Permissions
The actor, assignment and requested action must match the current job state.
Security
Safe AI is an execution boundary, not a model claim.
Kinemica treats plans and analysis from a model as untrusted proposals. Deterministic application and policy code decides whether an assignment, action or completion transition is allowed.
Model and Kinemica authority boundary
MODEL OUTPUT
Suggestion
- Proposed tasks
- Recommended worker
- Short rationale
- Evidence requirements
KINEMICA POLICY
Decision
- Identity and status
- Qualification and capability
- Location, risk and state
- Approval and evidence
Controls
Permission is checked at the moment of action.
A plan review is not a permanent pass. Time-sensitive facts are checked again before dispatch, task start, completion and job closeout.
Qualification checks
Human qualifications and machine capabilities are explicit, active facts.
Policy enforcement
Deterministic rules return ALLOW, BLOCK or REQUIRE APPROVAL.
Approval gates
Sensitive permitted work waits for a current, bounded approval from an authorised person.
Evidence requirements
Tasks cannot complete until the declared proof is present and acceptable.
Independent validation
Model output is schema checked and then evaluated separately from permission.
Audit records
Material actions and decisions are appended with actors, facts, timestamps and outcomes.
Human approval
Approval is specific, attributable and limited.
When policy permits an action only after approval, Kinemica checks the approver’s active identity and authority, then binds that approval to the task, worker, action and facts being reviewed.
Approval cannot convert a prohibited action into an allowed action. A BLOCK remains a block until relevant facts change and a new evaluation permits reconsideration.
Accountability
The record explains what happened and why.
Operational audit
Plans, assignments, policy decisions, approvals, dispatches, evidence, verifications, exceptions and closeout remain visible in sequence.
Integrity boundary
Audit events are append-only for normal users and hash-linked to detect retained-log changes. Hash chaining alone is not legal notarisation or independent proof against a database owner.
Current deployment boundaries
Real-world integrations require a controlled operating scope.
- The public robot workflow is simulated and controls no real machine.
- Model and software verification cannot prove every real-world fact.
- Real machine adapters require a separate safety architecture and review.
- Pilot deployment, retention and operational controls must be agreed before use.