M0 menu-responsive demand rewrite (Conditional-Pass; owner decision on residual)

[HSpector1]

What this is

The authorized headless demand-model rewrite following the PR #5 adjudication. Replaces menu-independent arrival composition (the root of the cross-market dominator) with a menu-responsive consideration model: the menu/prices shape who visits and how many, while purchase (WTP/affordability in PickBest) stays separate and unchanged.

Do not merge. This is a review + owner-decision surface. No human tests. No M1.

What it achieves

DemandModel.Capture derives per-segment consideration from relevant-option depth + positioning focus + occasion variety (BreadthAffinityBp) + a representative price-position gate, and uses it to set both arrival volume and composition (MakeParty samples the captured mix; the forecast reuses the same Capture).

Result — the per-market frontiers are now distinct regimes, which the old fixed-composition model could not produce:

• value/throughput wins the value lunch
• premium wins the enthusiast evening
• mixed wins social

Manipulation-resistant (cheap filler bounded, never-ordered dishes zero-influence, a lone anchor gives some not full premium draw, a broad menu does not capture every segment at full strength — all price-level, never a variance/breadth penalty), legible (plain-language causes), deterministic. 139 tests pass (54 Core + 31 Determinism + 54 Scenario, incl. 14 new demand tests).

The honest residual (owner decision needed)

The strict "no plan within 10% of the frontier in every market" gate is borderline-missed on held-out seeds: a premium-anchored generalist's worst-market regret is ~6–13% (median ~9%, sweep over 4 seed bases; dominator on 2/3 fresh bases).

Decisive cause is economic, not a demand defect: consideration is market-name-agnostic, so a premium menu serves the same fraction of whatever premium slice a market's pool holds, and premium multi-course checks (~5× a value cover) make that profitable everywhere. The regret is fixed by (pool mix) × (per-cover economics) — both outside the demand model — so no composition candidate (C1–C4) can close it. Closing it robustly needs an economic rebalance of value-vs-premium margins, or the excluded systems (reputation / repeat visits), or an extreme value-heavy lunch pool. One principled fixture change was made (lunch pool 65→82% value, matching its stated "value-heavy" identity, achieving 3 of 4 gates); pushing to ~95% to force the 4th was declined as gate-chasing.

Gate recommendation

Verdict: Conditional-Pass
Action:  Continue   (independent review + owner decision)

Owner decision: (a) accept the borderline all-rounder (specialists still win each pole — arguably healthy), (b) authorize a targeted economic rebalance follow-on, or (c) accept a strongly value-heavy lunch pool.

Where to look

• reports/m0/M0-DEMAND-REWRITE-REPORT.md — full report (§A–H incl. exact experiments to re-run)
• reports/m0/demand-rewrite/candidate-1.md — iteration log (R0–R6, failed square-sharpen preserved)
• docs/DECISION-LOG.md — D-029 / D-030
• Diagnostics to reproduce: --probe, --manip, --attr, --fdir, --searchbase

Note: the §10 design sub-agent fan-out was blocked twice by transient API 529 overload; the design + adversarial reasoning were done in the main loop and are recorded. Re-running that adversarial pass is listed in the report's §H for the independent reviewer.

🤖 Generated with Claude Code

[HSpector1]

Final Independent Adjudication — PR #7 (M0 Menu-Responsive Demand Rewrite)

A. Review identity

• Repository: HSpector1/Restaurant
• PR: M0 menu-responsive demand rewrite (Conditional-Pass; owner decision on residual) #7 (base fix/m0-economic-coherence)
• Base commit: 7bd8060 · Head reviewed: e5baa541f18cf9021ba706ae0b7247428518f8e0
• Date: 2026-07-30 · Env: Darwin arm64 25.5.0, .NET SDK 8.0.423, Release
• Sub-agents: A correctness · B manipulation · C search · D economics · E legibility (parallel, read-only, own held-out seeds). Lead independently reproduced every High finding and ran an independent equal-compute frontier/robustness search + economic decomposition. Repo tree verified clean and HEAD unchanged after review (strictly read-only).

B. Executive verdict

• Technical verdict: PASS. The demand rewrite is structurally correct and in scope. dotnet build -c Release clean; dotnet test = 139/139 (Core 54, Determinism 31, Scenario 54), 0 failures, stable across two runs.
• Product-property verdict: PARTIALLY established. The value-lunch-vs-premium-dinner regime split is real and viable; menu/price genuinely shape arrival composition; manipulation is bounded. BUT three documented gaps remain (H1, H2, M2 below): a borderline fragile soft-dominator, an un-surfaced attraction mechanism, and softer-than-claimed social/enthusiast distinctness.
• Action: Conditional-Pass / Defer.
• Human-test authorization: NOT AUTHORIZED yet — one High legibility defect (H2) means a 5-player test cannot honestly evaluate the positioning hypothesis, and a justified economic correction (M1) is pending.
• M1: Not authorized (unchanged, separate).
• Merge: PR M0 menu-responsive demand rewrite (Conditional-Pass; owner decision on residual) #7 → PR M0 correction 3: economic coherence — High A (forecast) FIXED, High B (dominator) NOT resolved (Fail/Rewrite) — do not merge #5 branch: do not merge yet (land the two scoped corrections first, then merge as the demand-rewrite candidate). Correction chain → foundation: do not merge (pending human gate). Foundation → main: do not merge (pending human gate; M1 separate).

C. Rewrite correctness (§5) — CONFIRMED (Sub-Agent A + lead)

The funnel is genuinely staged in code: pool = MarketScenario.SegmentMixBp used only as a weight (Demand.cs:124); captured demand = CompositionBp = pure ratio Bp*w[s]/wsum (Demand.cs:126-127); arrivals = MakeParty sampling that composition (Simulator.cs:281-286). Composition responds to menu AND price via SuitBp = appeal x ResistBp x AffordBp (Demand.cs:28-35) plus the mean-main-price position gate (Demand.cs:89,106). SegmentMixBp is never used as final arrivals (raw only in the wsum==0 fallback). Forecast reuses the same DemandModel.Capture (Forecast.cs:21) — no parallel demand model. Purchase (PickBest) never reads consideration and vice-versa; a captured party can still NoOrder. No branch on market/strategy name anywhere in the demand path (SegmentId used only as an array index). No flat mixed-menu penalty / breadth tax / price-variance penalty — posFit uses price level (mean main price), the variety term is a strictly-positive reward reweighted by BreadthAffinityBp, and there is no dispersion/stddev term. Lead reproduced the no-name-branch and no-variance checks by grep.

D. Manipulation results (§6) — PASS-WITH-NOTES (Sub-Agent B + lead --manip)

All required cases bounded and economically reasonable: one/several cheap fillers keep value consideration at floor (100→343→395 vs pure-value 7399); a never-ordered dish is byte-identical to no dish (7399/8845/3413); a lone premium anchor gives partial (E 3413→5483 < pure-premium 7784) and costs value diners; broad menu under-draws every specialist; two-item menus degrade smoothly; extreme price spread reads as incoherent (all segments depressed); same-average/different-distribution is NOT treated identically (depth+variety respond per-dish) — refuting a plausible concern. Notes: (N1) a cheap filler on a premium menu can inflate social consideration slightly above a genuine social menu (BreadthAffinity quirk; self-defeating at purchase) — Low; (M3) duplicate RecipeId throws an unhandled ArgumentException at Demand.cs:54 (ToDictionary) — a latent input-validation gap, not a demand-gaming vector; (N2) Capture ignores staffing (documented Candidate-1 scope: operational-promise/pace deferred) so demand can attract a crowd the kitchen can't serve.

E. Search evidence (§7,8,9,10,11,12)

Methods: shipped harness dominance search across fresh bases (Sub-Agent C) + an independent equal-compute multi-start hill-climb I wrote (pool ~6140, 6 restarts × 160 steps, EQUAL budget for frontiers and generalist, held-out bases), plus a robustness perturbation sweep.

Worst-market regret is seed-dependent and straddles 10% — pooled across builder, verifier, Sub-Agent C, and my equal-compute runs: ~5%, 5%, 6%, 6%, 8%, 8%, 9%, 13%, 14%, 17%, 17%, 20%. Dominator (≤10%) on roughly half of held-out bases; no hard dominator anywhere (never <2%, the old failure) and no robust clean pass. Worst market is enthusiast-evening in every case.

§8 unequal compute — CONFIRMED (High): the shipped harness hill-climbs each frontier 80 steps but the generalist 120 steps (Program.cs:330 vs :339). This under-searches the frontier denominator and biases regret downward — i.e. the shipped tool is biased against the product (over-declares a dominator). Evidence it matters: the committed search shows the generalist's social result ($5938.97) exceeding its own social "frontier" ($5861.80). My equal-compute frontiers are materially stronger ($1659–1861 lunch vs the shipped $1390), and under equal compute the generalist never exceeds a frontier.

§11 robustness — DECISIVE (lead): even where the equal-compute generalist reaches 5% everywhere, it is a fragile knife-edge: only 1/12 (base 525252521) and 0/12 (base 636363631) two-mutation perturbations stay within 10% in all three markets. It is not a broad safe basin — a player cannot casually default into it and small changes cost >10% in some market.

§12 structural distinctness — REFINED (Medium, C3): the sharp split is lunch (value) vs dinner (premium) — the lunch frontier differs on ≥4 load-bearing dimensions (menu, price regime, expected meal cost, capacity). But the social and enthusiast frontiers largely overlap (both premium-anchored Roast Chicken/Ribeye/Risotto, differing mainly in price level and seats). So the honest picture is two sharp regimes plus a softer social↔enthusiast gradient, not three fully-distinct regimes.

F. Economic decomposition (§13,14) — rebalance JUSTIFIED (Sub-Agent D + lead, independently agree)

per representative cover	revenue	contribution	station-min	seat-min	$/station-min	$/seat-min
value (burger, ~1 course)	$21.0	$16.5	7	~23	$2.36	$0.72
premium (ribeye+starter+dessert)	$104	$81	30	~27	$2.70	$3.00
premium/value ratio	~5×	~5×	~4.3×	~1.2×	~1.15×	~4.2×

The "~5× a value cover" claim is confirmed (revenue and contribution), and per station-minute (the kitchen — the forecast's own named binding ceiling) premium leverage is only 1.15× (proportionate — this is why value wins the kitchen-bound lunch). But per seat-minute premium has ~4× leverage, because EatMin is a flat 12 minutes for every party (Simulator.cs:19,201) and a party's courses cook in parallel and deliver together — so a 3-course premium dinner occupies its seat essentially as long as a 1-course value lunch. This is a real, correctable capacity-duration omission, not a design non-goal. Sensitivity (D4): course-scaling the eat/dwell time (12→24) cuts premium seat-leverage from ~4.2× toward ~2.9×, which would shrink the generalist's cross-market edge. Classification: capacity-duration rebalance justified (targeted, not gate-chasing).

G. Player-strategy judgment (§15,16) — healthy generalist, but currently illegible

Against the §15 checklist the generalist reads much closer to a healthy generalist than a strategy-collapsing soft dominator: it is never the best (wins no market under strong search), specialists beat it by a material, seed-dependent 5–20%, it carries real operational commitment/risk (a full premium kitchen: high wages, slow dishes), and its advantage disappears under perturbation (0–1/12 robust). The one caveat cutting the other way is legibility (H2): because the positioning/attraction signal is invisible (§H), a player using rough comparisons could default to the premium build without understanding the tradeoff — which would make an otherwise-healthy generalist function as a soft default in practice. Fixing legibility is therefore the highest-leverage change for the product hypothesis.

H. Forecast and explanation (§17,18) — CONCERNS (Sub-Agent E + lead, both High reproduced)

Forecast directionality is sound (existing EconomicCoherenceTests/ForecastDirectionTests pass; forecast reuses Capture). BUT the core mechanism this rewrite adds is not surfaced to the player:

• H2a (High, reproduced): ForecastSnapshot (Result.cs:4-12) carries no per-segment consideration/composition field; the forecast collapses all of attraction to one line ("Conversion of attempted visits: 78%"). A player committing a plan cannot see "you attract value 78% / enthusiast 6%" — the positioning decision is opaque pre-service.
• H2b (High, reproduced): the post-service autopsy causes are exactly {seating, kitchen, menu-fit, weak-demand} (Simulator.cs:608-620); the only attraction-related cause ("weak-demand") fires only below 35% total capture. Attracting the wrong MIX (a full room of the wrong segments) has no category and the funnel is aggregate, not per-segment — so a low-value-covers outcome in a value market cannot be attributed to attraction vs conversion vs service.

Net: the product question (§2) asks for understandable strategies with visible consequences. The consequences of positioning are currently invisible, so a human playtest would test a game where positioning silently matters but cannot be understood — not an honest test of this design's central claim.

I. Regression and scope (§19) — CLEAN

Accounting reconciles; determinism holds (Determinism suite 31/31; identical checksums across runs); goldens re-baselined and documented (D-030); RNG streams isolated; forecast immutable; causal attribution logic intact (thresholds recalibrated, cause strings not hashed → checksum unaffected); exact-money parsing, FIFO, labor, numeric guards untouched; assets remain quarantined; changed surface confined to demand seams + fixtures (BreadthAffinityBp, lunch pool) + tests + docs. No non-goal systems introduced.

Key findings (severity format)

ID: H1  Severity: High  Status: Refined  Category: strategy integrity / search
Claim: a fixed premium-anchored generalist violates the 10%-everywhere gate on ~half of held-out bases.
Evidence: pooled worst-market regret 5-20% across ~12 bases; my equal-compute search reaches 5% (base 525252521).
Why: the locked strategy-integrity property is not robustly established.
Required action: treat as borderline+fragile (see H1-mitigant), pair with M1 economic correction; re-measure.
Gate effect: prevents "Pass"; does NOT force "Fail" because of the robustness mitigant.

ID: H1-mitigant  Severity: (note on H1)  Status: Confirmed
Claim: the violating generalist is a FRAGILE local optimum, not a robust safe basin.
Evidence: 0-1 of 12 perturbations stay within 10% everywhere (independent equal-compute robustness sweep).
Why: distinguishes a healthy jack-of-all-trades from a strategy-collapsing dominator.

ID: H2  Severity: High  Status: Confirmed  Category: legibility / product hypothesis
Claim: the positioning/attraction mechanism (per-segment consideration) is computed but never shown to the player.
Evidence: ForecastSnapshot (Result.cs) has no per-segment field; autopsy has no wrong-mix attraction cause (Simulator.cs:608-620).
Why: a human test cannot honestly evaluate "visible positioning consequences" (§2) if they are invisible.
Required action: surface per-segment consideration/attraction in the forecast + add an attraction-mix cause to the autopsy.
Gate effect: blocks human-test authorization until addressed.

ID: M1  Severity: Medium  Status: Confirmed  Category: unit economics
Claim: premium has ~4x contribution-per-seat-minute leverage due to flat EatMin=12 (multi-course seat time under-represented).
Evidence: Simulator.cs:19,201; decomposition in §F; D4 sensitivity (12->24 cuts leverage 4.2x->2.9x).
Why: partly drives the generalist's cross-market competitiveness; a correctable M0 omission, not a non-goal.
Required action: course-scaled seat/eat duration (capacity-duration rebalance); re-measure regret.
Gate effect: recommended before human gate; not gate-chasing.

ID: M2  Severity: Medium  Status: Refined  Category: strategy distinctness
Claim: "three distinct regimes" overstates it — social and enthusiast frontiers overlap (both premium-anchored).
Evidence: fresh-base frontier tables; my equal-compute (social has some value items, enthusiast pure premium, but heavy overlap).
Why: the real split is value-lunch vs premium-dinner; honest framing matters for the product claim.
Required action: report as two sharp regimes + a social/enthusiast gradient; consider strengthening the social occasion.

ID: M3  Severity: Medium  Status: Confirmed  Category: robustness
Claim: a menu with a duplicate RecipeId throws an unhandled ArgumentException.
Evidence: Demand.cs:54 ToDictionary on duplicate key; reproduced by Sub-Agent B.
Why: malformed input crashes the sim rather than validating; latent, not a demand exploit.
Required action: validate/dedupe menu input (defensive; low effort).

ID: N1  Severity: Low  Status: Confirmed  Category: search methodology
Claim: the shipped harness gives the generalist more hill-climb steps (120) than the frontiers (80).
Evidence: Program.cs:330 vs :339.
Why: biases regret DOWNWARD (against the product); the builder's reported regret is if anything pessimistic.
Required action: equalize search compute in the harness (or note it); already reproduced with equal compute.

J. Final gate

Verdict: Conditional-Pass
Action:  Defer

• The demand rewrite is technically correct, in scope, deterministic, manipulation-resistant, and a genuine improvement (a real value-vs-premium regime split the old fixed-composition model could not produce) — hence not Fail/Rewrite.
• It is not Pass: the strategy-integrity gate is borderline-violated (H1), the core mechanism is illegible (H2), and a justified economic omission remains (M1).
• Defer = two small, targeted, justified corrections before re-review: (1) attraction legibility (surface per-segment consideration in the forecast + an attraction-mix cause in the autopsy) and (2) capacity-duration rebalance (course-scaled seat dwell). Then re-run the equal-compute regret + robustness. Defer does NOT authorize human testing, M1, or any merge.

K. Required owner decision

Howard & Aaron: authorize the builder to make the two scoped corrections — (a) make per-segment attraction visible in the forecast/autopsy, and (b) course-scale seat/eat duration so multi-course premium meals occupy seats longer — and then return for a re-review that decides the human-test gate? The alternative is to run the 5-player human gate now with those two as documented limitations. Recommendation: do the two corrections first — without (a) the playtest cannot evaluate the very mechanism this milestone adds, and (b) directly attacks the economic driver of the residual soft-dominator without any gate-chasing. Neither correction is M1; both are in-scope M0 finishing work.

Read-only review complete. No branch modified or merged. No human tests run. M1 not begun.