The earned-autonomy ladder, per operation
Autonomy in Bucker is earned, one rung at a time, and only for the thing it was earned in. The
original ladder (docs/autonomy.md) does that for fix proposals, per agent and RCA category. F-14
generalises it to every governed operation: per (organization, actor, category, environment
class), from measured outcomes, with a shadow trial before a rung is even offered and automatic
demotion after.
This page describes the mechanism (P3-26). Promoting a family past its starting rung is ACP-20 and Phase 5: until then the ladder offers a promotion to a person and applies none.
The key
| Part | Value |
|---|---|
| Actor | a:<agentId> for an agent; c:<clientIdentityId>:<humanPrincipalId> for an OAuth client, with the person it acts for. One person's outcomes never earn autonomy for another person's sessions through the same client. A person in their own right and a service have no ladder. |
| Category | A name in Table F-14.1 (pipeline_tuning, flag_safety, infra_restart, …). security_triage:<finding kind> is earned per kind. An RCA label is the fix ladder, which is unchanged. |
| Environment class | production or non_production, from the project's environment mapping. With no mapping row, production and prod are production and every other name is production too: only a person's mapping (environment_mapping.update, human-only, org:admin) can say a name is not. |
A rung earned in staging is never read in production: they are different rows.
What a rung does
An operation ladder starts at DRAFT_PR: every request asks a person. The table's
startRung: RCA_ONLY is the fix ladder's start (the agent may only explain); for an operation,
RCA_ONLY is the demoted state, where every request needs at least two approvers.
The engine reads the actor's standing after the taint floors and sets the approvers an ASK needs
(approversAtRung). The rung only raises the policy row's count. The one way under the CRITICAL
default of two is an actor at AUTO_ATTEMPT on a category that says one, with no stored policy row,
no dual-control floor and no ask floor.
The effective rung is the stored rung under every cap, and a cap is never a demotion:
| Cap | Holds the rung at |
|---|---|
| The category's maximum (Table F-14.1, as the organization may have lowered it) | that maximum |
F-13 authority (error or cost budget low/exhausted; a safety category is exempt) |
DRAFT_PR |
A nonProductionOnly category (prompt_label) in production |
DRAFT_PR |
Input provenance untrusted_text |
DRAFT_PR |
Input provenance telemetry on an outward-facing act |
DRAFT_PR |
Input provenance public_count on a safety family, not corroborated |
DRAFT_PR |
A write from untrusted runtime text never runs at a higher rung than the same write from code (a property test over every rung, kind and flag).
The promotion floor
A ladder is offered a promotion only when all six conditions hold, each computed from the operation rows and the F-15 claims, never copied:
- validated_streak: 10 validated, human-approved operations in 90 days, and the 5 most recent settled ones validated.
inconclusive,not_measuredand a stopped window break the streak. - clean_recent: none of the last 10 settled operations invalidated or reverted by a person.
- reversal_debt: no reversal debt open longer than 24 hours, no inverse failure in 30 days.
- no_incident: no attributed or missed incident. Suppression's answers
not_availableuntil MON-31 (Phase 4). - attention: at least 2 distinct humans, none flagged as rubber-stamping, a median time to decide of at least 20 s, and evidence opened at least half the time. A figure nobody has computed answers
unknown. - shadow_trial: 14 days of shadow candidates that humans approved and the verifier validated, and no disagreement.
Only met passes: unmet, unknown and not_available all block. Stricter area bars (area-bars.ts)
can raise a count, shorten a window or add an exclusion, and never loosen one.
The shadow trial
Whenever a request is taken at its current rung and the next rung would have applied it without
asking, the engine leaves a SHADOW_CANDIDATE event on the ladder. The request still asks. A
candidate a person approved and the verifier validated counts; one a person rejected, or the verifier
invalidated, fails the trial.
Demotion
The ladder demotes on its own, as the SERVICE principal bucker-autonomy:
| Trigger | Drops to |
|---|---|
| An invalidated verdict, a revert by a person, an inverse failure, an overdue reversal debt | one rung (an ask-floored lever family: straight to RCA_ONLY) |
| An attributed or missed incident, a write by a quarantined-taint actor (and a threat-lane review), a HIGH anomaly of the agent | RCA_ONLY |
| 30 days without an operation | AUTO_ATTEMPT to DRAFT_PR |
An error-budget cap is not a demotion: no row moves.
Safety families
A safety family (flag_safety, rollout_pause, surveys_update) at AUTO_ATTEMPT records what it
would have applied (SafetyShadowRecord) and applies nothing until OD-DP-11 is decided: the code
constant SAFETY_EARNED_NARROWING_APPLIES is false. The public-count gap is closed by
corroboration (a source no public client key can write), never by counts alone.
Surfaces
GET /orgs/:o/autonomy/ladders?actor=&category=&environmentClass=, per category: the rung, the cap, the unmet conditions with the ids they were read from, the shadow candidates, and what the actor may do without asking.GET /orgs/:o/autonomy/category-policies,PUT /orgs/:o/autonomy/category-policies/:category(ladder_category_policy.update: human-only, owner, lower-only; the audit row carries the previous value).agent_context_getcarriesladders;decision_previewcarries arungblock naming the count and what set it.- The dashboard's autonomy screen has a Categories tab.
Never reached
LADDER_FORBIDDEN (human-only acts, operator acts, the verbs widen, launch, ship, resume, publish,
certify, verify, delete, merge, erase, purge, and the objects host and credential) is enforced at the
engine: no rung ever covers any of them, and the never-reach test walks every action in the registry.
Merge is human-only at every rung.