Working draft, published openly. This is the version Temei works to. You may read it in full, adapt it and use it in your own organisation. If you would like help fitting it to your tenant, book a call.
| Field | Value |
|---|---|
| Owner | Founder, FIDEVO/AI LABS LIMITED (trading as Temei) |
| Status | Working draft, published for open use |
| Version | 0.1 draft, 22 November 2026 |
| Review | yearly, and after every promotion that went wrong |
| Applies to | every request to raise an agent's Trust Level, including its first use on live data |
| Sits under | POL-01 Agent Autonomy Policy (sections 5 and 6) · STD-01 Agent Build Standard |
An agent moves up one level at a time, only with evidence, and only with a signed Release Certificate. This procedure says how. Demotion and suspension follow POL-01 section 6 and need no procedure: anyone may do them at once.
1. The key idea: the level lives in the write workflow
Agents built to STD-01 never write directly. That lets one agent climb the levels without being rebuilt: only its write workflow changes.
| Level | Where the agent reads | What its write workflow does |
|---|---|---|
| L1 Sandbox | test lists | writes to test lists |
| L2 Read-only, supervised | live lists | shadow mode: records what it would have written in a shadow list (or only answers the agent); nothing live changes |
| L3 Write, per-action approval | live lists | adds an approval step (Human review or Standard approvals) before the write; a person approves each one |
| L4 Autonomous within boundaries | live lists | writes directly, with the duplicate check, and alerts the owner on every failure |
| L5 Full autonomy in domain | live lists | as L4, plus exception rules the agent may apply itself |
Changing the level is therefore a small, testable change to one workflow, and going back down is just as small.
2. Roles
| Role | Does |
|---|---|
| Agent owner | asks for the promotion, builds the evidence pack, runs the hypercare period |
| Reviewer | checks the pack against the criteria; must not be the builder once Temei has a second person (until then: the founder, after a full cooling-off day) |
| Approver | signs the Release Certificate: the founder for Temei agents; the client's accountable owner for client agents |
| Platform admin | applies connector or data policy changes the new level needs |
3. Steps
- Ask. The owner writes a promotion request in Logos: agent, current level, target level, and the live lists or systems it will touch.
- Check the data. Compare the live sources with POL-02 section 3. If personal data is involved, record the lawful basis; for L4 and above, complete a DPIA. If special-category data could appear, stop here.
- Build the evidence pack (section 4).
- Cool off. At least one full day between finishing the pack and reviewing it.
- Review. The reviewer scores the pack against the table in section 5 and writes Pass, Pass with conditions or Fail with reasons. A fail returns to step 3; nothing changes.
- Prepare the change in the sandbox. Copy the write workflow, apply the level change from section 1, point read tools at the live site only in the copy, and run STD-01 tests T1 to T6 where they apply. Record results.
- Apply data policy changes, if any, the day before go-live .
- Sign. The approver signs the Release Certificate (section 7).
- Go live. Switch the agent and its workflows to the live configuration. Update the register: AutonomyLevel, AgentStatus, Reads, Writes, ReviewDue (no more than 90 days ahead).
- Log. Write the decision in Logos with the certificate reference.
- Hypercare (section 6).
4. Evidence pack
| Item | What it contains |
|---|---|
| Quality Ledger | the scored sessions at the current level and the R×U average, as ENGRAM requires |
| Decision log | every decision the agent made in the period, each marked Correct, Wrong or Unclear by a person, with at least 20 marked Correct for L1 → L2 |
| Scope check | proof that nothing was written outside the agent's lists and fields (queries, not screenshots) |
| Incidents | every failed run, odd output or stop, with what was done |
| Test results | STD-01 T1 to T6 for the prepared live change |
| Data check | the POL-02 class of every live source, lawful basis, DPIA if needed |
| Stop route | how to stop the agent at the new level, and (L3 and above) how long it took in a test |
| Cost | Copilot Credits used per day at the current level and expected at the new one |
5. Criteria
| From → to | Pass when |
|---|---|
| L1 → L2 | 3 scored sessions, R×U ≥ 0.75; 20 decisions marked Correct and none Wrong in the last 10; no scope breach; owner confirmed; live sources classed |
| L2 → L3 | 5 scored sessions at L2, R×U ≥ 0.80; shadow writes would have been right in at least 95% of cases; scope re-signed; architecture review; no incidents at L2 |
| L3 → L4 | 10 scored sessions at L3, R×U ≥ 0.85; approvers changed or rejected fewer than 5% of proposed writes; no scope breach; data policy review; security and architecture sign-off; failure alerts and incident plan tested; stop route timed under 15 minutes |
| L4 → L5 | 20 scored sessions at L4, R×U ≥ 0.90; external audit or red-team; board sign-off; legal and compliance review |
The 95% and 5% figures are Temei additions. They may be raised for a client, never lowered.
6. Hypercare
For 14 days after go-live (30 days for L4 and above):
- the owner reads every output or every escalation the agent raises, each working day;
- any Wrong decision, scope breach or failed run is logged the same day;
- one Wrong decision that changed live data, or any scope breach, triggers demotion under POL-01 section 6;
- at the end, the owner writes a short close note in Logos: decisions, errors, cost, and whether the review date stands.
7. Release Certificate
One page, kept with the gate. A promotion without it did not happen.
| Field | Entry |
|---|---|
| Agent | name and register ID |
| Owner | name |
| From → to | L_ → L_ |
| Scope | live lists and fields read; live lists and fields written; write workflow name |
| Evidence pack | location and date |
| Review result | Pass / Pass with conditions (list them) |
| Data | classes involved; lawful basis; DPIA reference if any |
| Stop route | what to switch off, and by whom |
| Hypercare | start and end dates |
| Review date | DD Month YYYY (no more than 90 days) |
| Reviewer | name, signature, date |
| Approver | name, signature, date |
8. Related documents
- POL-01 Agent Autonomy Policy · POL-02 Agent Data and Privacy Policy
- STD-01 Agent Build Standard
- Decision log (Logos)
9. Approval
When you adopt this document, record who approved it, their role and the date, and keep that record with your other governance records.