global-proof-review

Reviews global proof diagnostics and routes failures to the smallest valid repair stage.

65|2|Updated Aug 10, 2026
One-click install
npx skills add https://github.com/DechenZhang/VALG-ML-Theory-Agent --skill global-proof-review-dechenzhang
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: global-proof-review
Source: https://github.com/DechenZhang/VALG-ML-Theory-Agent/tree/main/skills/global-proof-review
Command: npx skills add https://github.com/DechenZhang/VALG-ML-Theory-Agent --skill global-proof-review-dechenzhang

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? When developing machine-learning theory proofs, a global proof diagnostic can hide theorem-level obstructions, shallow closure claims, or broken generated-output flows that only surface late in step-level proving. This Skill acts as an early review gate that validates the diagnostic before any step-level proof work begins, deciding whether it is accepted, needs a global rewrite, requires proof-sketch repair, or exposes an idea-level obstruction. ## Core Features & Use Cases - Contract and fidelity validation: Checks that global_proof.md contains all required sections, stays faithful to setting.md and the accepted proof sketch, and respects the diagnostic boundary. - Multi-gate audit pipeline: Applies theorem-critical mechanism witness, noncircular closure, entry-state trace, step-locality, baseline invariance, residual-to-target adequacy, and gate evidence row checks before acceptance. - Failure-localized routing: Classifies every hard obligation as step-local, sketch/interface defect, or idea/theorem-contract defect, and routes to /global-proof, /proof-sketch, or /subagent-idea-generator. - Use Case: After a global proof diagnostic is written for a theorem branch, run this review to catch a circular recurrence closure or a missing mechanism witness before step workers waste effort on unprovable local steps. ## Quick Start Ask the agent to review the branch's global_proof.md against setting.md and the accepted proof sketch, and write global_proof_review.md with a status of ACCEPTED, REVISE_GLOBAL, REVISE_SKETCH, or IDEA_FAIL.

Frequently Asked Questions about global-proof-review

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I review a global proof diagnostic before step-level proving?

Run this review after global_proof.md is written and before any proof-step runs. It validates required sections, sketch fidelity, closure certificates, and gate evidence, then writes global_proof_review.md with a status and smallest retry target.

What review statuses can a global proof review produce?

The review outputs exactly one of ACCEPTED, REVISE_GLOBAL, REVISE_SKETCH, or IDEA_FAIL. Each status maps to a retry target: none, /global-proof, /proof-sketch, or /subagent-idea-generator respectively.

When does a proof failure route to sketch revision instead of global rewrite?

REVISE_SKETCH applies when the proof roadmap, step decomposition, dependencies, closure mechanism, or generated-output flow must change before step proving can be sound. REVISE_GLOBAL is used only when the accepted sketch already contains the needed evidence and the diagnostic merely omitted or misstated it.

Does the review trust a previously accepted proof sketch review?

No. The accepted proof_sketch_review.md is treated as an input gate and provenance only. The review independently re-audits theorem viability, mechanism witnesses, source adequacy, and closure claims exposed by the global diagnostic.

What files does the global proof review read and write?

It reads setting.md, the accepted proof_sketch.md and proof_sketch_review.md, and global_proof.md, then writes global_proof_review.md. It never modifies the diagnostic, sketches, step artifacts, trackers, or accepted results.

Why would a global proof review return IDEA_FAIL?

IDEA_FAIL occurs when the target theorem appears false, materially mis-scoped, or repairable only by changing primitive assumptions, the algorithm or model, theorem scope, exposed dependence, or the success criterion. The review must identify which setting or goal component must change.