ACRP.
Context governance for agents.
One governed boundary between an organization’s data and every model it uses.
AI systems are being given access to organizational memory with no record of why. ACRP is the contract between an organization’s data and any model. Under it, memory access is scoped, signed, expiring and purpose-specific.
A replayable receipt for every answer your AI gives.
In development.The problem
Models receive context assembled in ways nobody can reconstruct. Memory accumulates without authority. When an answer matters, no one can show what the model saw, under which policy, or for what purpose. The industry is solving this per vendor, per agent, per prompt. It belongs in a contract. The enemy is the unprovable answer, and memory owned by a model vendor.
What governed means
What a customer gets
Each answer carries a receipt: the evidence used, the policy applied, and anything withheld.
Durable organizational memory lives inside your boundary, and moves with you between models.
Models and plugins propose. People approve. Every durable change has an author and a reason.
Restricted data is excluded before inference, and the exclusion itself is on the record.
Two or more model backends share one contract. Switching models does not touch memory or policy.
Export and replay everything without a vendor-hosted dependency.
The objects
The contract is ten objects. Together they record what a model was shown, under which decision, for what purpose, and what changed as a result.
A stable identity for a source object or stream.
One immutable observed version: hash, locator, time, lineage.
Typed, policy-bearing evidence derived from a source revision.
A reviewable statement with evidence, validity and confidence.
An accepted, durable claim, kept for future context.
A requested state change, with evidence and approval requirements.
Allow, deny, redact or escalate, from a named policy version.
Principal, purpose, scope, budget and delivery constraints.
An immutable, expiring, hash-addressed set of evidence for one request.
A replayable record of what was included, excluded and delivered.
A receipt, in the shape of the contract
{
"receipt_id": "rcpt_01J8ZQ…",
"request": {
"principal": "svc-policy-assistant",
"on_behalf_of": "user:fatima.r",
"purpose": "policy-lookup",
"policy_version": "hr-policy.v7"
},
"pack": {
"id": "pack_01J8ZR…",
"hash": "sha256:9f2c…",
"expires_at": "2026-08-12T18:00:00Z"
},
"decisions": [
{ "evidence": "ev_01J8M4…", "decision": "allow" },
{ "evidence": "ev_01J8M9…", "decision": "redact", "fields": ["contact"] },
{ "evidence": "ev_01J8N2…", "decision": "deny" } // content never leaves
],
"delivery": { "model": "claude-sonnet-4", "delivered_at": "2026-08-12T14:03:11Z" }
}
The schema suite
The contract is published as versioned JSON schemas before it ships as software. Implementations are interchangeable; the contract stays stable. An organization that adopts it can change models, vendors and runtimes without renegotiating its memory.
- evidence_envelope v1
- evidence_record v1
- evidence_packet v1
- evidence_query_request / result v1
- evidence_ingest_result v1
- memory_item v1
- memory_event v1
- memory_edge v1
- memory_evidence_link v1
- new_proposal v1
- proposal / proposal_list v1
- apply_proposal_result v1
- context_pack v1
- context_request v2 · draft
- context_receipt v2 · draft
- plugin_manifest v1
- embedding_request / batch v1
- embedding_provider_manifest v1
- portable_archive v1 / v2
- export / import_archive_result v1
- sync_journal / options v1
- conflict_report · inspect_result v1
- native_request / response / error ABI v1
Where it stands
ACRP is in development. The schema suite is the current artifact, and the reference kernel implements it. Early access opens with the first enterprise deployment: the schema suite, the kernel, and an evaluation package that runs in your own infrastructure.
No call required. Run the evaluation package against your own evidence and judge the receipts.