Documentation

Autopilot & control tiers

Deterministic validation, action leases and human gates.

Autopilot is a loop, not a model with credentials. Deterministic detectors watch provider-level deferrals, bounces and complaints. An incident opens. The model (or a rule fallback) proposes exactly one structured action. A policy controller validates it, and — if allowed — applies it as a leasethat expires on its own. The model never touches a database, shell, cloud API or MTA configuration.

loop
detectors → incident → proposal → validation → apply | hold for approval
                                   ↓
                              signed lease (5–240 min)
                                   ↓
                          expiry / rollback → audit

Tiers describe authority, not confidence. A high-confidence A3 proposal still needs a human; a low-confidence A1 proposal is rejected by the confidence floor.

TierAuthorityActionsGate
A0Observe & proposeSummaries, root-cause notes, structured proposalsAlways permitted, even with autopilot off
A1Reduce riskReduce scheduler concurrency, pause stream, pause provider within a tenantAutomatic after validation; short lease
A2Bounded recoveryRestart stateless worker, scale within reserved capacityAutomatic only at A2; otherwise approval. Lease rollback.
A3Expand privilegesPause tenant, raise caps, lift suspension, resume tenant/provider; reassign IP/change DNS catalogued onlyOwner/admin approval, separation of duties

What the controller checks

  • • Action exists in the catalog and matches the configured autonomy ceiling.
  • • Policy version is the active one; proposed version must not be stale.
  • • Confidence is above the floor (default 0.6) and evidence refs are present.
  • • Telemetry is fresh — older than 15 minutes freezes mutations.
  • • Numeric bounds: concurrency reduction ≤ 90%, lease 5–240 minutes.
  • • Scope is non-empty and the blast radius is allowed (no global provider pause at A1).
  • • Cooldown: no identical action already active.
  • • Lease expiry and rollback condition are recorded; every decision is audited.

What runs today

A0 and A1 are live: scheduler batch limits, stream pauses and tenant-scoped provider pauses. Proposals are model-generated when a DeepSeek key is configured, otherwise the deterministic rule fallback is used. A2 restart/scale executors are wired but this deployment caps automatic application at A1; A3 actions require approval and execute for cap changes, suspensions and resumes. Tenant-wide pause is A3. `REASSIGN_IP`, `CHANGE_DNS` and `QUARANTINE_POOL` remain catalogued but are rejected because no executor is wired.

Changing the ceiling

TRANSMITLY_AUTOPILOT_TIER sets the maximum tier that may be applied automatically. A3 always goes to the approval queue regardless of this value. Setting TRANSMITLY_AUTOPILOT_ENABLED=false stops all mutation while keeping A0 analysis and proposals.

Invariants

  • • Nothing can override suppression, a suspended domain, tenant quota or provider rate limits.
  • • Pool isolation is never used to launder rejected traffic onto healthy IPs.
  • • Missing or stale telemetry freezes increases and applies conservative limits.
  • • The model has no shell, SQL, cloud credentials or direct MTA write access.

Need the full API surface? OpenAPI 3.1.