enterprise · the pack
Agents your auditor can read.
Every run in a department leaves evidence, an approval record and a cost line. The auditor reads the same receipt the operator did.
Most AI products ask an organisation to trust a log the product writes about itself. Q is built the other way round. The run is held in a journal it cannot rewrite, the decisions a person made are records in a table of their own, and the money is reserved before a model is called. The auditor's pack is assembled from those, not from a summary.
This page is laid out as that pack. Each exhibit says what is enforced, where it is enforced and what it looks like on the record. The field names in the specimens are the real ones; the values are illustrative.
the pack, assembled
From the journal to the seal.
01 · the journal
The run writes itself down.
Every step, decision and cost is appended to a journal as it happens. Nothing is written afterwards.
the exhibits
What is enforced, and where.
Six controls, each with the record it leaves behind.
A. Isolation
Row-level security is forced in PostgreSQL with a runtime role that is not a superuser. A request without tenant context sees no rows. Isolation is a property of the schema, not of the prompt.
The tenant is set per transaction, and every table that holds a run, a decision or a cost carries it on each row.
enforced today
the request context
- tenant_id
- ten_7f3a…
- principal_hash
- sha256 · 64 chars
- runtime role
- not a superuser
- row-level security
- forced
- rows without tenant context
- 0
B. The approval record
When a run reaches an action that matters, it stops and writes a record: tenant, principal, run, the action and the version of that action. A person decides. The record keeps the decision, when it expires and the moment it is spent.
Terminal records are retained. Plan capacity and an archival policy before large-scale operation.
in service today · archival policy to plan
q_service.approvals
- run_id
- run_01j9…
- approval_id
- apr_4c1e…
- action_id · version
- send_email · 3
- state
- consumed
- decision
- approved
- expires_at
- 2026-10-14 09:31:00
- consumed_at
- 2026-10-14 09:12:42
specimen · values are illustrative
C. The journal
Every step, decision and cost of a run is appended to a journal in sequence. An entry is never rewritten; an append that conflicts with what is already there is refused. Runs replay from the journal, and audit evidence is assembled from it.
The journal is kept per run and per tenant, under the same row-level security as everything else.
in service today
q_service.event_journal
- run_id
- run_01j9…
- sequence
- 17
- cursor
- run_01j9…:17
- type
- tool.result
- evidence
- doc_88e2… · 3 pages
- size_bytes
- 1 842
specimen · values are illustrative
D. The spend line
Cost is reserved before a model is called and settled after, then reconciled against the provider's own records. An attempt that never settles stays held against the budget rather than forgotten.
Reconciliation is implemented. Qualification under representative load is pending.
implemented · load qualification pending
the spend attempt
- model · provider
- named on the receipt
- reservation_usd
- 0.0184
- accounted_usd
- 0.0171
- state
- reconciled
- provider_reference
- inv_9b…
- settled_at
- 2026-10-14 09:12:40
specimen · values are illustrative
E. The guards
Input guards refuse injection and malformed requests. Output guards hold back leaked personal data and claims with no source. Cost guards keep each tenant inside its budget and its rate limit.
A hard budget is never extended by the run. Extending it is a decision a person makes, and it is recorded like any other.
enforced today
guard results for one run
- input
- passed
- output · personal data
- none found
- output · grounding
- 4 sources cited
- cost
- within budget
- hard budget
- extension needs a decision
F. The provider, under policy
Requests go to a provider that meets your compliance requirement for that task. The policy names the providers allowed for the task; the run uses one of them, and the receipt says which.
Changing the provider is a change of configuration, not of application code.
in service today
the routing decision
- task
- contract classification
- policy
- providers allowed for the task
- provider · model
- named on the receipt
- application code
- unchanged
the auditor's view
A department, for a period, on one page.
What its roles did, which runs waited for a person, what was rejected or expired, and what it cost. Assembled from the journal and the approval records, filed by department.
assurance pack · cover sheet
specimen · values are illustrative
Finance, October 2026
The export format is in design. The records it reads from exist today.
- department
- finance
- period
- 2026-10
- runs
- 142
- waited for a person
- 23
- rejected
- 2
- expired before dispatch
- 1
- evidence attached
- 142 / 142
- spend
- USD 61.20
sectors
Where the receipt matters most.
Work that touches money, tenants and regulators. Policy generation, compliance gap analysis, contract risk, personal-data tagging and audit evidence packaging run as durable workflows with approval where it matters.
- Property
- the workThe chase, the renewal, the listing, the compliance pack.
- the departments it falls inOperations and finance.
- Legal
- the workContract classification, precedent retrieval, contract risk, gap analysis, personal-data tagging.
- the departments it falls inRecords.
- Insurance
- the workClaims analysis, policy wording search, compliance documentation.
- the departments it falls inRecords and finance.
- Financial services
- the workRegulatory documentation, audit evidence packs, report analysis.
- the departments it falls inRecords and finance.
- Professional services
- the workKnowledge search, report generation, proposal assembly, the chase.
- the departments it falls inRecords and finance.
- Public sector
- the workPolicy search, compliance analysis, briefing documents, information requests.
- the departments it falls inRecords.
where it stands
Measured, and not yet accepted.
No production gate has yet been passed. Local qualification, publication and production acceptance are reported as distinct states, and the readiness report says which has been reached for each part.
Approvals are per run. Standing mandates, under which a role may act without a fresh decision each time, are in design and not yet live.
The assurance export is in design. The journal and the approval records it will read from are in service.
Spend reconciliation is implemented; qualification under representative load is pending. Terminal approval records are retained, and an archival policy is yours to set before large-scale operation.