docs product pilot plan - AlexandreZanata/Abastevo GitHub Wiki
Controlled community pilot — launch definition (P10-T09)
State: DEFINED, NOT LAUNCHED, 2026-10-05. Owner task P10-T09 (issue #53). This document defines the pilot; it launches nothing. Execution stays BLOCKED until the entry prerequisites below are met. No city, cohort, window or result is claimed here.
Entry prerequisites (all unmet — none waived)
- Scoped real G09 RELEASE_CERTIFIED on an immutable candidate
(record:
../release-evidence/p09-g09-scoped-release.md; currently UNCERTIFIED/BLOCKED, tracker #13 OPEN). - G24 manual/device batch closed (accepted BY USER DECISION 2026-10-02;
batch OWED — record
../release-evidence/p24-g24-user-acceptance.md). - Separately authorized production deployment + tag; no rollout follows any plan merge automatically.
Launch scope (fixed at authorization, TBD below — never invented)
- One bounded city: municipality + state, directory station list and
survey-week cadence recorded before opening (criteria in
../operator/pilot-city.md). City: TBD. No simulated members, prices or activity, ever. - Cohort/capacity/window: agreed with the owner at authorization from measured staging capacity. Cohort: TBD. Window: TBD.
- Staffing: named moderation coverage with response target + triage
cadence, staffed support route, reachable privacy operations
(export/erase, 24-hour evidence expiry, 12-month audit purge).
Names/targets: TBD (checklist in
COMMUNITY_METRICS.md).
Conduct and rollback (procedure, rehearsed at launch — not executed)
- Operate per
../operator/pilot-city.md: opening checklist first, stop/pause triggers owned there (backlog, abuse, privacy incident). - Rollback rehearsal at launch: pause onboarding/contribution, keep ANP browsing available; product changes revert only through versioned releases; moderation audit rows stay append-only.
- Monitor freshness/disputes/cost against the frozen definitions in
COMMUNITY_METRICS.md; sparse cells show the dated ANP reference.
Results
- Pilot report goes to
pilot-results.mdonly after a real run, with evidence, limits and next product decisions. No run exists: no results, no metrics and no launch decision are recorded here.