Mitigation triage for 6 residual risks: fix Risk (a) NOISE-list expans…
Decision
Mitigation triage for 6 residual risks: fix Risk (a) NOISE-list expansion in-session (done — added HOME/PATH/USER/IFS/LOG/ARGS/etc. to aj-doc-release NOISE; re-test cut shell-noise from 19→15 ENV findings, remaining are project-specific real symbols). Defer Risks (b) hardcoded decay rate (c) token heuristic accuracy (d) post-ship rate-limit (e) taste-memory dormancy (f) model-slug staleness to 30/90d outcome review via review_decisions(tag=gstack-comparison).
Rationale
First-principles test (3 questions per risk): silent-vs-visible failure × 30d-likely-vs-rare × in-session-cheap-vs-deferred-expensive. Risk (a) was silent (advisory log fills with junk → real drift hidden) AND cheap to fix in-session (5 min NOISE additions, syntax-validated, smoke re-tested) — so fix now. Risks (b)(c)(f) are non-silent (visible weights/costs/errors). Risk (d) is visible (multiple log entries). Risk (e) taste-memory dormancy IS silent but mitigation requires skill-flow integration which is bigger scope than fits this session — track explicitly: “if zero records by 30d, integrate or prune”. This triage held to the 3 zero-X principles (no compromise on the silent-and-cheap risk; no half-baked premature mitigation on the deferrable ones).
Alternatives Rejected
Outcome
Pending