Per-segment adaptive thresholds

Status: core (Apache-2.0), offline, human-gated. Like the global adaptive threshold learner this only recommends — route the recommendation through change management before promotion.

The right halt threshold is not global. A clinical-domain answer and a casual chat tolerate very different false-positive rates, and two models hallucinate at different rates. SegmentedThresholdLearner wraps the global AdaptiveThresholdLearner with a per-segment routing layer, so each segment — any string key: a domain, a model id, a tenant id, or a composite — accumulates its own human-labelled evidence and earns its own threshold.

Cold start

A global pool sees every observation. A segment with fewer than promote_after (default min_samples) observations of its own falls back to the pooled recommendation, so a brand-new segment is never starved of guidance; once it has earned enough evidence, its own learner takes over.

Usage

from director_ai.core import SegmentedThresholdLearner

learner = SegmentedThresholdLearner(
    candidate_thresholds=[0.3, 0.5, 0.7, 0.9],
    current_threshold=0.5,
    min_samples=20,
)

# Replay human-labelled feedback, tagged by segment.
learner.observe(score=0.91, human_approved=True, segment="clinical")
learner.observe(score=0.44, human_approved=True, segment="chat")

rec = learner.recommend(segment="clinical")
rec.source                          # "segment" once promoted, else "global"
rec.recommendation.recommended_threshold
rec.recommendation.to_profile_overlay(profile="clinical-adaptive")

recommend returns a SegmentRecommendation carrying the source, the per-segment feedback count, and the underlying AdaptiveThresholdRecommendation (with its safety constraints, expected lift, and rollback threshold).

Measured

On a synthetic three-segment workload where each segment has a different latent approval boundary (python -m benchmarks.segmented_threshold):

Policy	Mean approval accuracy
Per-segment thresholds	0.89
Single global threshold	0.83
Segmentation lift	+0.06

The largest gain is on the strict (clinical-style) segment, where the global compromise threshold is far too lenient (0.91 vs 0.72 accuracy). A segment whose optimum happens to coincide with the global one gains nothing — segmentation helps exactly where segments diverge. Numbers come from the committed benchmark and benchmarks/results/segmented_threshold.json.

Backend

This layer adds only routing and bookkeeping; the Beta-posterior arithmetic and its Rust rust_beta_posterior_mean fast path are inherited unchanged from AdaptiveThresholdLearner. See Rust Acceleration.