Shadow-mode proof framework · §16 / §21
Proof over autonomy. Authority is earned, not assumed.
The first valuable build is a reliable shadow operating system that can reconstruct calls, Vans, routes, bookings, warranty/membership context, outcomes, and overrides with defensible receipts. This site documents that ladder. It does not claim any rung is live.
Deployment ladder
- Observe
Ingest and normalize data without influencing decisions. Analytics mode name: Watcher.
- Shadow
Produce sealed counterfactual recommendations while humans and existing software operate normally.
- Recommend
Surface decisions to staff with no autonomous execution.
- Human Approves
One-click acceptance / override workflow.
- Limited Autonomy
Narrowly scoped actions with hard constraints and rollback.
- Autonomous Operations
Only after sustained evidence, governance, and operational approval. Not claimed here.
Proof metrics — not revenue alone
- Gross contribution delta and margin delta
- First-trip completion
- Callback / rework rate and diagnostic accuracy
- Turnovers created and sold
- Travel / fuel / time efficiency
- Customer wait and displacement impact
- Runtime forecast error and inventory prediction accuracy
- Warranty / callback classification accuracy
- Quality / workmanship indicators
- Manager intervention prediction quality
- Human override win / loss / neutral rate
- Exploration results and newly discovered technician capability
Minimum viable shadow system
BYO: the user points this local runtime at their own ServiceTitan and ProBooks read sources. FragGate admits + hashes on that machine. No central dump. Normalize customers, equipment, calls, Vans, bookings, jobs, quotes, parts, turnovers, and sales. Create canonical timestamps and call IDs. Map identities to Van identifiers. Implement Chains A/B/C, baseline Call-Fit and rebasing, role tags and multi-window dashboards, 30-day callback/warranty support, warranty lookup adapters where authorized, media ingest with initial visual QA, reconstructable receipts, and actual-versus-shadow comparison without future information leaking into the original recommendation.