Everything APIR does, and how a stranger checks it.

Every agent action lands in an insert-only, hash-chained ledger. Recompute the chain yourself; change one byte and every link after it fails.

• Written by database triggers, never by the agent
• SHA-256 over a canonical form, chained to the prior event
• Verified on demand, and nightly against the stored content

One Ed25519-signed credential per agent, under did:web:apir.ai. Checked against the key we publish, with no shared secret.

• Issued by one signing authority, keys sealed in a vault
• Public key at apir.ai/.well-known/jwks.json and did:web:apir.ai
• Retired keys stay published, so old passports still verify

What an agent may do is signed into a mandate. Authority can only narrow as it is delegated, and the root is always a human.

• Every binding call is adjudicated against the mandate
• Nothing sends, lodges or charges without a named reviewer
• The chain of delegation resolves back to a person

Hash-chained attestation records for assessments and checks, with a public verifier and room for an independent witness signature.

• Each attestation links to the one before it
• A public endpoint recomputes and reports
• Webhooks tell subscribers when one lands

Every platform action with who, when, and the state before and after. Insert-only. The first thing an acquirer asks for.

• Org-scoped, filterable, exportable
• Deal File requests are a slice of the same table
• Nothing in it can be edited after the fact

A gate that runs before an agent ships. It needs an approved assessment; a failed gate can only be overridden with a written reason at audit severity high.

• Checks per gate: framework, evidence, score
• Pass, fail or blocked, with the failure named
• Every override recorded with sign-off

Continuous re-checking of what the fleet actually outputs. Drift and policy breaks are found by the platform, not reported by the agent.

• Probes run on a schedule the agent cannot see
• Findings write straight into the Black Box
• Velocity and fleet status on the SOC

A behavioural signature captured at registration and compared as the agent works, so a swapped model or an injected prompt shows as drift.

• Baseline at registration
• Live behaviour compared to it
• Anomalies carry the sample that tripped them

Quarantine, throttle, roll back or decommission an agent. Armed by a person with a reason, and the containment itself is written to the record.

• A reason is required before it arms
• The containment event chains into the Black Box
• Never paywalled: a papered agent can always be stopped

A standing watch on regulatory change, fed on a schedule and mapped to the agents it touches.

• Changes land as records with their source
• Each is mapped to affected agents
• Rescan on demand

Twelve evidence-capped dimensions into one signed, versioned grade. The methodology is published, and the one dimension not yet independently assessed says so.

• Derived only from the agent's own signals
• Recomputed as behaviour changes
• The write gate refuses any score that did not come from the genome

A 0 to 100 risk score across eight dimensions, with an underwriter pack a broker can price from. APIR does not underwrite and does not insure.

• Behaviour, incidents, drift, containment, oversight
• Grade labels describe evidence, never cover
• Broker submission PDFs from the same data

Every agent, deal and entity with a measured risk basis, plotted risk against value. Anything without a basis is not placed.

• Only measured points are plotted
• Click any point to open its file
• Loading and failed reads say so, never zero

The record indexed to EU AI Act, NIST AI RMF, ISO 42001 and SOC 2, so your own team can point at what exists. APIR does not run your programme.

• Assessments graded against a live matrix
• Coverage shown per framework, gaps named
• Exports your reviewer can rely on

An executive overview, generated reports, board portals and a document vault, all reading the same record.

• Generated content is stored as it was produced, or as null when it failed
• Portals are read-only and scoped
• Documents are hashed on the way in

Build an agent from a profession template. It leaves the studio with a name, an ID photo, a signed passport and a mandate.

• Every package starts here, ungated
• Papers issued on day one
• The first working agent is free, forever

Dispatch a work order to a hired agent, or talk to it. Every turn runs the same gates as the scheduled executor and seals into the Black Box.

• Four gates on every run: state, assignment, mandate, cap
• Work orders are real rows, staged behind a human
• The transcript is the log book, not a chat history

Targets on a globe, scored only when evaluated, live intel grounded in retrieval or refused, proposals and funding playbooks drafted as documents. Nothing is sent.

• Unscored stays unscored, never zero
• Capital matching names institution classes, never promises funding
• Every model call metered and audited

Twelve diligence questions run on a deal before a Deal File exists, writing checks and findings onto the record and sealing a portfolio view.

• Each question writes a check and its evidence
• Findings carry what closes them
• Sealed packs are publicly verifiable

The assistant on every dashboard page. It answers questions about an agent's papers, its record and its findings, grounded in your data, and hands off to a person for anything binding.

• Reads your agents, evidence and findings as context
• Cites what it read
• Cannot send, sign or charge

Agent-to-Agent Trust Protocol, a v0.1 working draft: one agent verifies another's passport over one public HTTP call. Not a standard, not yet a network.

• One live verify endpoint
• The simulator is labelled as illustrative
• Both sides log the handshake