Falsify Audit Skill

July 5, 2026 · View on GitHub

Audit quant backtests before deploying capital. Receipt anchored on X Layer.

Install

npx skills add chris/falsify-skill

Or pin to a specific version:

npx skills add chris/falsify-skill@0.1.0

Use

Tell your AI agent:

"audit my backtest at results/"

Or call the HTTP endpoint directly:

curl -X POST https://falsify.example.com/audit \
  -F strategy_script=@strategy.py \
  -F backtest_results=@backtest_results.csv \
  -F contract=@risk_contract.yaml

For paid audits (production), the first request returns 402 Payment Required with an accepts JSON describing the x402 payment challenge. Sign the payment with your OKX Agentic Wallet and retry with the X-Payment header:

curl -X POST https://falsify.example.com/audit \
  -H "X-Payment: <base64-encoded-payment-payload>" \
  -F strategy_script=@strategy.py \
  -F backtest_results=@backtest_results.csv

Output

{
  "verdict": "BLOCK",
  "reason": "PBO=0.81 (gate5); universe contaminated by glob(*_daily.csv)",
  "receipts": [
    {
      "anchor":      { "file": "strategy.py", "line": 42, "symbol": "build_universe" },
      "artifact":    { "json_path": "gates.gate5.pbo", "equals": "0.81" },
      "refutation":  "glob(*_daily.csv) includes strategies from other experiments"
    }
  ],
  "receipt_hash": "sha256:9a3f...e1c2",
  "onchain_tx_hash": "0xabc123...def456",
  "onchain_explorer": "https://www.oklink.com/xlayer/tx/0xabc123...def456"
}

Possible verdicts:

VerdictMeaning
PASSGates passed; next action allowed within stated scope
PASS_WITH_DEBTUsable; debt cannot invalidate the next action; names owner + next check
BLOCKMust Fix closes before proceeding; next action is a concrete gate
KILLThesis dead; do not rescue or retune on the same evidence window

Verify the receipt

Receipt hash is anchored on X Layer. Anyone can verify:

python judge_verify.py --tx 0xabc123...def456 --expected-hash "sha256:9a3f...e1c2"

Or check on OKLink: paste the onchain_tx_hash into https://www.oklink.com/xlayer/tx/<tx_hash> and compare the calldata receipt_hash field against your local sha256(canonical_json).

5-Layer Audit Pipeline

  1. L1 Context — assemble audit context (contract + repo rules + manifest).
  2. L2 Red-Light — deterministic mini-Semgrep rules for known-bad patterns.
  3. L3 LLM Adversarial — Brooks-Lint + cutline verdict across normal / production / quant tiers.
  4. L4 Finding Ledger — append-only JSONL of all findings.
  5. L5 Output — verdict + receipt JSON + onchain tx hash.

See V1_ARCHITECTURE.md for full system view.

Examples

License

MIT — see LICENSE.