Auditor Objections — forge-pm-work-order-sim

FICTIONAL / SYNTHETIC DEPLOYMENT MODEL. Hostile objections, answered from evidence.

Objection 1: "This is a fake customer dressed up as a real deployment." Answer: It is explicitly labeled a FICTIONAL / SYNTHETIC DEPLOYMENT MODEL in the README, every report, the dataset card, the social copy, and the public page. The customer-outcome seam is declared to the Proof Layer, which therefore withholds CERTIFIED/PRODUCTION_VALIDATED. No real deployment is claimed anywhere.

Objection 2: "The accuracy numbers are inflated by easy synthetic text." Answer: Agreed that synthetic text is cleaner than real intake — this is disclosed in LIMITATIONS.md and the dataset card. The corpus deliberately injects hard cases (ambiguous two-trade orders, duplicates, over-cost, missing-location, missing-field) and the metrics are measured against a ground-truth key, not asserted. Real-text accuracy is explicitly listed as unverified.

Objection 3: "The ROI is just made-up savings." Answer: The ROI is a transparent model whose every input is a stated assumption in assumptions.json and whose only measured input is the auto-action rate from the simulation. The arithmetic is shown line-by-line in evidence/roi.json and the impact study. It is presented as an illustrative model, not a realized result.

Objection 4: "You didn't really run 100,000 — you extrapolated from a small run." Answer: verify.mjs processes the full corpus through the live ecosystem (default 100,000), and a MUST_PASS check asserts the processed count equals the target. Real DB row counts (work_orders, audit_log, dispatch_records) and the execution log in proof/evidence/verify.log corroborate the full-scale run.

Objection 5: "The infrastructure is fake." Answer: Persistence is a real PostgreSQL engine (PGlite) with real SQL/transactions, and dispatch crosses a real gRPC/HTTP2 wire — verified by reachability, idempotency, and malformed-rejection checks. What is a seam is that it runs in-process on localhost rather than against an external Postgres/Go service; this is disclosed.

Objection 6: "The classifier is not the LLM you'd ship." Answer: Correct and disclosed. The classifier is a deterministic lexicon engine chosen for auditability; the LLM swap is a documented, unverified seam.

Objection 7: "Numbers can't be independently reproduced." Answer: Everything is seeded (20260625); REPRODUCE.md lists exact commands; a corpus-fingerprint check guarantees identical inputs; CHECKSUMS.json pins every shipped file.