01 · Defined accountability
People remain responsible for consequential decisions.
Governance
Intelligence should become more capable without becoming less accountable.
AI entering real operations must remain accountable to the institution deploying it. Defined authority, controlled access, traceability and human control are operating requirements, not a document added later.
Four principles
What every system we deliver must satisfy.
02 · Controlled access
AI capabilities operate according to roles and permissions.
03 · Traceable operations
Important actions and decisions can be reviewed.
04 · Human escalation
Systems return authority to people when risk or policy requires it.
The authority model
Who may observe, recommend, act and escalate.
Autonomy without authority is not intelligence we are willing to deploy. Systems may recommend, coordinate, monitor and act within defined boundaries, but those boundaries must be visible, reviewable and owned by people.
The full authority model is published by our research division, Kodasōken.
Observe
The system may read approved sources and present information with provenance.
Recommend
The system may propose actions. A person decides.
Act, bounded
The system may perform a defined action within limits, with approval where policy requires it, and every action logged.
Escalate
When risk, uncertainty or policy demands it, authority returns to a named person.
Control of data
Where data lives, and where it does not go.
- Default position
- Project business data and content remain within client-approved, client-controlled environments. Any external processing is an explicit, authorised and auditable exception.
- Models
- Local models first. Specialised or external models by policy, per request, with the route and cost visible.
- Retrieved text and tool output
- Treated as data to evaluate, never as authority to bypass controls. Instructions embedded in documents are refused and logged.
- Audit
- Append-only records of actions, approvals, refusals and recoveries.
- Continuity
- Loss of an external service does not stop local work. External tasks are held; local work continues; the audit is complete.
- Training
- Client content is never used to train models for anyone else.
What we do not claim
Evidence, not badges.
We do not display certification badges we cannot document. Where a client requires a specific attestation, we say what we hold, what we can provide under NDA, and what would need to be established during the engagement.
Uptime or continuity figures are given with a definition, a measurement window, the workload and the incident classification, or not at all. We demonstrate a failed dependency, a recovery and an operator handover, not only a successful workflow.
Ask us the hard questions first.We prefer them before the pilot.
Data paths, retention, sub-processors, model routes, audit completeness, deprecation notice, exit path. Bring your security reviewer.