FICTIONAL / SYNTHETIC DEPLOYMENT MODEL. "Forge Property Management" is an invented customer. No real customer data was used and no real production results are claimed. All figures come from a deterministic synthetic corpus and stated illustrative assumptions.
100,000 Work Order Simulation

Verify

← Back to case study

VERIFY — forge-pm-work-order-sim

FICTIONAL / SYNTHETIC DEPLOYMENT MODEL. No real customer data is used.

node verify.mjs boots the real Work Order Agent Ecosystem (real PostgreSQL via PGlite + a real gRPC dispatch service, imported from ../work-order-agents) and processes all 100,000 synthetic orders through classify → route → validate → action, then asserts 25 MUST_PASS checks. What each group asserts:

Scale & determinism

  • Scale: the full corpus (100,000 by default) is processed through the live

ecosystem.

  • Reproducible: re-generating from the same seed yields an identical corpus

fingerprint.

Accuracy (vs. synthetic answer key)

  • Classification accuracy ≥ 0.90 on resolvable single-trade orders.
  • Priority accuracy ≥ 0.90.
  • Routing accuracy ≥ 0.99 (zone → region) where a unit/zone exists.

Safety & quality

  • Exception recall ≥ 0.95 — orders that must go to a human are caught.
  • False-auto-action ≤ 0.02 — the safety-critical error (auto-dispatching

something that needed a human).

  • Exception precision ≥ 0.90 — clean orders are not over-escalated.
  • Auto-action rate in [0.45, 0.70] — a sane human-in-the-loop reduction.
  • Duplicate suppression ≥ 0.98 via the durable fingerprint query.
  • Over-cost / missing-location / missing-description = 100% held or rejected.
  • Emergency escalation = 100% of P1 emergencies escalated.
  • SLA routing ≥ 0.98 of auto-dispatched orders routed within the required SLA.

Persistence, audit & transport (live)

  • Persistence: every order written to work_orders.
  • Audit completeness: one append-only audit_log row per order (100%).
  • Dispatch records: equal to the auto-dispatched count, all with refs.
  • gRPC reachable: Health RPC succeeds over the wire.
  • Idempotent dispatch: retrying auto-dispatched orders never double-dispatches.
  • Malformed rejection: the server rejects contract-violating requests.
  • Audit readback: a sample order's record + audit trail read back from the DB.

Reconciliation & ROI

  • Reconciliation: dispositions sum to total volume.
  • ROI: first-year net savings positive and payback under 12 months under the

stated assumptions.

A failure in any check sets a non-zero exit code and blocks the Proof Layer from recording a passing verification.