Auditor Objections — Starship Build Engineering System

Pre-written hostile objections, each answered from evidence.

Objection 1: "These are made-up numbers dressed up as SpaceX data."

Response. Correct that the data is notional, and it is disclosed everywhere (README.md, proof/LIMITATIONS.md, verification-report.json#/disclosedSeams, and an on-screen banner in the planner). The claim is not about real data; it is that the *engineering math* is correct. That math is verified on fixtures with hand-checkable answers (checks 1–17), independent of the model.

Objection 2: "Line balancing is a heuristic — you can't prove it's right."

Response. We do not claim global optimality. We assert provable invariants: no precedence violation, every station ≤ takt, station count between the theoretical minimum (ceil(work/takt)) and the operation count, and a deterministic result. Check 3 also proves the balancer hits the known optimum on a case where the optimum is obvious (4×30 min @ takt 60 → 2 stations, 100%).

Objection 3: "The capability PPM assumes normality you never tested."

Response. Disclosed as an assumption in proof/LIMITATIONS.md. The Cp/Cpk *indices* are exact algebra (check 11). Only the PPM/yield conversion uses the normal model, validated against textbook anchors (Φ(1.96) ≈ 0.975, ±3σ ≈ 0.9973, check 12–13).

Objection 4: "Tolerance stack-up is trivial / could be wrong."

Response. It is exact (check 14: WC 0.5, RSS 0.3 for [0.1,0.2,0.2]) and guarded by an invariant that RSS ≤ worst-case for every model stack (check 15).

Objection 5: "How do I know verify.mjs actually ran and wasn't faked?"

Response. The Proof Layer re-ran it; raw stdout is in proof/evidence/verify.log with a sha256 in proof/EXECUTION_TRACE.json, and every shipped file is checksummed in proof/CHECKSUMS.json (node tools/forge-proof-verify.mjs re-validates).

Objection 6: "The benchmark is rigged against an artificially bad baseline."

Response. The baseline is the most common real anti-pattern — one operation per workstation with no consolidation — scored at the *same* takt as the optimized line. Check 18 only asserts the optimized line is never worse; the magnitude (31.25% / +45.5%) is read directly from the report, not asserted.

Objection 7: "Marginal Megabay weld at Cpk 1.21 — is that a bug?"

Response. No — it is intentional. The closeout-weld sample is set to fall just below the Cpk ≥ 1.33 bar so the acceptance logic visibly flags a not-capable weld, demonstrating the quality gate works rather than always passing.