An agent may observe, propose, act with approval or act within limits. The level is set per action type, per agent, and an agent climbs only through a gate: a passing evaluation on golden cases, an escalation route that is wired, and a human stamp.
Promotion is a change like any other: it is drafted, checked and stamped, and it is in the audit trail.
The gate is the same discipline every publish already has: a change set, a check, a stamp. It applies per action type, so a corrector agent can be at L3 for country-code standardisation, where the model is at 99.9% and the field is public, and at L1 for bank details, where a person always decides.
Demotion is one action with no gate: the owner lowers the level, the agent's pending actions return to the queue, and the reason is recorded. An agent whose acceptance rate drops below the owner's threshold is demoted by policy, and the owner is told.
Because the evaluation is replayed after every change to the directives, an edited rule earns its level again. This is what keeps a promoted agent honest as the rules around it change.
| Guard | What it checks | When it fires |
|---|---|---|
| Confidence floor | The proposal's confidence against the floor set for the action type | Below the floor the action goes to a person, whatever the level |
| Daily budget | The number of writes the agent may make in a day, per action type | At the budget the agent stops and reports; the owner raises it or reviews |
| Four eyes | Fields marked sensitive (bank details, tax ids, credit limits) need two people | Always a person, and a second one, whatever the level |
| Scope | The action stays inside the agent's tables, rows and fields | Outside the scope the action is refused and logged |
| Write-back acknowledgement | A write is complete only when the owning system confirms it | Until the acknowledgement arrives the change stays pending and retries |
Policy floors are the limits no configuration can lower: a blocking rule is waived only with a written reason of at least fifteen characters, provenance is always shown, and write-back is off per field until a test write has passed. Floors are set by the platform, and the register shows them as locked.
This is what an auditor reads.
| when | who | field | before | after | why | entry |
|---|---|---|---|---|---|---|
| 02:14:08 | corrector L3 | supplier 100482 · country | United Kingdom | GB | ISO 3166 · conf 0.99 · budget 212/500 · guards pass | #88,102 |
| 02:14:09 | corrector L1 | supplier 100913 · bank account | GB29 NWBK ···· 1234 | proposed: GB29 NWBK ···· 1298 | verifier mismatch · four eyes → treasury screen | #88,103 |
| 02:16:40 | duplicate finder L2 | supplier 100482 ↔ 55901 | two records | merged · approved by j.park | tax id equal · address 0.94 · steward stamp | #88,140 |
| 06:31:02 | verifier L2 | supplier 100777 · bank account | verified 2024-06-02 | verified 2026-09-18 · ref BV-77120 | aged 27 months · timeliness rule | #88,911 |
Four entries from the showcase ledger: the agent and its level, the field, before and after, the guard verdicts and the entry id. Showcase estate.
The steward, from decisions already made in the queue: an accepted proposal becomes a positive case, a rejected one a negative case. Thirty to fifty cases per action type is the usual size; the eval page shows the coverage.
No level bypasses four eyes on a field marked sensitive; the mark is a policy floor. The agent prepares the change and the evidence, and two people decide.
Per agent and per action type, by the owner, from the dry-run counts and the acceptance rate. A budget is a brake, so it starts low and is raised on evidence.
For standardisations and list mappings, yes, once the model is trained on the estate's own decisions. For merges and verified corrections, L2 is the usual ceiling; for sensitive fields, L1.
Bring one domain and two systems. In a week the sources are profiled, the ten costliest fields are under rules, the first corrections are written back with their acknowledgements and the publish gate is on.