Shared memory and context tools for agentic work.
Code Rooms
# Learning and calibration time units
The clock repair and v1 calibration containment shipped in 1.6.5 at
`363bc97f2336c73e90802916804dccd2109c5746`. The relevance-only event separation
described below is **Unreleased** PR580 behavior on `main`, absent from sealed 1.6.5.
The envelope calibration path consumes defect history recorded in `TrustLedger`.
The direct `TrustLedger::record_defect` producer remains supported; generic
`learn(correct|wrong)` now supplies relevance feedback only and writes no defect
history. See [the event boundary](LEARNING-EVENT-SEMANTICS.md).
`TrustEntry` defect timestamps and `TrustLedger::compute_trust` use Unix seconds;
`CalibrationRow::calibrated_at_ms` and historical seek calibration receipts use Unix
milliseconds.
The calibration core accepts one integer millisecond timestamp, converts it to
fractional seconds for trust decay, and preserves the original integer in the
row. A fresh defect therefore keeps its recent-defect weighting; the default
30-day half-life is measured in days, not accidentally in milliseconds treated
as seconds. The existing synthetic corpus also records seconds, matching the
direct defect producer.
## Regression contract
- A fixed corpus with three defects and one false alarm per node produces
trust-band reliability 0.2 immediately, 0.5 after 30 days, and 0.8 after
120 days. At 30 days its recency factor is 0.5 and trust score is 0.5125.
- A direct `TrustLedger::record_defect` event followed by envelope calibration
retains the fresh-defect band. This replaces the earlier generic
`learn(correct)` producer regression, whose defect attribution is no longer
permitted.
- Calibration retains exact integer epoch milliseconds, including subsecond
digits, in the row; persistence and reload retain that same row.
These are clock and persistence regressions, not quality benchmarks.
## Compatibility and limits
The persisted ledger and calibration formats are unchanged. Existing rows are
not rewritten on load. A subsequent authorized calibration recomputes from the
recorded ledger using the corrected clock; this change does not add a public
calibration or learning consumer or relax an authority floor.
The clock repair itself left feedback vocabulary, label interpretation,
and the threshold algorithm unchanged. The subsequent event-boundary correction
separates relevance from defect evidence without changing the seconds/milliseconds
contract. Correct clock units do not establish
that relevance feedback is a verified defect, that the corpus is independently
held out, or that a calibration threshold satisfies a statistical error budget.
The subsequent [calibration qualification containment](CALIBRATION-QUALIFICATION.md)
keeps all v1 rows structurally readable but marks their runtime evidence
`legacy_evidence_unqualified`. Prediction abstains, and the seek envelope stays
at `reverify` or `unprovable`; running a diagnostic calibration does not enable
`act`. Scores, ranking, persistence, and historical receipt digest compatibility
are retained. Current envelopes emit no calibration receipt from these rows.
Continue checking code and executed evidence.