M0 correction 3: economic coherence — High A (forecast) FIXED, High B (dominator) NOT resolved (Fail/Rewrite) — do not merge

[HSpector1]

Third bounded M0 correction (PR #4, on top of PR #3). Do not merge. Honest outcome up front.

Result

• High A (forecast pricing direction) — FIXED. New per-cover affordability primitive + a forecast order model that reuses the sim's own primitives (Wtp/ResistBp/CoherenceWeightBp/AffordBp). Audited value-lunch 1.0×→1.5× now agrees (forecast down, sim down); seats 46→66 still down; ≥85% price-direction matrix. Checksum-free.
• High B (cross-market dominator) — NOT RESOLVED. The dead per-cover budget primitive is now wired in (PriceModel.AffordBp, smooth, name-agnostic, no cap/cliff), which defeats premium-only dominators and makes value diners genuinely budget-bound. But a strengthened search (multi-start hill-climb, per-dish prices, review-champion seeded) surfaced a residual mixed "value-chassis + premium-anchor" dominator: a cheap Burger feeds the value majority while premium Ribeye/Scallops skim the high-budget minority present in every market. On held-out seeds it is within ~1–2% of the best per-regime plan in all three markets (lunch ~$2178 where pure-value gets ~$735; social ~$4879; enthusiast ~$4793).

Why High B can't be closed in M0

The premium skim of the always-present high-budget minority is additive profit in every market. Affordability kills premium-only menus but not a mixed menu whose cheap chassis serves value diners. Market composition (tested to 80% value) doesn't close it; lowering premium margins/prices destroys premium play; a value-market penalty is a forbidden hack. This is the single-service / fixed menu-independent arrival mix / no-repeat-visit boundary — an M1 redesign (menu/positioning-dependent arrival mix and/or repeat-visit reputation).

Gate recommendation (per §18)

Verdict: Fail
Action:  Rewrite

Keep the affordability + forecast corrections (correct; resolve High A). But strategy integrity cannot be honestly established in M0 scope. Full detail + required reviewer recheck: reports/m0/M0-CORRECTION-3-REPORT.md. The residual is locked honestly as a regression fixture (EconomicCoherenceTests.KNOWN_RESIDUAL_a_mixed_generalist_still_dominates_every_market) and the harness now reports "DOMINATOR FOUND."

Verification

125 tests pass (Core 54, Determinism 31, Scenario 40), determinism byte-identical, accounting reconciles, scope clean, Debug=Release. Labor investigated → no change. Goldens deliberately re-baselined (D-027). Docs corrected: D-028 supersedes the over-optimistic D-025.

No PR merged. Human gate not run. M1 not begun.

🤖 Generated with Claude Code

[HSpector1]

Independent Adjudication — M0 Fail / Rewrite Boundary (PR #5)

A. Review identity

• Repository: github.com/HSpector1/Restaurant · PR: M0 correction 3: economic coherence — High A (forecast) FIXED, High B (dominator) NOT resolved (Fail/Rewrite) — do not merge #5 (fix/m0-economic-coherence → fix/m0-price-forecast-integrity)
• Exact commit: 7bd8060f0a0a7cbbdcd96a6d06e28e328abc9ce2 (base 7fd1fbc). PR Foundation + M0 Headless Service Lab (stops at M0 gate) #1/M0 audit corrections (H1-H3 + M4-M7; NEW-1 disclosed) — do not merge pending re-review #2/M0 second correction: strategy integrity (price elasticity) + forecast direction — do not merge pending re-review #3/M0 correction 3: economic coherence — High A (forecast) FIXED, High B (dominator) NOT resolved (Fail/Rewrite) — do not merge #5 all OPEN, unmerged; main 06f0a7e untouched; no binaries; no M1/persistent/reputation/visual code added.
• Date: 2026-07-29 · Env: macOS Darwin arm64, .NET SDK 8.0.423 (fresh clone, read-only)
• Method: 5 parallel read-only sub-reviewers (correction-integrity, residual-dominator, search-strength, root-cause, rewrite-boundary), each self-cloning 7bd8060 with its own held-out prime seeds; lead independently reproduced the dominator and the arrival-composition code path.

B. Executive decision

Technical-correction verdict: Corrected (affordability + forecast fixes are sound; keep them)
Product-hypothesis verdict:   Fail (the locked M0 strategy-integrity question is not met)
Verdict: Fail
Action:  Rewrite   (a HEADLESS M0 demand-model experiment — NOT M1)
Human tests: Not authorized
M1:          Not authorized

The builder's Fail/Rewrite is upheld. Two of the builder's supporting claims are corrected, and the prescribed rewrite is narrowed: it does not require M1 (reputation/repeat-visits).

C. Claims table

Claim	Status	Note
Affordability correction (AffordBp)	Confirmed	smooth, segment-sensitive, no cap, no name branch; value diner keeps ~9% at $85 on a $22 budget; premium-only/old-hybrid dominator defeated (loses lunch)
Forecast pricing-direction fix (High A)	Confirmed	value-lunch 1.0×→1.5× down in both forecast and sim; reuses the sim's own primitives; name-agnostic
Seat-forecast (H3) preserved	Confirmed	46→66 still predicted down
Residual mixed dominator exists	Confirmed (Blocker)	reproduced on multiple disjoint held-out seed sets; robust; 0% loss-prob in every market
"Worst-market regret ~1–2%"	Refuted	true regret ~12–14% vs a properly-searched frontier; still within the 15% dominator bar and wins the value lunch outright
"Premium anchor has NO present-service cost"	Refined	real cost (crowds ~28 value covers off the Grill, +failures) but ~4× dominated by the premium upside
Search strength	Confirmed	independent stronger searches also find the dominator; not a search artifact (caveat: harness climbs the generalist harder than the frontier, so its own "−1%" number is unreliable — the honest figure is ~12–14%)
Root cause = menu-independent arrival composition	Confirmed	MakeParty samples segment from raw SegmentMixBp; per-segment segConv is computed then discarded
"Fixing this requires M1"	Refuted	a pre-service headless consideration→arrival-composition rewrite is legitimate M0; ~90% of the machinery already exists
Determinism / accounting / scope / regression	Confirmed	125 tests pass ×2, Debug=Release checksums, cent-exact reconciliation, no float/Random/wallclock in Core, scope clean

D. Residual-plan evidence

Fixed plan (unchanged across markets): Ribeye(6)@$81.56, Burger(4)@$21, Scallops(3)@$44.45, Fondant(11)@$33.06, 48 seats, full 8 staff.
Held-out medians (multiple disjoint prime seed sets): lunch ~$2157, social ~$4825, enthusiast ~$4791, 0% loss probability in every market. Against a properly-searched per-market frontier (lunch ~$2505, social ~$5202, enthusiast ~$4835 — themselves all mixed-anchor plans) the fixed plan's worst-market regret is ~12–14% (lunch), well inside the milestone's ≥85% dominator bar, and it wins the value lunch outright vs pure-value and pure-premium regimes. Robust under perturbation (burger/anchor price, seats, staff, segment share); some perturbations (fewer seats, one fewer FOH) improve it — a robust generalist, not a narrow optimum. Nuance: the two dinner markets slightly prefer dropping the burger (pure-premium edges the mixed plan there), so "context" amounts to "keep or drop one cheap main" — not a regime change.

E. Economic decomposition (why it stays strong)

Adding the premium anchor to a value chassis raises contribution in every market (+$1.2k lunch / +$3.7k social / +$4.6k enthusiast). The premium dishes are ordered only by the high-budget minority (affordability correctly stops value covers buying them); that minority is present in every market's fixed arrival mix (lunch 65/25/10, social 15/60/25, enth 5/30/65). Per-cover contribution: value ~$5, social ~$29, enthusiast ~$47 — so skimming the minority is 5–9× more valuable per cover than serving value diners. The anchor's present-service cost (Grill contention, ~28 lost value covers, +failures) is real but ~4× dominated. Multi-layer (fixed demand composition + premium recipe economics + affordability sorting); the primary lever is demand composition.

F. Arrival-model assessment — the load-bearing defect

Demand sequence: market fixture → ConversionBp (menu-sensitive scalar) → ArrivalCurve scales TOTAL volume → MakeParty picks segment from raw SegmentMixBp (FIXED) → order → cook → pay. Arrival VOLUME already responds to the menu (cheap lunch menu convBp ~6471 vs premium ~1989); arrival COMPOSITION does not (served mix tracks the fixture regardless of menu). Critically, ConversionBp already computes a per-segment segConv (appeal × price-resistance) and then throws it away by collapsing to one mix-weighted scalar. So the deterrence signal that should reshape who walks in is computed and discarded — the same class of dead-primitive gap as the earlier unused budget. This menu-independent arrival composition is load-bearing: it guarantees the premium anchor a minority to skim in every market.

The existing menu-coherence deterrent (2.0× median) does not engage on the residual plan (Ribeye $81.56 < 2× the $44.45 median = $88.90 → zero penalty), so it is toothless against this exploit.

G. Rewrite feasibility — headless M0, not M1

The rewrite is achievable inside a headless, single-service, deterministic M0 and does not need reputation or repeat-visits. It is two coupled pieces:

1. Menu-responsive arrival composition (consideration). Reuse the already-computed per-segment segConv as pre-service consideration; sample MakeParty's segment from a consideration-reweighted mix (not the raw fixture), over a finite market pool so broad positioning has opportunity cost. This makes an expensive-reading menu suppress value arrivals and a value menu suppress enthusiast arrivals — so a premium anchor in a value market is paid for by lost value volume.
2. A positioning deterrent that engages on realistic prices. Consideration alone is insufficient (a mixed menu reads attractive to both audiences — high consideration from value via the cheap burger and enthusiasts via the ribeye — so a naive model lets it attract everyone, reproducing the exploit). Add a positioning/coherence term (price tier + premium concentration + expected meal cost) so a contradictory "cheap + premium" offer loses consideration from both segments — modeling muddled positioning that fits no occasion's peak. This must not punish mixed menus merely for being mixed: the deterrent is a credible occasion-fit tradeoff (a focused menu wins its occasion's peak; a broad menu gets moderate draw from several but the peak of none), not a flat penalty.

H. PR preservation recommendations

• PR M0 correction 3: economic coherence — High A (forecast) FIXED, High B (dominator) NOT resolved (Fail/Rewrite) — do not merge #5 → PR M0 second correction: strategy integrity (price elasticity) + forecast direction — do not merge pending re-review #3 branch: Safe to merge as preserved correction + failed-experiment evidence — the affordability primitive, the forecast price-direction fix, the stronger search, the honest KNOWN_RESIDUAL fixture, and the doc corrections (D-028 supersedes the over-optimistic D-025) are all sound and should be carried forward. (I did not merge it.)
• Correction chain → foundation: Superseded by a rewrite branch — the M0 demand model is the thing being rewritten; fold these corrections into that branch rather than sealing the failed model as foundation.
• PR Foundation + M0 Headless Service Lab (stops at M0 gate) #1 → main: Merge only after a successful rewrite. The code is clean, but merging to main would enshrine a prototype that failed its product question; do not signal M0 success. Preserve on the correction chain until the rewrite establishes the strategy-integrity property.

One honesty note (Low): two green test names (No_fixed_plan... history, Premium_is_not_always_best) read stronger than their assertions; the KNOWN_RESIDUAL fixture itself is honest (documents, does not hide, the dominator). Correct the builder's report prose from "~1–2%" to "~12–14% worst-market regret."

I. Final gate

Verdict: Fail
Action:  Rewrite

Rewrite here authorizes planning a new headless M0 demand-model experiment (menu-responsive arrival composition + positioning deterrent). It does not authorize M1, the human gate, or any merge to main.

Narrowest next experiment (rewrite boundary)

• Product uncertainty: does menu/pricing/positioning create market-specific restaurant strategies, or only additive market coverage?
• Expensive-to-reverse assumption being tested: that a headless single-service model with menu-dependent arrival composition (not repeat-visits) is enough to make context select a different best complete plan.
• Cheapest valid prototype: the current headless engine + a pre-service consideration/arrival-composition stage + a realistic positioning deterrent; preserve affordability, forecast, pricing/WTP, determinism, accounting, the fixture set; replace only MakeParty's fixed-mix sampling and the collapsed ConversionBp; explicitly absent: reputation, repeat-visits, persistent memory, marketing, graphics, spatial.
• New pass condition: on held-out seeds, no fixed complete plan is within ~10–15% of the best per-market plan in all three markets, AND at least value/premium/mixed each win a distinct market, AND mixed menus remain viable (not punished for being mixed).
• New fail condition: a fixed plan (mixed or otherwise) stays near-optimal everywhere, or the positioning term makes mixed menus non-viable.
• Timebox: small (the machinery mostly exists). Required evidence: the strengthened frontier search + the KNOWN_RESIDUAL fixture flipped to assert absence.

J. Required owner decision

Howard & Aaron must choose the next experiment: authorize the narrow headless M0 demand-model rewrite above (menu-responsive arrival composition + positioning deterrent — recommended, since the defect is a menu-independent arrival mix and the fix is single-service, not M1), or formally re-scope M0 to accept that one broad flagship menu can be broadly optimal (only defensible if you accept that "the menu is the strategy" does not require market-specific positioning — the evidence argues against this), or abandon the restaurant-strategy concept (not warranted — the concept is salvageable at the demand-model layer). Do not merge PR #1 to main until the rewrite establishes the strategy-integrity property; M1 remains unauthorized.

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