h-verify
h-verify enforces the FPF verification loop (baseline, measure, evidence, record) by identifying a decision, reading its predictions, optionally baselining affected files for drift detection, gathering evidence for each prediction through tests or metrics, and attaching verdict-tagged evidence artifacts. Use it when you need to validate whether previously declared predictions still hold, detect file drift, or surface stale decisions past their valid_until date.
git clone --depth 1 https://github.com/m0n0x41d/haft /tmp/h-verify && cp -r /tmp/h-verify/packages/haft-pi/skills/h-verify ~/.claude/skills/h-verifySKILL.md
# h-verify — Evidence against reality When an FPF evidence or validity distinction is material, inspect a known SourceID/UnitID with `haft_query(action="fpf", mode="inspect", identifier="...")`, or query the exact concern and inspect the direct pattern body. Retrieval is not evidence or a verdict. Recover the exact record and its claims, thresholds, validity window, and planned evidence. Gather current evidence, state context transfer and expiry, and attach it with `haft_decision(action="evidence", ...)`. Surface weakened, refuted, stale, or drifted claims honestly. Any rebaseline, supersede, deprecate, or reopen mutation requires explicit operator action.
|
|
Manual-only skill that creates bounded execution authority from an explicit operator grant. Never auto-trigger or infer approval from prompt text or tool schemas.
Compare two or more existing candidates under an explicit characteristic space, parity basis, and predeclared selection policy. Return trade-offs and a non-dominated set; persistence is conditional and a binding choice requires a direct unambiguous operator request.
Route one direct, unambiguous operator request for a bounded binding choice. A manual h-decide invocation is a compatible shortcut, not an authorization receipt.
Diagnose a concrete failure with rival-hypothesis testing. Stabilize the symptom, generate distinct explanations, test discriminating observations read-only in parallel, and rank by evidence while retaining losing rivals.
Generate 3-5 genuinely distinct candidate approaches for a current question, with the weakest link of each visible. May work from an inline question or a durable ProblemCard; persistence is conditional.
Shape an under-articulated engineering problem without assuming a solution or forcing a project phase. Default to a conversational frame; create a ProblemCard only on explicit save intent or when current Work supplies a concrete operator-named or agent-inferred receiving use that needs a durable accepted problem statement.