A cryptographically signed Trust Receipt. A public verify endpoint. An audit trail you can hand an LP — so a claim holds up when it's challenged, not just when it's read.
When a run finishes, you can download a Trust Receipt — a signed, verifiable record of the rubric score, per-section provenance, and the exact prompt version that produced the work.
Every run is scored 0–100 against a structural, content, and calibration rubric. The score travels in the receipt.
Each section is hashed with SHA-512. Change one comma after the fact and verification fails.
The whole receipt is signed with a platform secret. The signature is what makes the receipt the receipt.
Verification recomputes the signature over the stored payload and re-derives the chain from the sections. If either fails, the receipt is tampered. Deterministic, dependency-free, and mathematically sound.
Your LP doesn't need to log into Zyphv to check the report you sent them.
They open the PDF, scan the QR code, and a public endpoint confirms the receipt is intact — or doesn't.
The verify URL is rate-limited per IP and leaks no internal information. It is, on purpose, the most boring API we run.
Every memo in your IC packet carries a receipt. When an LP asks where a number came from, you can show the chain — from the report section to the prompt version that produced it.
Hand a target's lender the Trust Receipt alongside the DD memo. They confirm the report hasn't been edited — without you handing over access to your tooling.
Sell-side and buy-side verify against the same public endpoint. The receipt is the reference — and neither side controls it.
Prompt changes are logged, deploys timestamped, and every score chains to its prompt-version hash — mapped in our compliance posture documentation.
The receipt is the surface. Underneath is a closed loop that keeps quality up — with a human in the loop every time it changes.
Structural, content, and calibration rubric. A second-pass LLM judge. Per-agent scores, all chained to the exact prompt version that produced the work.
Golden fixtures re-scored against today's prompts. Drift alerts when fleet quality dips — the system catches its own regressions before customers do.
A model proposes prompt edits; an admin approves; an A/B test runs against fixtures; a 10% canary deploys with auto-rollback. Nothing promotes without a human.
Scoring coverage, judge latency, queue depth, alert backlog — traffic-light KPIs with alerting on the red ones. The eval system has its own observability.
A Trust Receipt proves integrity, not truth. It proves the report you're reading is exactly what the system produced — sourced and scored — and that nobody edited it after. It does not, and cannot, prove every claim is correct. That's why we score, judge, and improve every run, and why the rubric travels in the receipt.
Start with 150 free credits. The receipt is on the house.
Traceable by design.