v3.10.0 — Evidence-aware Goal Operating Layer

[3.10.0] - 2026-07-16

Repositioning release: the project is now framed as an Evidence-aware Goal Operating
Layer for Claude Code, not just a trust layer. Trust/evidence is repositioned as the
control system that makes more agent autonomy safe. Minor bump (semver): everything below
is additive — no install, hook, or behavior break.

Added

• PRODUCT_CONSTITUTION.md — North Star, the Core Loop (Goal → Plan → Capability-
  Selection → Execute → Verify → Remember), autonomy + evidence models, non-goals, and a
  Contribution Gate (a new component must name the Core Loop stage it improves).
• Registry capability v3-lite — optional capability: (provides / risk_tier /
  verification_required) and pack: fields on the 15 most-routed skills, validated by
  TestRegistryCapabilitySchema.
• /release-scout — propose-only, Green-tier self-development scout (Claude Code
  releases + watched repos + papers), with FL pre-gates and inbox output; plus durable
  weekly schedule setup scripts (cron + Task Scheduler, dry-run by default).
• docs/research-sources.yaml — external-idea register (adopted / not_adopted /
  internal_validation) so an external claim can't silently become an internal fact.
• RFC-001 (claim pipeline) — Claimify front-stage + 5-route claim-type routing for
  claim-decomposer, with a bilingual (RU+EN) benchmark corpus; design decisions D1–D3
  resolved.
• Prompt-injection guard corpus — benign / malicious / heldout labelled sets,
  and an FL baseline experiment for the response-guard FP/FN.
• PyYAML pinned as a dev/CI dependency, so registry-reading gates actually run in CI
  instead of skipping vacuously (runtime hooks stay stdlib-only).

Changed

• README hero + docs/positioning.md (§1/§3/§6) + plugin/marketplace descriptions carry
  the new identity; count strings preserved for the CI gate.
• rules/skeptic-triggers.md — triggers reframed as signals-to-check (REQUIRES-CHECK),
  not verdicts-of-falsity; Trigger 2 now requires a sourced baseline; constants labelled
  as [WEAK] recall heuristics.
• [VALIDATED:] lifecycle tag renamed to [REVIEWED:] (it meant "last reviewed", not an
  evidence claim — the name collided with integrity.md's evidence vocabulary); the 60-day
  staleness rule is now CI-enforced (39 stale-confirmed skills flipped to review).

Fixed

• Metadata count drift (87 vs 88 hooks in .claude-plugin/marketplace.json); the CI
  count-gate now covers all 3 metadata files and catches adjective-evading phrasings.
• Phantom evidence citations in docs/positioning.md (two experiment dirs never
  committed) — artifacts imported, one overclaim ("each with a red-team pass") corrected.
• One-way depends_on edges — 14 missing upstream backlinks added, both directions gated.

Rejected (recorded as null_results, not shipped)

• Regex-composition response-guard scoring — 0/0 on calibration, 6/8 on held-out; regex
  can't classify context.
• LLM-judge response-guard — sec-auditor red-team found the core premise unsound (a weak,
  injectable model suppressing the only control on the highest-value attack class), killed
  at the design phase before any code.