Algorithm

Drift Detector uses a three-layer pipeline: a deterministic lexical/structural point engine, a trajectory layer that classifies drift over time, and a real-corpus ExtraTree classifier + DCD pipeline for production eval.

Point engine (src/lib/drift_score.py)

Answers: "how bad is this single turn?" Returns a 0–100 score.

Pipeline

1. Strip code. Fenced and inline code blocks removed — prose-only scoring, so emitting normal code never counts as drift.
2. Tokenize. Lowercased, accent-stripped word tokens; sentences split on .!?.
3. Component scores (each normalized 0..1):

Component	How measured
hedges	Marker density via saturating curve 1 - e^(-rate×18)
filler	Same saturating curve — one stray filler barely moves score; marker-saturated reply hits ceiling
hype	Positive hype vocabulary density
meta	Meta-narration markers ("I'll now explain…", "Let me walk you through…")
verbosity	Words-per-sentence ramped between profile's target_wps and max_wps
length	Prose word count ramped between target_words and max_words
complexity	Fraction of long words (≥ long_word_len) up to max_long_ratio

1. Noisy-OR aggregation. One strong signal carries; many weak ones accumulate. Weighted by the active profile's weights map.
2. Polysemous corroboration gate. Polysemous words (leverage, robust, etc.) only count toward a class when another lexical marker fires or real length pressure exists — eliminates technical-prose false positives.
3. Structural-only damping. When total lexical hits == 0, structural components (verbosity, length, complexity) are scaled ×0.8 — dense jargon without hedges/filler is not drift.

Profiles

A profile defines the contract: threshold (0–100), sensitivity multiplier, per-component weights, custom lexicons, and calibration parameters. See Profiles for authoring.

Trajectory layer (src/lib/drift_trajectory.py)

Answers: "is this session drifting, adapting, or at a chronic subclinical level?"

Takes a sequence of per-turn scores and classifies:

• Adaptive drift — isolated spike that immediately falls back. System was self-correcting; no nudge fired.
• Degenerative drift — sustained climb, plateau near threshold, or repeating-spike cycle. Correction nudge fires.
• Chronic subclinical — narrow band of scores parked just under threshold across multiple turns. Flagged even though no single turn crosses threshold.
• Repeating-spike — session that alternates above/below threshold ≥3 times. Adaptive turn-by-turn but degenerative as a cycle; bypasses the adaptive gate.
• Velocity alarm — rate of score increase over recent turns exceeds a per-profile ramp threshold, catching slow creep before it crosses threshold.

Why trajectory matters

A single-turn snapshot lies about dynamics. A session can score 85 once and recover (not drift) or score 65 consistently (chronic subclinical drift). The trajectory layer is what keeps false positives at zero: corrections fire on patterns, not on individual twitchy turns.

ExtraTree classifier (scripts/backtest_real.py)

43-feature ExtraTree (n_estimators=500, max_depth=10, class_weight='balanced'), GroupKFold(5) by session_id, threshold t=0.58. Features include:

• Point scores, trajectory flags, drift rate, max streak
• classify_user_reply() labels on following user messages (correction_style, correction_substance, frustration, approval, continuation, new_task)
• Turn position, burst length, interrupt flags

DCD pipeline (Deferred Correction Detection)

Scans N+1…N+10 turns following a candidate for user correction signals. Critical for recovering cases where user feedback is delayed (e.g., continues working for a few turns before explicitly correcting).

steps=10, t=0.58 — tuned on the real corpus to maximize F1 without introducing false positives.

Eval results

Metric	Value
F1	0.9973
Precision	1.0000
Recall	0.9947
tp	375
fp	0
fn	2
tn	906
Corpus size	1,283 labeled entries

Two irreducible FNs: one credential provision in ok context (3 overlapping ok patterns prevent safe classification) and one bare "Try now" indistinguishable from ok entries without session context.

See Eval for the full 21-round scientific-method campaign.