DTL Verify — independently test DTL, sign your verdict

A public registry of independent, cryptographically-signed audits of the Deterministic Taxonomy Lanes (DTL) verification API.

Most "trust us" pages are the vendor's own claims. This is the opposite: you run the audit yourself, you sign your own verdict with your own key, and it joins a public list that anyone can verify. Nobody — including us — can forge or alter a signature in the registry. It's Certificate Transparency, applied to the auditors.

If you are a researcher, lab, or firm evaluating DTL: this repo lets you verify our claims in ~5 minutes and leave a signed record of what you found — pass or fail. We want the failures published too.

🔴 Live proof (zero install)

• Verification wall — every signed audit, its signature checked in your browser: https://kyal102.github.io/dtl-verify/ (enable GitHub Pages → docs/).
• The standard — this is the reference implementation of SVR-1, an open standard for making any AI-verification claim independently provable. See SPEC.md.

---

What you're verifying (6 falsifiable claims)

#	Claim	Checked by
C1	Determinism — same input → identical signed certificate_hash	run_audit.py
C2	Replayability — sealed evidence re-executes bit-for-bit	run_audit.py
C3	Correctness — verdicts agree with reference tools (sympy, pint)	run_audit.py (skipped if libs absent)
C4	Conservatism — never a false "verified" on bad input	run_audit.py
C5	Failure honesty — unavailable gates 503, tokens refunded, no fabricated output	run_audit.py
C6	Non-repudiation — every verdict Ed25519-signed + tamper-evident chain, verifiable offline	run_audit.py + dtl_client.py

Full methodology: AUDIT_PROTOCOL.md.

---

Quickstart (≈5 minutes)

git clone https://github.com/kyal102/dtl-verify
cd dtl-verify
pip install -r requirements.txt

# 1. Run the audit against the live API (ask Kyal for an evaluation key)
python run_audit.py --key jv3_live_YOUR_EVAL_KEY
#   → writes audit_report.json  (what you observed, per claim)

# 2. Sign YOUR verdict with YOUR key (a keypair is generated on first run)
python sign_report.py audit_report.json \
    --name "Prof. Jane Doe" --affiliation "MIT CSAIL"
#   → writes signed/prof-jane-doe.json  (your signed attestation)

# 3. Submit it: open a pull request adding your file to registry/
#    (or email it to admin@jvi3.com and we'll add it)

# Verify the whole registry yourself — every signature, offline:
python verify_registry.py

You never send us your private key. You sign locally; only your public key and signature travel. Anyone can re-run verify_registry.py and confirm.

---

Why this is different

• Self-generated proof has a ceiling. A vendor's own benchmarks are discounted. Independent, signed verification is not.
• The registry is tamper-evident. Each entry is signed by its author's key. We cannot add, edit, or remove a verdict without it failing verification.
• Negative results are welcome. If a claim fails on your run, sign that. A published, honest failure is more credible than a wall of green.

Honest scope

• This audits the public metered API — the black-box behaviour, not our source. That's deliberate: you verify claims without trusting our code.
• Some gates are marked preview in the catalog and are expected to 503 (that's claim C5, not a failure).
• Correctness cross-checks (C3) need sympy/pint; if absent they're reported as skipped, never as passed.

Contact

admin@jvi3.com · no NDA required for the audit · DTL patents AU 2026905289, 2026905506 (provisional).