M0 second correction: strategy integrity (price elasticity) + forecast direction — do not merge pending re-review

[HSpector1]

Second bounded correction inside M0. Not M0.5, not M1. Closes the two findings the first pass left open. Do not merge — for independent re-review, then owner-run human playtests.

H2 / NEW-1 — strategy integrity (price elasticity)

• New PriceModel: willingness-to-pay anchored to each dish's own suggested price (not the menu median), widened by segment tolerance + dish quality; smooth hyperbolic resistance above WTP, segment-scaled, floored at 3%. No hard cap, no name/recipe branch, no cliff. PriceResistScaleBp=450 from an elasticity target, not the dominance outcome.
• Suggested prices recentered (~+40%) and segment sensitivities retuned for credible anchors → golden checksums deliberately re-baselined (documented, DECISION-LOG D-022).
• No cross-market dominator — proven by a strengthened frontier search, not a comparison to the under-optimized named strategies. The best generalist plan is ~$1170 below the per-market frontier in its worst market; the per-market optima are distinct, opposed regimes (value lunch → fair-priced/high-capacity; enthusiast → premium/lean). Premium pricing stays viable contextually (wins only the enthusiast market, restoring hypothesis H1).

H3 — forecast direction

• Per-station kitchen bottleneck (min over stations, not averaged), realization keyed off kitchen throughput only, and over-acceptance waste for seats past the kitchen wall. The forecast no longer reverses the seat decision (was 46→66: forecast +$483 / actual −$377). Held-out seat-direction agreement 11/12; band coverage 66%/73% vs stated 55%.

Verification

• 117 tests pass (Core 54, Determinism 31, Scenario 32), identical across two runs. Determinism PASS.
• Three independent adversarial verification agents re-ran the claims; the price verifier caught that the original "0 dominators" test was low-power against a weak bar, which drove the frontier-based rewrite.
• Contracts: docs/design/PRICING-CONTRACT.md, docs/design/FORECAST-CONTRACT.md. Full report: reports/m0/M0-CORRECTION-2-REPORT.md.

Disclosed M0.5 residuals (not hidden)

• Named strategy set is illustrative / under-optimized vs the frontier — enrich it.
• Lean strategies underperform heavily — labor/throughput economics deserve a calibration review.

Recommendation: technical Conditional-Pass pending the owner-run human playtests. No M1 work; the Builder does not declare M1 readiness.

🤖 Generated with Claude Code

[HSpector1]

Focused Independent Re-Review — M0 Pricing & Forecast Integrity (PR #3)

A. Review identity

• Repository: github.com/HSpector1/Restaurant
• PR: M0 second correction: strategy integrity (price elasticity) + forecast direction — do not merge pending re-review #3 (fix/m0-price-forecast-integrity → base fix/m0-audit-corrections)
• Exact commit reviewed: 7fd1fbc652fe1660781b6cd514fb229771a28b69 (base 88c1c0b)
• PR Foundation + M0 Headless Service Lab (stops at M0 gate) #1 head beb9289, main 06f0a7e — confirmed unmodified.
• Date: 2026-07-29 · Env: macOS Darwin 25.5.0 arm64, .NET SDK 8.0.423 (fresh clone, read-only)
• Method: 5 independent read-only sub-reviewers (A price, B frontier, C forecast, D regression/scope, E labor), each self-cloning the exact commit with its own held-out seed bases and an independent method. The lead independently reproduced every Blocker/High on fresh seeds (bases 55555557 / 88888883 / 13372099 / 13370000, none used by the builder).

B. Executive verdict

Verdict: Conditional-Pass
Action:  Defer

• PR M0 second correction: strategy integrity (price elasticity) + forecast direction — do not merge pending re-review #3 merge: merge into the correction chain after the two specified High corrections (or an honest re-scope of the claims they falsify).
• PR Foundation + M0 Headless Service Lab (stops at M0 gate) #1 → main: do not merge until the human gate passes (deferred).
• Human-test authorization: NOT authorized yet.
• M1: remains unauthorized.

The engine is genuinely sound and the two chartered fixes are partially real: the specific H2 uniform-overprice exploit is closed, the H3 seat-direction reversal is fixed, determinism/accounting/scope/purity all pass, and price elasticity is principled (single-peaked, smooth, per-dish, substitution/decline work). But two load-bearing claims are not earned and would contaminate the human verdict: (1) the forecast recommends the wrong pricing direction in the value-lunch market, and (2) the flagship "no cross-market dominator / structurally-opposed regimes" claim is refuted by stronger independent search — a single fixed premium plan is broadly near-optimal in every market and wins the value lunch. Both are correctable without an engine rewrite, so Defer (not Rewrite).

Note on my own prior role: an earlier builder-side pass (which I authored) concluded "no dominator, opposed regimes." That conclusion does not survive stronger independent search and is corrected here. The builder's strengthened frontier search was still ~2× too weak.

C. Finding table

ID	Sev	Status	Claim tested	Verdict
B-1	High	Still reproducible	"No cross-market dominator"	A fixed premium generalist is within 9–24% of every market's (premium) optimum
B-2	High	Still reproducible	Frontier search is strong/conservative	Builder frontiers ~2× too low; search under-explores per-dish-priced premium corner
B-3	Medium	Still reproducible	"Opposed regimes; lunch wants value"	Premium wins the value lunch (+68% vs the test's ValueVolume); tests use a 2-item straw-man
B-4	Note	Corrected	Uniform 2.5× overprice no longer dominates	Confirmed non-dominant (−$755/−$561/−$160), still legal, no name-branch
C-10.1	Note	Corrected	Seat 46→66 no longer reversed	Confirmed: forecast −$373, actual −$523 (both down)
C-10.3	High	Still reproducible	Forecast pricing-direction correct	Sign reversal in lunch: forecast +$491 UP vs actual −$133 DOWN at 1.5×
C-10.6	High	Still reproducible	Forecast ranking useful	Lunch top-choice wrong, regret ~$1170; social/enth fine
C-10.5	Medium	Partially corrected	Coverage ≥55% by market/class	Aggregate ~66–68% OK, but lunch 41–55% and capacity-class ~49% breach the floor
A-1	Medium	New	Suggested price is not a hidden cheat	Semi-circular: WTP anchor and strategy prices share SuggestedPriceCents; model measures deviation, can't validate anchors (acceptably data-sensitive, not fragile)
A-2	Low	New	No forced purchase	Confirmed (decline/substitution work); dead Party.BudgetPerCover code, no pay-side ceiling (documented D-012 boundary)
D-*	Pass	Corrected/OK	Tests, determinism, accounting, H1, scope	117 tests ×2, Debug=Release, accounting to the cent, no scope leak, docs otherwise honest
E-1	Medium	New	Labor is a meaningful decision	Max staffing wins 15/18 cells (10–40× marginal return); overstaffing penalty only at seats ≤25; does not block

D. Price-model evidence (Domain A — reproduced)

• Contribution is single-peaked, not unbounded. Complete-menu multiplier sweep (my held-out seeds), median contribution: lunch 1.0×=−$96 / 1.25×=+$803 / 1.5×=−$172 / 2.5×+=−$755 (0 covers); social peaks 1.25×=+$1893; enthusiast peaks 1.25×=+$1926. Argmax 1.25–1.5× in every market across two seed families; 2.5×–5× collapse to the fixed-cost floor. The uniform-overprice exploit is dead.
• Smooth, no cliff: steepest 1%-of-WTP step ~1818 bp (a kink at WTP, ~72 bp per 1-cent step), far from a ~9000 bp discontinuity. No 1-cent/1-bp exploit, no hard cap, no name/recipe branch. WTP anchored per-dish (value=identity, social 1.27–1.44×, enthusiast 1.42–1.69×) and matches the contract table.
• Substitution/decline work: overpriced main ordered 0×/run while a fair peer absorbs demand; all-overpriced menu ⇒ 0 covers. Seated customers are not force-fed.
• Caveat (A-1, Medium): because the strategy menu prices and the WTP anchor both derive from SuggestedPriceCents, resistance is exactly scale-invariant in the suggested-price level; the model measures deviation from authored prices and cannot validate them. Perturbation shows this is acceptably data-sensitive (±10%/2× preserve the verdict; halving makes the game unprofitable because ingredient cost is fixed; mispricing one anchor below cost correctly dethrones its strategy). This is a real limit on what the green tests prove — "do the suggested prices feel right?" is a genuine playtest question the sim cannot self-answer.

E. Strategy-frontier evidence (Domain B — reproduced, contradicts the central claim)

Independent multi-restart hill-climbing (a different algorithm than the builder's random+archetype sweep) found a single fixed, legal plan — genB = Ribeye@$84.96, Burger@$21, Scallops@$54, Fondant@$41, 28 seats, full acceptance, 6 staff — with these medians on my fresh held-out seeds (400 sims/cell):

Plan	lunch	social	enthusiast
genB (fixed premium generalist)	$4019	$8203	$7824
ValueVolume (the DominanceFrontierTests "value" plan)	$2387	$3564	$2626
Builder's reported frontier (dominance-search.md)	$2350	$5005	$4657

• genB beats the builder's entire reported frontier in all three markets, and beats the test's ValueVolume plan including in the value lunch (+68%) — directly refuting "lunch wants a value regime / structurally-opposed regimes."
• genB regret vs my strong (seeded) per-market optimizer: lunch 11%, social 9%, enthusiast 24% (frontiers $4500 / $9060 / $10310 — all premium plans). So a single premium generalist is within ~10% of two markets and ~24% of the third; the per-market optima are the same premium regime re-tuned, not opposed regimes.
• B-4 (corrected): the former universal-overprice dominator (exactly 2.5× suggested, 44 seats) is now decisively non-dominant (−$755/−$561/−$160, ranks last/near-last), remains a legal ungated config, determinism holds. The specific H2 exploit is genuinely closed.
• Root cause of the false claim: the harness search uses blind random draws + an archetype sweep that holds staffing fixed and uses a single global price multiplier (never per-dish), so it never finds the per-dish-priced premium optimum; and DominanceFrontierTests compares a fair plan against a crippled 2-item "PremiumLean" straw-man.

F. Forecast evidence (Domain C — reproduced)

• Seat direction (H3) — FIXED (C-10.1): lunch 46→66 forecast −$373 vs actual −$523 (both down; old bug was +$483 vs −$377). Forecast and actual both single-peak near ~30 seats. Verified byte-identical twice.
• Pricing direction — REVERSED in the value market (C-10.3, High): lunch 1.0→1.5× forecast +$491 (UP) vs actual −$133 (DOWN). Social/enthusiast agree (premium helps there). Root cause: DemandModel.ConversionBp samples only the single top-appeal main's resistance while forecast revenue scales with all menu prices, so it under-penalizes moderate overpricing in the price-sensitive market. This violates the FORECAST-CONTRACT's own directional guarantee V0 — 3D Restaurant Readability Sandbox (visual track — do not merge, Art PM review) #4 and is a sibling of the very reversal H3 just fixed.
• Ranking (C-10.6, High): lunch top-choice wrong (forecast floats premium/high-price plans above the actual best), regret ~$1170; social/enthusiast rankings are good (Spearman 0.85/0.97).
• Calibration (C-10.5, Medium): aggregate coverage ~66–68% (I reproduced 67.6%) vs stated 55% — conservative in aggregate — but lunch (41–55%) and the capacity-constrained class (~49%) breach the 55% floor, driven by a systematic −$330 pessimistic bias (22/27 cells under-predict). The player-facing "55%" label matches the constant, but the "≥55% in every market/class" claim is false. Bands are also very wide (1.5–6.8× the true outcome spread).

G. Labor evidence (Domain E — non-blocking)

Understaffing is severely, legibly costly (a starved station ⇒ walkouts, 90–100% loss). A genuine lean strategy is viable with menu/station discipline (2 cooks + FOH on a 2-station menu: +$1186…+$1462, 0% loss). But on a fixed menu/seats, adding cooks to the maximum wins 15/18 market×seat cells (a bottleneck cook unlocks ~$1000–$2800 of contribution for a $60–$180 wage; 10–40× marginal return); max staffing is the profit-max cook count in all three markets at seats ≥30. The overstaffing penalty is real but only appears at seats ≤25. Gate ruling: does not block — understaffing is costly, overstaffing eventually loses money, staffing is not removable as a decision — but staffing level does not vary by market, which reinforces the Domain-B finding that context matters less than claimed. Recommend the playtest script include a small-capacity (~20–25 seat) scenario.

H. Regression & scope (Domain D — PASS)

Independently reproduced: 117 tests (Core 54, Determinism 31, Scenario 32), 0 fail/skip/warn, ×2 identical; Debug vs Release checksums byte-identical; accounting reconciles to the cent (contribution = revenue − ingredients − labor − overhead, verified incl. an 89-failure service); H1 attribution preserved (no 88%-util flip); integer-cents parsing, FIFO, labor, float/wall-clock/RNG guards all intact; scope clean (no M1/persistent-world code, no committed binaries, bin/obj/assets gitignored). Goldens were re-baselined for a genuine intentional change (D-020/D-022), not to suppress a behavior change. Docs are otherwise honest (cross-OS determinism correctly scoped as OPEN). One low nit: ADR-003 phrases portable determinism a touch more strongly than ADR-002.

I. Human-test ruling

NOT authorized. Two defects would contaminate the human M0 verdict: the player-facing forecast recommends the wrong pricing direction in a core market (teaching false cause-and-effect), and the balance claim the gate would be entered under ("no dominator / read-the-market-and-flip-regimes") is false (a premium generalist is broadly near-optimal and wins the value lunch). Neither is a determinism/accounting/scope defect — the technical foundation is sound — but both must be corrected or honestly re-scoped, then re-reviewed, before uncoached testers are exposed to the loop.

J. Final gate

Verdict: Conditional-Pass
Action:  Defer

Required corrections before the human gate (re-review after):

1. Forecast pricing direction (C-10.3/C-10.6): make the forecast's demand/revenue response to price consistent with the sim (apply per-dish ResistBp/coherence across the ordered course mix, not just the top main in ConversionBp), or narrow the FORECAST-CONTRACT to disclose that the forecast under-penalizes 1.25–1.5× overpricing in value markets. Re-verify lunch top-choice/regret.
2. Strategy-integrity claim (B-1/B-2/B-3): either strengthen the economics so a fixed premium plan cannot stay within ~10% of the value-lunch optimum (e.g. bind premium contribution in lunch via the segment's budget/price-sensitivity or a peak-throughput ceiling), or retract the "no dominator / structurally-opposed regimes" claim and replace the frontier tests with an optimizer-derived frontier and a percentage near-optimal band, honestly restating the property as "a coherent premium generalist is broadly strong; context tunes rather than reverses the best plan."
3. Calibration wording (C-10.5): stop presenting per-market/per-class "≥55%" as established; state 55% as aggregate-only and disclose the lunch/capacity breach and the ~−$330 bias.

K. Required owner decision

Howard & Aaron must choose the path before the human gate: (a) direct the builder to fix — align the forecast's price response with the sim, and change the economics so premium does not broadly dominate the value lunch — then re-review; or (b) accept an honestly re-scoped balance claim (premium generalist is broadly strong; context tunes not reverses) plus a disclosed forecast pricing-caveat, and re-review a corrected-claims PR before the gate. Either way the two Highs must be resolved and re-verified before the five uncoached M0 playtests begin. Do not merge PR #1 into main until that gate passes; M1 stays unauthorized.

— Independent Reviewer (read-only; no branches modified, no PR merged, no human tests run, M1 not begun)