v1.1.0 — Review tiers: proposes vs disposes

v1.1.0 — Review tiers: stochastic judgment proposes, deterministic verification disposes

Highlights

• Review-tier doctrine (#134): every check now belongs to exactly one tier, sorted by reproducibility. Tier 0 (deterministic validators) is the only tier that hard-blocks; Tier 1 (LLM judgment with a mandatory evidence artifact) gates only through that artifact; Tier 2 (pure judgment) never gates — it is advisory and goes to the findings ledger. Includes the RATCHET: recurring Tier 2 findings get investigated and encoded as Tier 0/1 checks, leaving the stochastic layer forever.
• New skill: findings-ledger (#134): append-only JSONL ledger for unevidenced review findings, with a stdlib-only ledger.py (add / tally / triage / promote / retire). Findings are fingerprinted (sha256 of file + normalized claim) so the same defect rephrased across runs collides; recurrence counts distinct run ids, so an agent repeating itself within one run cannot cross the promotion threshold. Gitignored by default, with a documented team-shared mode.
• New skill + agent: adversarial-claims-reviewer (#133): cold-context, read-only adversarial review of documents that make formal or technical claims. Inventories every equation and quantitative claim, verifies each AS NAMED via deterministic scripts, and reports VERIFIED / REFUTED / UNVERIFIABLE / VACUOUS counts. A REFUTED verdict requires a failing script or explicit counterexample — this is the Tier 1 pattern the tier doctrine generalizes.
• Reviewer agents rewired (#134): code-reviewer, security-reviewer, library-reviewer, and adversarial-claims-reviewer each gained a "Tier discipline" section. Unevidenced findings are emitted as ledger appends instead of blocking language; hold / fix-before-merge verdicts riding only on Tier 2 findings are proposals to the operator, not gates. Severity is not tier.
• Routing-collision sweep (#131, #132): 33 confirmed routing collisions fixed with reciprocal "Not when" deflections across the skill library, so deliberately shared trigger keywords disambiguate instead of colliding.

Improvements

• CONTRIBUTING.md (#133): skill/agent conventions, the two review gates, the tier rule, and the ship-allowlist contract, all in one place.
• Library docs sweep (#129): frontmatter description quality (trigger vocabulary, cross-reference trailers, length trims), reciprocal cross-references for asymmetric seams, fill-in templates relocated to assets/, dead idea-refine script removed, portable telemetry paths.
• validate.sh: new Tier 0 invariant — any file declaring a "Tier discipline" section must reference the tier doctrine, and the findings ledger (if present) must be valid JSONL with known status values. The wiring polices itself.
• New maintainer command /triage-findings (repo-local, not installed): tallies the ledger and proposes a ratchet target per recurring finding — a concrete validate.sh rule, a named Tier 1 evidence script, or stays-advisory with reasoning. The command only proposes; the human disposes.
• README: branded as Engineering Heresy — Agentic Framework (#130); new Commands table.

Fixes

• Fingerprint normalization treated contraction apostrophes as quote delimiters, deleting prose between them so distinct defects falsely collided (caught by external review, fixed test-first with a new test_ledger.py regression suite).
• routing-collision-sweep workflow: dropped top-level allOf from its verdict schema, which silently returned empty results.
• ledger.py hardening from its own adversarial review gate: promote refuses PROMOTED without --evidence; malformed ledger lines exit 2 instead of crashing; appends survive a missing trailing newline; recurrence counts independent runs as documented.

Install

curl -fsSL https://raw.githubusercontent.com/LazyIsEfficient/agentic-os/v1.1.0/install.sh | bash

Asset: agentic-os-v1.1.0.tar.gz · SHA-256 ec9c493b147779d4f1f2354608a5c72716d8fb44834c7cfad7767da4a43a8290

Full diff: v1.0.0...v1.1.0