Proofroom — evidence for AI agents
Close enterprise AI deals with proof, not promises.
A verifiable record of what your agents did — ready for security, procurement and governance. Hash chains are how it works; the commercial outcome is why it exists.
- Event type
- exception raised
- Authority
- in scope
- Approval
- not required
- Evidence level
- self reported
- Chain
- chain valid
- Recorded
- 2026-09-30 08:59 UTC
Client zero
This site is client zero. Audit us.
Live company vitals from named database queries — never model-written numbers. When in doubt we publish the unflattering figure.
Days since first internal evidence
112
Days since oldest evidence_event on an internal-agent room.
cq_days_agent_run
Evidence events (internal)
15,569
Count of evidence_events for is_internal agents.
cq_events_recorded
Days since human intervention
8
Days since last human-marked intervention event (metadata.human_intervention=true or approval_completed with actor human). Resets are themselves chain events.
cq_days_since_human_intervention
Tier-3 approval queue depth
5
Pending rows in approval_requests (status=pending).
cq_tier3_queue_depth
Open integrity incidents
0
Unresolved integrity_incidents (unflattering number on purpose).
cq_open_integrity_incidents
As of 2026-09-30 23:53:59 UTC · Skill file · Decision policy
Ninety seconds: one story
Request → action → receipt → verify → scope. Not a 300-line room as first contact.
Step 1 of 5
1. Request
A buyer’s security questionnaire asks: what did this agent do last week, and was it in scope?
How do we know the agent actually reported?
Capture grade is part of the contract. Arrow styling carries the meaning — dashed means chosen, solid through the SDK means automatic, receipt-before-action means gated.
Watched
Agent reports itself
Proves: Events that arrived are on the chain; silence publicly degrades the room.
Cannot prove: That every action was submitted — gaps are detectable via decay, not impossible.
Instrumented
SDK middleware wraps actions
Proves: Every wrapped action emits a receipt as a side effect — capture does not depend on the model remembering a call.
Cannot prove: Actions outside the wrapped tool layer, or failures before the middleware runs.
Gated
Receipt acknowledged before the action proceeds
Proves: Critical steps waited for a durable ack (require_ack) before continuing.
Cannot prove: Non-gated paths, or fitness/safety of the action itself.
Every room displays its capture grade so readers know what the receipts can and cannot prove. Full explainer · Observe vs Assurance
Declare
Register your agent and its use case: what it is allowed to do and what it must never do. Two minutes, one skill file or MCP call.
Receipts accrue
As your agent works, material actions become hash-chained action receipts. Fire-and-forget: never blocks your agent, metadata only.
Share
When a buyer asks, hand them a live proof room or a framework-mapped evidence pack. Structural facts public; full detail under review access.
This company runs on agents
Six internal agents operate Proofroom (plus Verification), and every material action they take is receipted in public. On day one, our own spend cap caught a duplicate instance double-firing every job. The blocked runs and the exception are still in the chain, because the record is append-only. That is the point.
Receipts across internal agents
6387
Chain integrity
9 chains intact
Oldest chain
112 days
Agent builders and agencies
You sell AI-assisted work into enterprises. Proofroom gives you a credential to attach to the deal: declared scope, receipts, honest evidence levels.
Buyers and governance teams
You assess vendors deploying agents. Start with our free AI Agent Vendor Assessment Pack: 25 questions, no signup.
What is missing elsewhere
Identity binds an agent to a person or organisation. Authorisation answers whether an action may proceed at a moment in time. Neither is a continuous, buyer-inspectable record of what the agent actually did against a declared scope. That continuous record is what a proof room holds. How the three layers fit together.
Identity proves who. Authorisation proves may. Proofroom proves did.
What a proof room is not
A proof room is not certification and not a safety claim. Self-registered evidence is labelled self-reported, the weakest of three evidence levels, and every room says on its face what is and is not evidenced. Judge the artifact because it is honest, not because it is green.
Everything free for the first 50 founding accounts — founding price locked forever.
50 founding accounts remaining (live · 0 of 50 claimed). $59/mo is the post-founding Pro anchor only — no payment processing on this site today. See pricing