§ Platform / Decision Receipts

Every AI decision gets a receipt. Signed, replayable, yours.

Most AI systems produce outputs. They do not produce proof. A Decision Receipt is the proof: a machine-verifiable evidence package showing the decision was made under known conditions, with traceable evidence, declared policy, and a reproducible result.

Open Decision Receipt ↗Read the Spec — v1.0 Verify a Receipt ↗

§ 01

The five questions a receipt answers

When a machine-generated decision is challenged — by an auditor, a regulator, a review board, or opposing counsel — these are the questions that get asked. A receipt answers all five, in writing, before anyone asks.

Question 1

Why was this decision made?

The receipt records what was decided, what changed, and the reasoning context the decision was produced under — not a reconstruction after the fact.

Question 2

What evidence supported it?

Every input is content-addressed and hashed. The evidence that informed the decision is enumerated, identified deterministically, and bound into the receipt.

Question 3

Can it be reproduced?

Receipts carry everything needed to replay the decision deterministically — same inputs, same policies, same result, bit-identical, including air-gapped.

Question 4

Was policy followed?

Deny-by-default rules are evaluated at decision time and the full evaluation log — which rules fired, which passed — is recorded in the receipt.

Question 5

Who is accountable?

Generation metadata, provenance, and an Ed25519 signature tie the decision to the system and process that produced it. Accountability is structural, not asserted.

If a system cannot answer these five questions, its decisions are not admissible. That is the standard receipts were built to meet — reviewable, reproducible, defensible.

§ 02

Anatomy of a receipt

A receipt is an evidence bundle, not a log line. Report, metrics, and integrity stamp are required; policy logs and supply-chain provenance ride along when the decision warrants them.

Deterministic evidence IDs. Identifiers are derived from canonical content — no timestamps, no hostnames, no random UUIDs. Identical content yields identical IDs, on any machine, forever.
Deny-by-default policy evaluation. Nothing is admissible until a rule admits it. The receipt records the full verdict trail, ready for export to your auditors.
Deterministic replay. Recorded inputs and declared policy versions make the decision reproducible on demand — the audit answer is a replay, not a recollection.
Ed25519 signature. The bundle is signed at issuance. Verification is independent — no Summit account, no Summit infrastructure, no trust in us required.

SUM-RCPT-v1-4c19…b3e7ADMIT

decision: procurement-eval.recommend.18
evidence: 5 artifacts · sha256 verified
evidence-id: SUM-EVID-v1-…-a91f44c2e07d3b18
policy: deny-by-default · 11/11 evaluated · 0 violations
replay: deterministic · bit-identical
signature: ed25519:2f8c…77d1 · valid
issued: 2026-06-12T09:41:53Z

fig. 1 — report, metrics, stamp; policy log attached.

§ 03

An open standard, not a vendor format

The Decision Receipt Specification v1.0 is published under CC BY 4.0 — structure, evidence-ID determinism rules, validation requirements, and trust model. Receipts are designed to be verified by people who never buy from us.

Issue

Issue your first receipt today.

The product site walks the full lifecycle — issue, inspect, replay, verify. When you're ready to instrument a real workflow, the 10-day pilot does exactly that.

Open Decision Receipt API Playground ↗