agentleFS
JuankCadavidJuankCadavid/akili-specs/.claude/skills/judgment-day/SKILL.md

judgment-day

Skill12 starsChanged 14 days agoMITView on GitHub

Trigger: judgment day, dual review, adversarial review, juzgar. Run explicit blind dual review with at most two scoped fix/re-judgment rounds.

3 other projects use a version of this, 1 of them edited.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

6 rules. 0 are common, 4 are distinctive, and 2 are very specific to this project.

  • 0

    Common

    Most of the 3 edited versions we read keep these.

  • 4

    Distinctive

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Never launch review-refuter.

    line 22
  2. Distinctive:

    Only the parent orchestrator merges/persists findings, launches the fix actor, launches scoped re-judgment, and updates native counters.

    line 23
  3. Very specific:

    Start review/start(target, mode=judgment_day) and persist the transaction.

    line 57
  4. Distinctive:

    Ask before round-one correction.

    line 60
  5. Distinctive:

    Run both judges again only over the frozen ledger plus immutable fix delta.

    line 61
  6. Distinctive:

    Prefer running the two judges on a model different from the one that authored the design (author ≠ auditor, per AKILI model routing).

    line 88

“Common” is measured: kept by at least 6 in 10 of the 3 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Activation Contract

Load only when the user explicitly requests Judgment Day or equivalent dual/adversarial review for a concrete target. Judgment Day replaces ordinary 4R for that target; never run both.

Hard Rules

  • Resolve matching project skills before starting and pass the same paths to both judges and any fix actor.
  • Build one complete immutable target, then launch two blind read-only judges in parallel with identical scope and criteria.
  • Each judge returns one neutral findings result and terminates. Wait for both; never accept a partial judgment.
  • Never launch review-refuter; two-judge agreement is the corroboration mechanism.
  • Only the parent orchestrator merges/persists findings, launches the fix actor, launches scoped re-judgment, and updates native counters.
  • Fix only severe findings confirmed by both judges. WARNING/SUGGESTION rows remain info.
  • Judges must contrast every count, total, and quantity the target asserts against the prose of the other in-scope documents. Documents agreeing with each other is not corroboration — it is often one wrong idea copied forward; a figure contradicted by any in-scope document's own prose is a finding.
  • Before any other reading, judges attack the design's Premise Ledger at its source: re-run or re-read every citation, try to refute every UNVERIFIED row with their own search, and look for premises the design depends on that carry no row. A premise is never accepted because the requirements and the design agree on it — only the source settles it. Row shape, classes, and triggers are defined in /akili-specify Step 2.2 — Premise Ledger block and are not restated here; this rule sets only what a bad row costs.

    A row the citations cannot settle is attacked, not merely scored. Action 1 needs a citation and action 2 is written for the marker, so two row states fall between them: a citation cell holding neither a citation as run nor the UNVERIFIED marker, and a citation naming a secondary source rather than the primary one. Which sources count as secondary is set by citation rule (d) of the /akili-specify Step 2.2 — Premise Ledger block, and is not restated here. Both states are attacked under action 2, as though the row carried the marker: the judge searches for the primary source itself and reports what that search returns, including not re-run where its host cannot reach that source. Re-reading a document settles that document, never the system it describes, so a document that agrees with the design is not confirmation. Recording a severity from the table below without attempting that refutation does not discharge this rule.

    Judge findsSeverity
    The source contradicts a premisesevere
    A citation does not reproduce at its verified-at commitsevere
    The design names existing code and has no Premise Ledger, or a false stated-empty linesevere
    A triggered class with no rowsevere
    A depended-on premise with no rowfinding, severity by Impact
    A premise confirmednot a finding

    Read-only is not no-search. A read-only judge writes nothing. Within that contract it may read and search the repository, and read its history where the host allows (the packaged judge tool set searches but has no shell). A judge that cannot run a cited command re-derives from the cited files where feasible and reports that row not re-run — never as confirmed. The protocol is unchanged: two-judge confirmation still gates auto-fix, and a premise contradiction reported by one judge is recorded as suspect with its command as run, so the architect settles it with a single re-run.

  • Permit at most two fix rounds and two scoped re-judgments. Re-judgment sees only the frozen ledger plus fix delta and may record fix-caused defects.
  • Terminal transaction states are only approved | escalated; never reset or extend an exhausted lineage.

Decision Gates

ConditionAction
Target unclearAsk one scope question and stop.
Both judges confirm severe findingAsk before round-one correction; then use the bounded fix actor.
One judge reports itRecord suspect; do not auto-fix.
Judges contradictEscalate for explicit human decision.
Scoped re-judgment fails before round twoParent may launch the final bounded fix round.
Any issue remains after round twoEscalate and stop.

Execution Steps

  1. Start review/start(target, mode=judgment_day) and persist the transaction.
  2. Launch both read-only judges against the same immutable target.
  3. Merge findings into the frozen ledger and persist it through the selected artifact store.
  4. Ask before round-one correction; run the fix actor only for confirmed severe IDs.
  5. Run both judges again only over the frozen ledger plus immutable fix delta.
  6. Repeat once at most, then run independent final verification and emit the terminal receipt.

Output Contract

Return target identity, round, confirmed/suspect/contradiction/INFO counts, correction work units, scoped re-judgment result, artifact references, skill resolution, and exactly one final JUDGMENT: APPROVED ✅ or JUDGMENT: ESCALATED ⚠️.

References

  • ../_shared/review-ledger-contract.md — canonical transaction, ledger, persistence, and lifecycle contract.
  • references/prompts-and-formats.md — compact judge/fix prompts and verdict shape.

These reference files are not packaged with AKILI-SPECS. When they are unavailable, proceed with the contract described in this document alone: persist the findings ledger inside the spec folder (see below) and derive judge prompts from the Hard Rules and Output Contract.

AKILI-SPECS Integration

AKILI momentHow to use this skill
/akili-specify Step 2.5 — Present & Approve, Review Design optionThe user-selected blind dual review of design.md before tasks are written; the target is the approved requirements + draft design, and judges attack the Premise Ledger first
/akili-review Step 3 — Merge and Verify, full tier (via the pr-review skill)Two blind judges per critical finding (blind protocol only, no fix rounds); a critical only one judge confirms becomes a warning with suspect: true. lite uses one verifier pass and trivial none, so they do not run this protocol
/akili-archive Kaizen MeasureSevere confirmed findings from Judgment Day runs are a signal row in the kaizen retrospective

Adaptation rules:

  • Persist the findings ledger as docs/specs/<spec-path>/judgment.md so /akili-archive can read it as evidence.
  • The two-judge blind protocol, the two-round fix ceiling, and the APPROVED/ESCALATED terminal states apply unchanged.
  • Under AKILI pre-approved the review still runs. A target corrected without re-judgment (/akili-specify Step 2.5 Fix only) closes judgment.md with the AKILI annotation fixed, not re-judged plus the fixed finding IDs, never JUDGMENT: APPROVED. The annotation is an AKILI-layer record; this skill's terminal states stay unchanged (see Hard Rules).
  • Prefer running the two judges on a model different from the one that authored the design (author ≠ auditor, per AKILI model routing).

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.