Capability Library

A real registry of reusable, domain-independent components — the direct answer to "don't build something once and forget it exists." Anyone starting new work should check here first, before building a similar capability from scratch.

The four real capabilities

Wilson Confidence Gate

wilson_confidence_gate.py — gates a state transition on real accumulated evidence, not a raw small-sample percentage. Two real presets: FOUNDATIONAL_FLUENCY (80% threshold, 95% confidence) and CONCEPT_NODE (75% threshold, 90% confidence) — deliberately more permeable, so students can advance to adjacent concepts and build knowledge holistically rather than getting stuck perfecting one node in isolation.

Already wired into: the live Assignment Tracker — both the per-node mastery gate (replacing a single-batch "4 of 5" threshold with real cross-assignment accumulated confidence) and the difficulty-selection engine itself, where it directly fixes a previously disclosed exploration bug. Real test history: 3 genuine bugs caught and fixed across iteration — a deterministic single-tier lock-in, then a premature lock-in from a single lucky attempt — before 5 real test scenarios passed. Could also apply to: the NFL Strategy Composer (which currently has its own separate, duplicate Wilson implementation) and exploratory_heuristic_42.py (which currently uses a strictly weaker fixed-N-only threshold) — both real, identified, not yet done.

Temporal Validity Warrant

temporal_validity_warrant.py — tracks a real premise's freshness against a stated TTL and half-life. Distinct from the Methodology page's Staleness Tracking rule, which tracks a verification claim's coverage going stale against new invariants; this tracks the world itself going stale — a measured fact that was true when recorded and degrades in trustworthiness over real time.

Not yet wired into anything live. Real test status: 4 assertions, all passing on first run. Could apply to: a Wayfinder student's mastery baseline aging between sessions, a Tesla sensor calibration's real validity window, any cached S-channel environmental assumption that shouldn't be trusted indefinitely.

Channel Divergence Monitor

channel_divergence_monitor.py — a real extension of the existing channel_correlation_diagnostic.py (imported directly, not reimplemented). The original module answers "what is the current pairwise correlation." This one answers "has that relationship drifted from its own historical baseline" — a genuinely different question, since a stable 0.6 correlation and one that just jumped from 0.1 to 0.6 are different situations even if the current snapshot looks identical.

Not yet wired into anything live. Real test status: 2 assertions passing after 1 real false-alarm caught and fixed — the first test used only 30 samples, and pure sampling noise on genuinely uncorrelated channels exceeded the drift threshold on its own; fixed by verifying a genuinely stable sample size (n=100) directly rather than assuming one. Could apply to: the contamination benchmark's own reported statistics (adding a real interval around "44.1% of trials had zero contamination"), any live system already using the base diagnostic that wants real drift alerting instead of a static snapshot.

Shadow Probe

shadow_probe.py — continuous adversarial self-testing. Injects synthetic perturbation into a duplicate copy of a state estimate and checks whether the estimator's own reported uncertainty widens appropriately. An estimator that stays "confident" even when fed real noise is caught here, before live data catches it instead.

Not yet wired into anything live. Real test status: 2 assertions passing on first run — a real, honest uncertainty function correctly widens under perturbation; a deliberately overconfident one is correctly caught failing to. Could apply to: any real estimator with an uncertainty function — the Wilson gate itself, the Tesla Authority State Model's risk scores, future domain estimators.

Where this honestly does not apply

The Tesla case analysis was checked directly against the Wilson gate specifically, and it doesn't fit: those are 16 distinct real incidents, not repeated trials of one consistent test — a different statistical shape entirely. Forcing a capability to apply somewhere it doesn't genuinely fit would defeat the purpose of building this registry rigorously in the first place.

What's real vs. what's a documented next step