Data quality Autonomy

Autonomy is earned, one level at a time.

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.

ObserveL0reads and reports; noproposals, no actionsProposeL1drafts the fix with evidence; everyaction waits for a personAct with approvalL2applies the fix the momenta person approves itAct within limitsL3acts alone while every guardholds; anything outside escalatesAUTONOMY IS EARNED, ONE RUNG AT A TIMEPromotion gate to L2 and L3eval ≥ 95% on golden cases · escalation route wired · a human stampGuardrails on every action, at every levelconfidence floor · daily budget · four eyes on sensitive fields · policy floorsa level is set per action type, per agent · it can be taken back at any time · the record of every action is what earns the next rung
The four levels and the gate between them. Guardrails apply at every level; policy floors are the limits nothing can lower.
4
levels
product
95%
pass rate on golden cases at the gate
product
3
conditions at the gate: eval, escalation route, human stamp
product
1
action type per level setting, so a corrector can be L3 for country codes and L1 for bank details
product
The four levels

What an agent may do at each

L0
ObserveReads and reports. It scores, it finds, it explains; it proposes nothing and changes nothing. The profiler and the explainer live here.
default for new agents
L1
ProposeDrafts the fix with its evidence and confidence. Every action waits for a person on the escalation screen. The default for anything that changes data.
most rule agents start here
L2
Act with approvalApplies the fix the moment a person approves it, including the write-back and its acknowledgement. The person decides; the agent does the work.
after the gate
L3
Act within limitsActs alone while every guard holds: confidence above the floor, inside the daily budget, outside the four-eyes fields. Anything outside escalates to a person.
after the gate, per action type
The promotion gate

Three conditions, all visible on the agent's page

Promotion is a change like any other: it is drafted, checked and stamped, and it is in the audit trail.

A passing evaluationGolden test cases, curated by the steward from real decisions, are replayed against the agent. The pass rate must reach 95%. The cases stay with the agent and are re-run after every change to its directives.

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.

Guardrails and policy floors

Hard checks on every action, and limits nothing can lower

GuardWhat it checksWhen it fires
Confidence floorThe proposal's confidence against the floor set for the action typeBelow the floor the action goes to a person, whatever the level
Daily budgetThe number of writes the agent may make in a day, per action typeAt the budget the agent stops and reports; the owner raises it or reviews
Four eyesFields marked sensitive (bank details, tax ids, credit limits) need two peopleAlways a person, and a second one, whatever the level
ScopeThe action stays inside the agent's tables, rows and fieldsOutside the scope the action is refused and logged
Write-back acknowledgementA write is complete only when the owning system confirms itUntil 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.

What the ledger holds

Every action with its verdicts

This is what an auditor reads.

whenwhofieldbeforeafterwhyentry
02:14:08corrector L3supplier 100482 · countryUnited KingdomGBISO 3166 · conf 0.99 · budget 212/500 · guards pass#88,102
02:14:09corrector L1supplier 100913 · bank accountGB29 NWBK ···· 1234proposed: GB29 NWBK ···· 1298verifier mismatch · four eyes → treasury screen#88,103
02:16:40duplicate finder L2supplier 100482 ↔ 55901two recordsmerged · approved by j.parktax id equal · address 0.94 · steward stamp#88,140
06:31:02verifier L2supplier 100777 · bank accountverified 2024-06-02verified 2026-09-18 · ref BV-77120aged 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.

Terms on this page

Autonomy levelWhat an agent may do for one action type: observe, propose, act with approval, act within limits.
Promotion gateA passing evaluation, a wired escalation route and a human stamp, required to reach L2 or L3.
GuardrailA hard check on every proposed action: confidence floor, budget, four eyes, scope, acknowledgement.
Policy floorA limit nothing in configuration can lower.
Escalation routeThe screen and the people a finding goes to for a decision, with a hop SLA.
Golden casesReal decisions curated as test cases and replayed against an agent.
Questions

Who curates the golden cases?

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.

Can an agent ever bypass a person on sensitive fields?

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.

How is a budget set?

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.

Is L3 common?

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.

See quality run on your own domain

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.