Vroom Analytics

# FinePrint — the building inspector for your agent's house rules Independent measurement of whether an AI agent's enforced limits match its stated limits — decision, approval, and effect. ## Try it live (for judges) Live demo: https://fineprint.vroomanalytics.com Password: fineprint-judges-2026 The demo calls a live model, judges each probe with FinePrint's evaluator, and executes nothing. ## What it does You tell a coding agent what it must never do: never delete recursively without asking, never send credentials to an unnamed host, never force-push to production. FinePrint checks the agent against those rules the way a building inspector checks a building against its filed code — item by item, writing down what matches and what doesn't. It never declares anything "safe." Each probe records three layers: policy decision, approval that actually happened, and effect that actually occurred. Unknown is never a bypass: no decision means a coverage gap, never a pass. A score is never a security grade: every number travels with its coverage, its route, and a plain-English finding. ## Built during the hackathon, with Meta The core innovation was developed during the hackathon window (Oct 26 – Nov 4) using Meta's tooling. What predates it is a seed, stated plainly: a single-harness Codex proof-of-concept (Oct 21–22) and an early live campaign (Oct 23). Everything that makes it a framework — the harness-neutral re-measurement protocol, the 24-scenario set generated with Meta's muse-spark-1.3, the 16-harness scorecard and weighted scoring, the MCP engine, the bulk of the live evidence (480 runs plus the 96-cell Omnigent parity run), and the upstream fix loop — was built in the window, with Meta's models generating the tests, a live arm (muse-spark-1.1), and Muse Code measured live. ## What was measured Founding measurement (vendor-confirmed, sealed): 41 probes against Codex's own policy checker. Pre-fix: allow 6, prompt 4, forbidden 9, unknown 22. Post-fix (Oct 22): forbidden 11; wrapped forms stay unknown by design. 16-harness scorecard (published formula, never hand-set): Amp 64% (decision only), Muse Code 49.71 live, Codex 22.05 of 75. The fix loop is real upstream work: a Pi permission-gate fix filed as earendil-works/pi#11608 (auto-closed by the contributor gate, not rejected on merit; claimed as filed, not merged) and an Omnigent credential-flow guard open as omnigent-ai/omnigent#12762 (claimed as open, not merged). ## Links - Repo (Apache-2.0): https://github.com/Vroom-Analytics-123/agent-boundary-conformance - Live demo: https://fineprint.vroomanalytics.com - Scorecard, method, limitations: repo README and docs/LIMITATIONS.md ## Reproduce it git clone https://github.com/Vroom-Analytics-123/agent-boundary-conformance.git python3 -m pytest tests/ -q (106 passed, stdlib only) Apache-2.0. The submission contains no closed-layer code, dashboards, configs, pricing, or customer data.