API Reference

Scanning

Mine feedback events out of transcripts, incrementally.

scan.scan(): Scans transcripts under roots for feedback, incrementally.
scan.ScanReport: The outcome of one scan pass.

Detection

The signal detectors and the candidate shape they emit.

detectors.detect(): Runs every detector over one transcript’s events.
models.FeedbackCandidate: A single piece of developer pushback extracted from a transcript.
models.dedup_key(): Returns the stable dedup key for parts.

Storage

The feedback event store.

store.FeedbackStore: Persistent store for collected feedback over the native mining engine.

Retraining

The weekly retrain lanes — gate and watcher — and their promotion machinery.

retrain.lexical.retrain_gate(): One gate retrain pass; returns the journaled one-line verdict.
retrain.watcher.retrain_watcher(): One watcher retrain pass; returns the journaled one-line verdict.
retrain.watcher.WatcherRecipe: Every knob of one watcher LoRA retrain, validated at parse so no degenerate value trains.
retrain.watcher.register_watcher_adapter(): Register (and by default promote) a built mlx-lm adapter as the watcher component.
retrain.watcher.seed_incumbent_probs(): Validate an external incumbent probs cache against the frozen frame, then write it through the store.
retrain.evalset.freeze_eval(): Copy the exported <view>/test.parquet into the eval root and merge its sha256 manifest.
retrain.evalset.freeze_steer_type(): Build then freeze the steer-type eval; returns the frozen file’s sha256.
retrain.evalset.freeze_pick(): Build then freeze the pick-prediction eval; returns the frozen file’s sha256.
retrain.evalset.EvalFrame: The frozen watcher eval as the arrays the promotion gate reads.
retrain.evalset.SteerTypeFrame: The frozen steer-type eval as the arrays a category classifier is scored on.
retrain.evalset.PickFrame: The frozen pick-prediction eval as the arrays an option classifier is scored on.
retrain.promotion.corrected_gate(): Evaluate the corrected paired gate over common rows at matched budget.
retrain.promotion.GateResult: The corrected paired gate over a candidate and incumbent at matched budget.
retrain.promotion.Verdict: One promotion decision: whether to promote and the human-readable reason.

Many-shot watcher

Score the frozen frame through a byte-constant, cache-amortized many-shot exemplar prefix.

retrain.manyshot.build_exemplar_system(): The byte-constant many-shot system prompt: seeded demonstrations under budget_chars.
retrain.manyshot.score_frame(): Score the frozen frame with a cached many-shot prefix and store its P(NO_STEER).

Encoder gate

Fine-tune a transformer encoder as the stage-1 gate and score it paired on the frozen gate frame.

retrain.encoder.EncoderSpec: Recipe for one encoder gate arm: the base model plus its fine-tune and calibration knobs.
retrain.encoder.EncoderModel: A fine-tuned encoder head plus its fitted temperature and held-out calibration error.
retrain.encoder.train_encoder(): Fine-tune the encoder head on train_frame, temperature-scaled on a held-out val carve.
retrain.encoder.score_frozen(): Score the frozen gate eval and persist the calibrated P(fire) through write_probs.

Instrument

Paired fast-DeLong AUC comparison and the instrument-card decision rule.

instrument.delong_se(): Return the DeLong standard error of a single AUC.
instrument.paired_delong(): Compare two score vectors on identical labels with a paired fast-DeLong test.
instrument.PairedDeLong: Result of a paired fast-DeLong comparison of two score vectors on shared labels.
instrument.bootstrap_ci(): Nonparametric bootstrap 95% CI of the AUC (E27 harness resampling).
instrument.mde(): Minimum detectable effect for a two-sided z-test at the given standard error.
instrument.actionable(): Apply the instrument card’s two-part actionability rule to an AUC delta.
instrument.InstrumentCard: The measurement-instrument card: the frame noise floors comparative verdicts consult.
instrument.Comparison: A card-governed comparative verdict between two checkpoints’ AUCs on one frame.
instrument.paired_verdict(): The card rule over a paired DeLong record — the production comparative verdict.
instrument.unpaired_verdict(): The card rule when a counterpart’s per-row probs were never persisted.
retrain.sweep.ArmScore: One arm’s frozen-frame measurement: the scalar metric and the per-row probs behind it.
retrain.sweep.compare_score_reports(): Card-governed verdict between two arms’ score reports — paired DeLong as the production default.

Decisions

Mine AskUserQuestion decision rounds from transcript trees into a parquet dataset.

decisions.mine(): Mine every AskUserQuestion decision round under root.
decisions.DecisionRow: One AskUserQuestion round: the question, its options, and the user’s pick.
decisions.MineResult: The mined decision rows and the quarantined, un-typed uses.
decisions.Quarantine: An AskUserQuestion use whose result carries no typed rounds.
decisions.write_decisions(): Write the mined rows to out as parquet, stamping the digest and quarantine count.
decisions.read_decisions(): Read a decisions parquet back into rows, its digest, and its quarantine count.
decisions.stats_of(): Aggregate mined rows into the counts the decisions stats command prints.
decisions.DecisionStats: Aggregate counts over a decisions dataset.