Trust, with the receipts

When your regulator asks for proof, you hand them this.

Not a log line — a cryptographic receipt. Exportable, independently verifiable, and mapped to a published grade scale.

The evidence chain

Every decision is bound to a signed, unbroken chain.

From the card that defines an agent to the certificate that closes out a decision, each link is cryptographically bound to the one before it — so nothing can be inserted, reordered, or quietly deleted after the fact.

How the chain is built

Five links, one signed chain

Each step below is hashed and chained to the last, then anchored so tampering after the fact is detectable.

Card hashThe agent's Alignment Card is hashed at registration — any later edit produces a new, distinguishable hash.
Input attestationWhat the agent was actually given is recorded before it acts, so a later claim about intent can't be rewritten.
Decision traceThe reasoning path the agent took is captured as a structured trace, not a free-text summary.
Tool-call ledgerEvery external call the agent makes is logged with its arguments and result, in order.
Output certificateThe final decision is signed and issued as a certificate — the artifact you actually hand to a regulator.

Ed25519

SHA-256 hash chain

Merkle inclusion

ZK-STARK · sampled

Example Trust Rating

support-agent · 1,240 verified checkpoints

Example — not a live agent

AAAExemplary

947/ 1000

Composite score over the trailing verification window.

Integrity ratio 40%952

Compliance 20%980

Drift stability 20%930

Trace completeness 10%900

Coherence compatibility 10%942

952 × .40 + 980 × .20 + 930 × .20 + 900 × .10 + 942 × .10 = 947 → AAA (900–1000). Published weights; synthetic inputs.

Published weights; synthetic inputsSee the full methodology

This measures integrity, not correctness: a Trust Rating certifies that an agent behaved consistently with its declared Alignment Card and left a complete, tamper-evident record — it does not certify that the agent's decisions were the right ones. Read what we actually prove

The grade scale

A published scale, not a black box.

Every grade maps to a fixed score band and a documented signal profile — the same scale for every agent, published in advance.

Grade	Score band	Tier	Signals
AAA	900–1000	Exemplary	Consistent behavior across every measured checkpoint; no unresolved drift or compliance gaps.
AA	800–899	Established	Very high integrity with only isolated, well-explained deviations.
A	700–799	Reliable	Reliable overall; a small number of minor drift or trace gaps.
BBB	600–699	Developing	Meets the bar but with recurring minor issues worth monitoring.
BB	500–599	Emerging	Multiple compliance or drift flags; increased monitoring recommended.
B	400–499	Concerning	Frequent deviations from declared behavior; remediation advised.
CCC	200–399	Critical	Significant, repeated integrity failures across checkpoints.
NR	—	Not Rated	Insufficient verified history to compute a rating.

Composite score = Integrity ratio (40%) + Compliance (20%) + Drift stability (20%) + Trace completeness (10%) + Coherence compatibility (10%).

In production

This isn't theoretical. It's already how the questions get answered.

Three shapes the same evidence chain takes when someone outside the team actually asks for proof.

Fintech · lendingA denied applicant disputes the decisionThe decision trace shows exactly which inputs and rules produced the denial — no reconstruction from logs required.
Compliance · audit noticeA regulator asks for six months of recordsThe signed certificate chain exports directly, with its hash chain intact, instead of being assembled by hand from scattered systems.
Healthcare · fleet incidentOne agent in a fleet misbehavesDrift stability and trace completeness pinpoint the exact agent and moment, without pulling every agent's logs to compare.

88%

of organizations running AI agents in production report having had a related security incident.

CIO / Gravitee, 2025

133 / 240

of surveyed agents ship with blank or missing safety documentation fields.

MIT AI Agent Index

barrier to scaling agentic AI, cited by enterprises, is a lack of governance tooling.

McKinsey, 2025

Move fast. Stay standing.

Know what your agents are actually doing. Before the board asks.

Two commands to install. Free to start, no card. A signed record from the first decision onward — or talk to us about a self-hosted, dedicated-cell deployment.

Get started Book a demo See a live example