A rule ledger for AI agent harnesses. Per-rule provenance for CLAUDE.md / AGENTS.md, plus ablation inquests that tell you which rules still earn their place — so a model upgrade retires rules with evidence instead of making you throw the harness away.
A verderer was the officer of the royal forest: he kept its law, held inquests, and answered for what the law was still for. That's this tool's job for your harness.
Every harness rule is a patch over a model weakness, frozen at the moment you wrote it. Models improve; the repo matures until the code itself carries the convention; the rules stay — costing tokens every session, and sometimes actively fighting a model that outgrew them. The common answer is ritual: delete everything every six months and see what breaks. Verderer replaces the ritual with a ledger and evidence.
Validated on a real repo before a line of this tool existed: of 3 long-standing CLAUDE.md rules probed by ablation, 1 was dead (the model derived the convention from neighboring code and said so), 1 was derivable (kept as cheap insurance against an expensive miss), and 1 reduced to a single clause — everything else it stated, the model already did unaided. ~1 hour, 3 headless runs.
npx verderer scan # inventory harness rules into .claude/verderer.json
npx verderer status # active/retired rules, provenance, verdicts
npx verderer checkpoint --model <id> # record the model by hand (tools without an adapter)
npx verderer verdict 9f2c dead --evidence "derived it from neighboring code"
# record an inquest (id prefix is enough)
npx verderer probe 9f2c --task "add a user store module"
# run the task with and without the rule, judge the diffsscan reads CLAUDE.md, AGENTS.md and GEMINI.md at the root, sub-agents in .claude/agents/*.md, skills in .claude/skills/*/SKILL.md, and Cursor rules in .cursor/rules/ (.md and .mdc). A symlinked harness (AGENTS.md → CLAUDE.md) counts once, and a rule repeated across files is one ledger entry, carried by the first file that has it.
A rule is a top-level list item (-, *, +, 1.) or a paragraph that opens with a bold lead (**Never trust a merged schema.** Every module…). YAML frontmatter, code blocks and spec-kit's generated sections (Active Technologies, Recent Changes) are skipped.
scan is idempotent and append-only: known rules keep their provenance and verdicts, new rules are dated, rules that left the file are marked retired — never deleted. The ledger is the history of your harness's law.
Rewording a rule doesn't reset it. A rule edited in place (same spot, mostly the same words, or a trim of them) keeps its provenance and verdict, and its old wording goes into revisions. The verdict judged the old wording, so it shows as stale until you re-run the inquest.
Verderer checks how git sees each harness file and adapts:
- tracked — provenance comes free:
git blamedates every rule (and the date tells you which model era it was written against). Pruning is a normal commit or PR; rollback is git. - ignored — a deliberate privacy choice some repos make. There is no history, so the ledger carries provenance itself: first scan records rules as
origin: unknown, everything after is dated. Consider tracking.claude/verderer.jsoneven while the harness stays private — the decisions get history without exposing the content.
verderer probe <id> --task "<text>" holds an inquest by experiment. It checks out HEAD into two throwaway worktrees, plants your harness files as they are on disk (ignored and uncommitted ones too), removes the rule from one of them, and runs the same task in both with headless Claude Code on the model the ledger records:
- control — full harness. If this run gets it wrong, the task doesn't test the rule.
- ablated — the rule removed (from every harness file that carries it, sub-agents and skills included).
Probes run on Claude Code, so a ledger whose model was detected from Codex sessions can't be probed yet (checkpoint a Claude model to probe on it).
Pick a task that forces the decision the rule is about: for "name files in kebab-case", ask for a new module. Each run is capped by --budget (USD, default 2), edits are auto-accepted, and only project settings load, so a copy of the rule in ~/.claude/CLAUDE.md can't leak into the ablated run. Probe sessions aren't saved, so they never count toward model detection.
Then a judge (same model, no tools, nothing but the rule, the task and both runs) proposes a verdict with a sentence of evidence. It answers inconclusive when the control run broke the rule too or a run failed: the task tested nothing, so try a sharper one.
judge · $0.02: needed
The control run created src/user-store.js; the ablated run followed the camelCase neighbors and created src/userStore.js.
read the diffs in the report, then record it (or overrule it):
verderer verdict 25a4806f needed --evidence 'probe: The control run created … (.claude/verderer-probes/…)'
Pass --record to save the proposal as the verdict directly. Either way nothing touches the harness: pruning stays your call.
Diffs, results, costs and the judgment go to .claude/verderer-probes/<id>/<timestamp>.json; the worktrees are removed unless you pass --keep.
Verderer never polls for releases. The question that matters is narrower: is this repo running on a different model than the one your verdicts were judged against? Every verdict records its model; scan reads the model id the repo actually ran on from the agent tool's own session logs and flags verdicts judged on another model as stale. Those logs carry concrete ids, so a stable "opus" alias doesn't hide an upgrade underneath:
- Claude Code —
~/.claude/projects/<repo>/*.jsonl - Codex —
$CODEX_HOME/sessions/**/rollout-*.jsonl(default~/.codex), sessions started in the repo or any folder under it
When both have sessions for the repo, the one that ran last wins.
model changed: claude-fable-5 → claude-fable-5-1
3 verdict(s) rendered against another model are stale — re-run their inquests.
For tools without a detection adapter yet, verderer checkpoint --model <id> records it by hand.
Then re-run the inquests in one go. A conclusive probe keeps its task on the rule, so verderer probe --stale re-runs every stale verdict with that task on the new model (add --record to save the new verdicts):
re-running 1 stale inquest(s) on claude-fable-5-1
25a4806f40cd Name every new source file in kebab-case (user-store.js, not userStore.js).
was needed (judged on claude-fable-5) · now needed · $0.04
…
Stale rules with no saved task are listed with the probe command to run once by hand.
.claude/verderer.json — one entry per rule:
{
"id": "9f2c01ab34de",
"text": "ESM only; imports use `.js` extensions even in TS source.",
"source": "CLAUDE.md",
"section": "Conventions",
"firstSeen": "2026-03-17",
"origin": "git:da89e1b",
"status": "active",
"lastVerdict": { "date": "2026-10-31", "result": "dead", "model": "claude-fable-5", "evidence": "ablation probe: model derived it from neighboring code", "ruleId": "9f2c01ab34de" },
"probe": { "task": "Add a TS module that imports the user store." }
}Verdicts are three-valued on purpose: dead (the model does it unaided), derivable (unaided but an expensive miss — keeping it is cheap insurance), needed. Blast radius is a judgment call, so verderer reports and you decide; it never auto-deletes a rule.
verderer verdict records one: it stamps today's date, the ledger's current model and the rule's wording, so a model change or a rewording flags it as stale. Evidence is required.
- Model-detection adapters beyond Claude Code and Codex; the ledger core is already tool-agnostic.
- Probes on Codex (
codex exec), so Codex-detected models can be probed too. - Nested harness files (
packages/*/CLAUDE.md,AGENTS.mdin subfolders).
npm test # node --testNo dependencies. Node ≥ 20.
Built by Victor Velázquez — COO at MagmaLabs and author of The Agentic Engineer. Sibling of banneret (sub-agent pipeline) and reeve (MCP guardrails): banneret disciplines how agents work, verderer audits what you tell them.
MIT — see LICENSE.