Skip to content

ADR-0204: A time-limited owner-reviewed email junk pilot

  • Status: Accepted — owner explicitly approved implementation and production activation on 2026-10-14
  • Date: 2026-10-14
  • Amends: ADR-0203 for this bounded pilot only

Decision

The owner reviewed 43 classifications from 50 live work-Inbox conversations: 30 correct junk nominations, six missed junk conversations and seven correct keep choices. Seven conversations were protected before classification. This is a small, selected sample, not the independent 100-case general-rollout gate.

Admit a separate, default-off pilot only with a private owner-reviewed packet bound to the exact questions, pinned Jev model, owner and work account. It requires at least 43 unique owner-reviewed classifications, at least 20 junk nominations, at least seven legitimate keep cases and no false-positive junk nominations. A packet is valid for at most seven days. The pilot may bypass normal assessment for at most 50 new conversations per UTC day; all other mail retains normal assessment. Expiry or invalid evidence stops new pilot calls and restores normal assessment of bypassed mail through bounded refresh.

Use the same closed question and model the owner reviewed. A valid spam or bulk choice qualifies without ADR-0203's 0.99 score cutoff or broad topic-probability veto. Preserve every pre-egress protection, including known correspondents, relationship memory, direct Important/Needs reply feedback, Not spam, kept subscriptions, incomplete sources, attachments, secret and injection detection. Keep and uncertain provider outcomes always retain normal assessment. The six missed junk cases are deliberately left on the normal path in this pilot.

Mail remains readable in Other/All with Important restoring priority. No Gmail labels change, and no email is archived, deleted, unsubscribed, sent, or used to create a draft or memory by the bypass. Existing cost reservations, recovery, audit events and budgets remain intact. Turning the mode off is the kill switch.

The 100-case calibrated filter and its thresholds remain unchanged. This pilot does not establish general rollout readiness or complete a milestone. A wider rollout or renewal requires further owner-reviewed evidence and authorization. Release still requires focused regressions, sidecar checks, exact-head hosted CI and CodeRabbit, successful deployment, and live release-identity verification.

Mechanism

pilot is a separate AGENT_EMAIL_JUNK_FILTER_MODE value. The private packet uses JUNK_PILOT_POLICY (binding ADR-0203's unchanged questions/model plus this pilot's choice rule, lifetime and cap), an owner/tenant identity, the literal work account, aware issue/expiry timestamps, selected-sample provenance and independent labels. It is loaded on every email slice. Missing or invalid pilot evidence disables calls rather than falling back to shadow. Scope and expiry are checked before dispatch, on receipt use and at assessment commit.

The existing principal lock and email transaction atomically save a content-free junk_pilot_day counter, a junk_pilot_admission keyed by a hash of account and provider conversation identity, and the bypass assessment. Repeated refresh or recovery of an admitted conversation consumes no further slot; changing evidence cannot reset the counter. New UTC days grant at most fifty new conversations. The count is checked before a paid request and again at save; losing a concurrent last-slot race resumes ordinary assessment. These records contain no mail content and require no schema migration. The general filter's calibration is unchanged.