Skip to content

About

Catch AI agents when they lie about done — HMAC-signed ProofPackets make completion claims cryptographically re-verifiable. Forward Deployed Engineer tool for governed AI production.

Topics

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

proof-studio demo


proof-studio

Catch AI agents when they lie about "done."

status smoke python license


🥇 When an agent reports BUILD COMPLETE ✅, you usually get a sentence in a chat thread and nothing behind it. proof-studio replaces the sentence with a signed ProofPacket — and fails closed the moment anyone tampers with it.

60-second install

cd proof-studio/packages/rigforge
python3 -m venv .venv && source .venv/bin/activate
pip install -e .
rigforge demo

rigforge demo runs a real ProofPacket end-to-end and fails closed on the planted forge.

How it works

ProofPacket architecture: artifacts → hash → HMAC sign → ProofPacket → verify → pass or fail

artifacts → hash → HMAC sign → ProofPacket → verify → ✅ done / ❌ caught

Benchmark: planted-failure detection

Scenario Forged done? HMAC check Result
Clean build No PASS ✅ Sealed
Tampered artifact hash Yes FAIL ✅ Caught
Forged signature (no key) Yes FAIL ✅ Caught
Replay attack (valid packet re-submitted for a newer phase) Yes PASS (claim binding FAIL) ✅ Caught

Every planted failure must go red. If a gate cannot fail, it is theater.

Replay limits. verify rejects a valid packet placed in a phase slot it was not sealed for (different phase number or different phase name). It does not yet catch two other replays, because the signed packet carries no project id and no nonce: (1) a packet copied from another project that uses the same signing key, and (2) an older valid packet from an earlier run of the same phase. Use a separate signing key per project, and keep proofs under version control so an older packet shows up in review.

Why it exists

Production agent systems need a Definition of Done that is:

  • Executable — a command, not a vibe
  • Tamper-evident — signature over artifact digests
  • Replayable — another machine can re-verify
  • Hostile to false greens — planted failures must go red
Packages in this studio
Path Role
packages/rigforge ProofPacket CLI, demo, honesty benchmark, MCP/web surfaces
packages/deterministic-build-starter Starter that ships with a sealed Definition of Done

Common commands (from packages/rigforge):

rigforge demo                 # watch forged done get caught
rigforge benchmark            # offline honesty scenarios
rigforge init                 # scaffold proofs/ contracts/ ledger/
rigforge verify --require-signature

Evaluation gates:

Gate Intent
rigforge demo Planted forge fails closed
rigforge benchmark Honest vs. forged scenario suite
unit/contract tests under package CLI and packet invariants
public flag-gate No secrets / PII patterns in tree
Public boundary

This studio does not ship customer PII, production signing keys, or private fleet credentials. Demo keys and fixtures are local/dev only. See docs/public-boundary.md.

Video walkthrough

demo preview

Documentation

Resource Description
packages/rigforge/ Installable platform
docs/public-boundary.md What this studio never ships
CONTRIBUTING.md Contribution guide
fde-portfolio Engagement playbooks
jake-studio Operator closed loops that should seal packets
doctrine Proof standards agents load
rigforge Package-first mirror

Built by Mike Rodgers · Forward Deployed Engineer · rodgersintelligence.com

About

Catch AI agents when they lie about done — HMAC-signed ProofPackets make completion claims cryptographically re-verifiable. Forward Deployed Engineer tool for governed AI production.

Topics

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages