proofloop-honest-agent-verification

name	ProofLoop Honest Agent Verification
description	A reusable ProofLoop skill for making coding agents prove completion with external receipts, anti-gaming gates, held-out tasks, and live or stratified browser evidence before they claim done.

ProofLoop Honest Agent Verification

Use this skill when an agent says a task is done but the user needs proof that the work actually survived the relevant scorer, tool contract, browser path, or product workflow. This is verification you run, not certification, until independent adoption makes the receipts meaningful outside the team that produced them.

Trigger Signals

• agent says done without proof receipts, official scorer output, or browser evidence.
• benchmark pass where the benchmark could be gamed by answer keys, hidden state, fixture leakage, or a shortcut harness.
• ship demo, hackathon, accelerator sprint, or release gate where a solo builder needs an honest yes/no answer.
• live UI proof, official scorer, ProofLoop, anti-gaming, held-out, no answer-keys, in-app transfer, or honest provenance.

Preconditions

• A repository, app, benchmark, workflow, or task family the agent is expected to complete.
• Permission to run local tests, local browser checks, and existing configured model/tool credentials.
• A written goal, acceptance criteria, and budget ceiling before any paid model or hosted browser work.
• A referee surface outside the worker agent: scorer, test, verifier, browser script, receipt ledger, or human review gate.

Strategy

1. Freeze the claim. Write the exact completion claim, target workflow, task family, allowed tools, budget ceiling, and proof standard before running the agent.
2. Separate capability from product proof. Run broad benchmark capability through the cheapest honest live harness first. Use full browser proof for workflows where browser semantics matter, and otherwise use stratified browser samples seeded from the commit SHA.
3. Block gaming surfaces. Enforce held-out tasks, no answer keys in the harness, no hidden memory shortcuts, no per-task writers, no self-edited scorer, and no final answer based only on the worker's self-report.
4. Collect receipts. Record command output, scorer result, model/provider route, tool calls, artifacts, browser screenshots or traces, cost ledger, and failure reasons. Redact secrets and avoid publishing private prompts or code paths without approval.
5. Fail closed on divergence. If a headless pass fails in the browser, label a harness shortcut. If a headless fail passes in the browser, label a harness bug. Either way, do not claim the benchmark result until the disagreement is explained.
6. Repair from evidence. Localize the failure, patch only the smallest relevant surface, rerun the same proof gate, and promote the failure into a regression if it represents a durable bug class.
7. Return an honest verdict. Report pass, fail, blocked, or partial with the exact receipts and the next most valuable fix. Never inflate a low score into a product claim.

Constraints

• Do not hardcode benchmark answers, expected outputs, fixture IDs, or per-task branches.
• Do not weaken scorer thresholds, test assertions, browser selectors, or receipt requirements to pass.
• Do not spend new money, publish externally, or expose credentials without explicit approval.
• Do not claim "prod browser verified for all tasks" unless every task actually ran through the prod browser.
• Do not call this certification; call it ProofLoop verification unless an independent certifier exists.

Validation

node evomap/validate-skills.mjs
node skills/solo-founder-nodes/conformance/conformance.mjs

The verification is acceptable when the skill file validates, the canonical Solo Founder Nodes conformance probe passes, and the produced ProofLoop receipt names the command, scorer or browser evidence, result, budget/cost note, and any unresolved blocker.