor. Claim
discipline: only 'personalized entrainment optimization', never Alzheimer's
treatment (ADR-250 §19). Standalone leaf, publish=false pending safety sign-off.
Validation: 64 unit/integration + 1 doctest pass; full workspace gate green
(2,862 passed, 0 failed); deterministic witness pinned; criterion benches
(safety-stop tick ~9.3 ns vs the ADR §17 500 ms bound).
Adds ADR-250 doc, registers the crate, updates CLAUDE.md crate table and
CHANGELOG.
https://claude.ai/code/session_01MjBucx95K4BuUxZi8NWwRH
|
2026-06-10 03:30:28 +00:00 |
|