The 90-Day Growth Control Pilot — Playbook

Status: ADOPTED as the standard commercial onboarding (owner ruling 2026-08-19-A). Source: docs/MIZOKI_SIGNAL_GROWTH_CONTROL_UNIFIED_SYSTEM_r2.0.md §4.2, reproduced verbatim below. Machinery: contracts/mizoki_contracts/pilot.py (per-tenant state machine; gate criteria enforced in code — an advance with the day window unreached or the checklist incomplete is refused with the missing items named) + the internal /api/v1/pilot/* routes on service-audit-replay. The pilot-report generator is a skeleton until a real pilot writes verified numbers — Preview labels flip only on those.


r2.0 §4.2 — verbatim

4.2 The commercial onboarding — the 90-Day Growth Control Pilot (standard, by owner ruling). - Days 1–30 · Observe: connect the agreed stack (Google Ads, Meta, Shopify, GA4, BigQuery), establish baseline data quality, define the target Decision Jobs. Zero changes to live execution. - Days 31–60 · Validate: synthetic-control geo experiments, CATE model calibration, propensity verification, margin reconciliation against the customer's own books. - Days 61–90 · Recommend: fully contextualized decision proposals with complete audit trails, routed through designated human approval, every prediction graded against outcomes.

The pilot rides the Three-Gate Evidence Maturity Framework — approved through Gate 3: Gate 1, demonstrable logic (the live public Signal Factory demo is Gate 1); Gate 2, pilot validation on live systems with calibrated models and verified margin economics; Gate 3, expansion — Level 3/4 execution authority within caps, granted when accuracy, policy, and safety requirements are met on real data. And one alignment that makes the machine self-reinforcing: the pilot that closes a customer is the same machine that produces the verified numbers that flip the public Preview labels. The sales motion and the proof motion are one motion.


Operating the machinery

Step Command / route Gate
Start a pilot POST /api/v1/pilot/{tenant}/start {"target_job_ids": ["J-01", …]} job ids validated against the J-01…J-06 registry
Record evidence POST /api/v1/pilot/{tenant}/checklist {"phase","item","evidence"} evidence REQUIRED — "not measured" is an acceptable honest value, silence is not
Advance a phase POST /api/v1/pilot/{tenant}/advance refused (409, items named) unless the day window is reached AND the phase checklist is complete
Read state GET /api/v1/pilot/{tenant} —
Generate the report skeleton GET /api/v1/pilot/{tenant}/report verified_numbers stays [] until a real pilot writes them
Generate the spec'd report (Wave 2 WS-8) GET /api/v1/pilot/{tenant}/report/{week-n\|day-90} (?format=markdown) — contracts/mizoki_contracts/pilot_report.py per docs/product/PILOT_REPORT_SPEC_v1.md assembled from stores only, "no data" rendered literally, ILLUSTRATIVE watermark by evidence class; Day-90 returns the ledger FEEDER (proposed row + readout markdown whose suggested path is docs/pilot/PILOT_READOUT_<tenant>_DAY90_<date>.md, per content_qa rule E(f)) — the ledger is never written

Checklists enforced in code (mizoki_contracts.pilot.PHASES):

Gate artifacts are recorded per phase transition (gate1 standing — the public Signal Factory demo; gate2 on entering Recommend; gate3 on completion). The Gate-3 artifact records evaluation inputs and grants nothing — L3/L4 authority within caps is a human grant on real-data accuracy/policy/safety criteria, per the Three-Gate framework and the autonomy-promotion constants (Brier ≤ 0.20, AUC ≥ 0.72, ≥ 2 purchase cycles, human decision).

Owner inputs on the critical path (report §15 of the completion run): pilot tenant selection · real config/net_yield_costs.yaml costs · config/treasury_constraints.yaml floors/covenants · config/f4_geo_candidates.yaml geos + caps · creative-asset access for F1 volume.

← All docsView source on GitHub →