prove-sop — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited prove-sop (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
When the user asks for a statistical proof ("prove X", "show that Y holds", "give a rigorous derivation"), or when a proposition node sits without a verified proof:
mcp__memory__match_signatures with the proposition text and domain="proof". If similar past proof errors exist, read them first before drafting.mcp__prove__propose_proposition to register it. Pass parent_id to a related question or hypothesis node so empirical and theoretical work share one tree (architecture.md §13).mcp__prove__retrieve_skeletons with lexical and semantic keywords you extracted from the proposition. If fewer than 3 candidates rank above similarity 0.4, ingest more material via mcp__prove__ingest_proof_corpus first or fall back to a from-scratch outline.mcp__prove__propose_proof_skeleton. Run $bt-tournament to BT-rank them; the proof-trunk leaderboard is mcp__memory__get_bt_leaderboard(kind="proof_skeleton"). Stop when the top-2 LCB / UCB intervals separate or 3 comparisons cover every candidate.mcp__prove__register_proof_draft. At this branch point (choosing one skeleton over the others), optionally call mcp__cockpit__narrate("why this skeleton") so the cockpit reflects the decision (v5.0; optional).mcp__prove__segment_proof(draft_id, snippets) -- this opens a fresh diagnostic manifest. For each snippet, call mcp__prove__diagnose_snippet to get historical proof-error candidates, decide is_flawed, and call mcp__prove__register_diagnosis. After all snippets are recorded, call mcp__prove__finalize_manifest.mcp__prove__compose_correction_prompt to build the global-fix prompt; generate the corrected LaTeX inline; call mcp__prove__apply_correction. The new draft is itself segmentable -- if you suspect more flaws, loop back to step 6 on the new draft_id. Before applying a global correction (a non-trivial decision among possible fixes), optionally call mcp__cockpit__narrate("...") so the activity pane shows the reasoning (v5.0; optional).$preregister. Lock metric, threshold, direction. The reviewer will later cross-link the theorem and the preregistration.mcp__prove__triage_for_formalization. If eligible=True, request a wallclock budget from the budgeter agent for long attempts (see prover.md § Budget); low-cost checks may proceed with an audit warning when no budget is configured. Then spawn the prover subagent. A successful lean_verify attaches as the strongest possible evidence; a failure feeds the cross-domain failure ledger via mcp__memory__record_failure(domain="proof").reviewer agent's proof checklist (P5) demands either a manifest.status='empty' plus a Lean verification, or an explicit unverified flag. Plan accordingly.update_bt_rating between a hypothesis and a proof_skeleton. Use parallel tournaments (architecture.md §13). Stop a skeleton tournament when the top-2 LCB / UCB intervals separate, or when every candidate has participated in at least one comparison.mcp__prove__diagnose_snippet is read-only; you must explicitly register_diagnosis for the entry to land in the manifest. This is intentional: the LLM judgment lives in the agent loop, not the MCP tool.segment_proof on the new draft_id -- this opens a fresh manifest. Old manifests stay around as audit history.prover.md § Budget). Missing budget context is an audit warning, not a reason to discard the NL proof.$preregister$bt-tournament~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.