Rationale: Mature, focused, working plugin that implements the frontier insight that verification/review-quality is the bottleneck in multi-agent systems: it closes the accept/dismiss feedback loop so high-signal review agents get dispatched first and noise-generators get deprioritized under tight token budgets. Code is clean, fail-open, sanitized, and fully tested (11/11 pass); the architecture (shared DB, self-contained library, progressive enhancement, never a gate) is sound. Not a candidate for modernize/deprecate; the SOTA gaps are incremental hardening, not a paradigm miss.
Library+command plugin (no hooks, no MCP). hooks/lib-trust.sh (365 lines, sourced not executed) records review-finding outcomes (accepted/discarded) into a trust_feedback table in the shared .interspect/interspect.db, then computes a per-(agent,project) trust score that is severity-weighted (P0=4.0, P1=2.0, P2=1.0, P3=0.5), time-decayed (~30-day half-life), and blended project-local vs global (blend weight min(1.0, project_count/20)), floored at 0.05, capped at 1.0. _trust_scores_batch feeds interflux dispatch; clavain:resolve calls _trust_record_outcome; /trust-status flags agents below 0.30 as suppression candidates. Has input sanitization, secret redaction, fail-open neutral 1.0, backward-compat shim. 11/11 tests pass.
The Agent-as-a-Judge / Gaming the Judge (Jan 2026) frontier warns outcome-driven agent reputation must guard against reward-hacking, self-preference bias, and feedback loops, and intertrust’s own PHILOSOPHY.md names prevent gaming or feedback loops as a decision filter, yet the code implements none. (1) Starvation ratchet: a low-trust agent is deprioritized/not dispatched, accrues no new evidence, and can never recover (no exploration floor). (2) No minimum-sample confidence gate before suppression (can flag below 0.30 on a few decayed P3 outcomes). (3) No detection of correlated/duplicate findings inflating one agent’s accept rate. An exploration floor plus sample-size confidence bound would satisfy its own stated north star.
Clean separation from interspect, confirmed by grep: interspect has zero references to trust_feedback, _trust_score, or lib-trust.sh. They share only the .interspect/interspect.db file, not schema or logic (interspect = evidence/routing/canaries; intertrust = reputation/trust). Split is intentional and defensible but conceptually adjacent (both score agents from evidence to influence flux-drive dispatch; interspect:calibrate also computes agent scores). Watch-item: if intertrust stalls or the boundary blurs, merge back into interspect is the natural consolidation.
[“Add an exploration floor / epsilon-dispatch so suppressed agents are occasionally re-dispatched to accrue fresh evidence and recover, fixing the starvation ratchet that violates PHILOSOPHY.md recoverable-score filter.”, “Add a minimum-sample confidence gate before flagging suppression candidates (require N>=5-10 weighted reviews, or use a Wilson/Beta lower-bound instead of raw accept ratio).”, “Fix stale docs: README claims lib-trust.sh is 233 lines but it is 365; align README Architecture section and headers with the actual file.”, “Extend the 11-test suite with regression tests for the recovery/exploration path and low-sample suppression behavior.”]