Convergent review
Run a finite, high-coverage formal review of an exact candidate against a frozen acceptance contract, using independent perspectives and executable oracles, batching repairs, and closing without moving the goalposts. Use when review convergence itself is the job. Do not use for open-ended exploration, root-cause hypothesis generation, option selection, or one quick opinion.From its SKILL.md
npx -y skills add SylphxAI/skills --skill convergent-reviewAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 28 days oldThe repository was created 28 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
3.4 KB, 600 tokens by cl100k_base, as published. Nobody here has run it
Convergent Review
Find material defects and reach a defensible terminal. Read references/convergent-review-method.md before running a formal review.
Method
Freeze the review contract
Record the objective, exact candidate, in-scope surfaces, non-goals, acceptance predicates, risk floor, selected perspectives, checks, and stop condition. Review evidence may reveal that a predicate was always necessary, but a new preference cannot retroactively become a blocker.
Review independently
Select perspectives with different failure models, such as semantic correctness, boundary and dependency integrity, data and state, security, operations, user behavior, and claim accuracy. Reviewers inspect the same candidate independently before findings are consolidated.
Verify orthogonally
Use the narrowest meaningful combination of type checks, tests, property or differential checks, mutation tests, static analysis, replay, simulation, or formal models. More reviewers repeating the same reasoning do not substitute for an executable oracle.
Classify and repair once
Apply the canonical finding and expansion classification from
scope-discipline. Batch terminal blockers and material same-cause repairs;
include positive-net improvements only through its bounded value scan, and do
not let independent or speculative expansion extend this review.
Close on the delta
After repair, review the changed delta, resolution of prior findings, interactions with untouched behavior, same-cause surface, regression evidence, exact candidate, and final claims. Do not restart an unbounded full review unless the repair changed the architecture or invalidated the frozen review contract.
Stop condition
Close when all frozen predicates pass, category 1 and 2 findings are resolved, the exact candidate passes the selected oracles, and reported claims do not exceed evidence. Do not claim a percentage of all possible defects unless the defect population and detection oracle are defined.
Output
Produce a Review Closure Record:
- Contract — candidate, scope, predicates, perspectives, and stop rule.
- Findings — severity, category, evidence, owner, and resolution.
- Verification — checks and exact subject.
- Closure verdict — pass or named unresolved blocker.
- Residual uncertainty — bounded and non-blocking unless it violates the frozen risk floor.
Boundaries
- This skill structures review; domain skills supply domain criteria.
scope-disciplineowns value and expansion control.evidence-and-claims-standardowns the truthfulness of the closure claim.- Independent reviewers increase coverage only when their perspectives or evidence channels differ.
Routing examples
Use for “audit all of these skills and finish the repair,” “review this migration without endless back-and-forth,” and “give a final accept/reject on this exact candidate.” Do not route a casual request for one quick opinion.