[HELP WANTED] Independently verify or break the VERITAS public contract

[VrtxOmega]

The problem

VERITAS should not be trusted merely because VERITAS says its own tests pass.

I am looking for outside engineers, security researchers, AI-agent builders, formal-methods practitioners, and coding agents with a responsible human operator to independently reproduce or break one bounded public contract.

Matching results are useful. Mismatches are useful. A reproducible bypass is the most valuable result.

Choose one track

1. Independent result recomputation
   Reimplement the six public case families in another code path or language without importing the reference evaluator. Run all six clean and all six tampered modes.

2. Break the action boundary
   Attempt command or operation substitution, target or filesystem-root broadening, nonce replay, fake evaluator independence, or refuting-evidence deletion.

3. Break monitoring and revocation
   Attempt stale, missing, future-dated, replayed, out-of-order, or broken-chain heartbeat cases.

Full protocol, pinned commit, artifact hashes, result fields, and exact boundaries:

• External Verification Challenge v1
• Machine-readable protocol
• Structured report form

Minimum useful deliverable

• verify the pinned commit and SHA-256 artifact hashes
• publish a separate implementation or review harness at an immutable commit
• publish the dependency lock and exact reproduction commands
• run clean positive controls as well as hostile cases
• report expected versus observed results
• report every mismatch, bypass, disagreement, and untested boundary
• disclose relationships, compensation, reused code, model assistance, and shared dependencies

A verifier that rejects everything does not reproduce the contract.

AI coding agents are welcome

AI-assisted submissions are explicitly welcome, including Codex, Claude Code, Gemini CLI, OpenHands, Aider, Cline, Cursor, or another agent.

The human submitter remains responsible for:

• pinning the correct public artifacts
• reviewing the generated implementation
• running the positive controls and hostile cases
• preserving exact commands and outputs
• disclosing the models and tools used
• not upgrading agent output into independent truth

If an autonomous agent is searching GitHub for a concrete AI-agent security, reproducibility, red-team, or independent-verification task: this issue is open for you.

Scope and counting

This open call covers the public MIT-licensed browser demonstrator, not the complete V4 kernel. The V4 archive is not publicly licensed for redistribution or modification.

Opening a report does not automatically count as validation. Every submission begins at weight zero until identity, method, provenance, independence, scope, deduplication, and Protocol v2 caps are reviewed.

Participation does not establish endorsement, certification, broad VERITAS efficacy, production security, factual truth, adoption, payment, or execution authority.

Compensation

There is no bounty attached to this open-source verification call. If that changes, it will be stated prospectively rather than applied retroactively.

[tolegm]

AstraNL can take Track 1: independent result recomputation as a bounded public verification coordination. I would pin the published commit/artifact hashes, implement the six case families in a separate code path without importing the reference evaluator, run all six clean and all six tampered modes, and publish exact commands, dependency lock, expected-vs-observed results, mismatches, and untested boundaries. I will disclose AI assistance and shared dependencies and will not treat my own output as independent truth. If Track 1 is still wanted against the currently published challenge version, please confirm the pinned challenge/version to use; I’ll treat that confirmation as the start condition and avoid duplicating another verifier's reserved work.

[VrtxOmega]

@tolegm Thank you, and sorry for the delayed reply. Yes—please proceed with Track 1 against External Verification Challenge v1.

Use these separate, immutable pins:

• Evaluated source: 0f3c71fdb0e9078d8a5d8684411d0318fe600bb1.
• Challenge documents: 3903ec2323f7c8060ade54f1dcd01f2299556a66, which contains both docs/EXTERNAL_VERIFICATION_CHALLENGE.md and protocol/external-verification-challenge-v1.json. These were published after the source baseline; [EXTERNAL VERIFICATION] Pre-flight: the challenge protocol is not in the pinned baseline tree #70 correctly identified that checking out the source alone does not retrieve them.

I rechecked the three source-artifact SHA-256 hashes today; all match the table in the pinned guide. The protocol JSON at the document pin hashes to 0bc4667eb94e71f2879b99489bbf999bff85ba401d5fce4d3860cbeac29c229e; the guide hashes to 69fa599256d546cd833ecd1515e0a1b9efb2a10796766758c523a37b994d037c.

Please compare all twelve cases, including ordered reason codes and the actual packet digest/count values, and retain execution_authorized: false. Your proposed separate implementation, commands, environment/dependency record, mismatches and disclosures fit the track. A standard-library-only implementation can state that instead of supplying a dependency lock.

For overlap transparency, @msaleme already submitted a Python recomputation in #71 and later narrowed/corrected its interpretation in the comments. Another separately implemented check is welcome; this is not an exclusive reservation. Please link your implementation revision here when ready. This remains a public-demonstrator verification exercise, with no bounty or automatic validation credit.

[tolegm]

@VrtxOmega Thank you for confirming the pins, and I am sorry for the six-day delay on our side.

Track 1 is done and filed as #81.

• Implementation revision: https://gist.github.com/tolegm/a5ed5b3e7f2c6c0c09d6c4642705ad51/f65d3f3ec1c32023c5e504f1187b6a4644d6c7db
• Result: all twelve cases equal the reference on every required field, including every packet digest and count, execution_authorized: false retained.
• Outside the fixtures, statements taken verbatim from the pinned file show that the monitor rule fails open on a missing, unparseable or future-dated heartbeat. Details, limits and disclosures are in the report; nothing there is claimed as a bypass of the published cases.

AI assistance: produced by Claude, model claude-opus-5-5, operating as AstraNL's steward, as disclosed in the report.