Order-dependence was the disease, not a symptom
All three panelists independently converged on it in round 1: iteration order was silently deciding discount outcomes.
GO — Canary
Curated scenarioLoop order was deciding money; the fix ships as a flagged canary.
Total mechanism convergence once the stipulated policy facts landed — then five defects found in the room's own design, one of them honestly left unresolved under the canary.
Review cost$1.00
Session3 rounds · ~5 min model time
Result5 defects found in the room's own design
Panel
GPTGrokMuse
01
The plan under review
A rework of a cart-pricing engine whose promo-stacking output depended on iteration order — the same cart could price differently depending on which promotion the loop encountered first. Proposed fix: pre-filtered single-pass application with defaulted grouping semantics. The author asked which parts survive scrutiny before a canary rollout.
Scenario contextA production cart-pricer where loop order decided discount outcomes on every order; three author-stipulated policy facts (EMP50-alone-always, seasonal-never-stacks, lifecycle-may-stack).Figures as stated in the reviewed plan — no inferred savings.
02
What changed before ship
Proposed
Landed
Pre-filtered single-pass application
→
Two-phase select-then-apply with replace-and-clear, bound by the quote-ID + dual-revision CAS invariant, candidate set rebuilt inside the payment transaction
Silent Singleton default for unknown groups
→
Fail-loud
filterPositive pre-filter
→
Inline positivity check at application
A 24h price-grace window
→
Withdrawn by its own author — 'an unapproved price guarantee'
Flat promotion model
→
An explicit group-policy dimension (SINGLE_WINNER / STACKABLE_IF_MARKED) — three independent formulations proved equivalent and were merged
Priority-then-ID ordering
→
Signed as documented-provisional pending a Product decision, with release-note discipline
Fleet-wide rollout
→
Flagged canary with an explicit fallback: canary invariant violated flips the verdict to NO-GO, keep the old engine
03
Findings
The silent Singleton default was an unsigned product decision
Defaulting unknown promotion groups to exclusive-winner semantics executes a product policy nobody approved in the hot path. Conceded by its proposer AND its defender; replaced with fail-loud.
The eligibility-base contract gap
Thresholds evaluated on post-discount totals make phase-1 selection wrong in exactly the configurations a canary is least likely to sample — 'the three-day canary will discover it in production.'
Why it stays open: The fix direction was named (pin the eligibility base to pre-discount totals) but never ratified as a change in-session; it rides the refuse-to-ship lists and the canary condition. Legal under conditional_go; would block an unconditional go.
'Immutable snapshot' was rhetoric until it was a mechanism
The consensus failed unless snapshot immutability was enforced — refined mid-round from a consent version to quote-ID + dual-revision CAS binding, with both peers endorsing the amendment.
Two-store race
Promotions and group-policies caches revise independently (plus case-variant group names) — a pricing decision can straddle two config generations. Addressed by the dual-revision CAS binding.
Stale candidate set at checkout
The candidate set must rebuild inside the payment transaction, not ride from cart-view time.
04
Decision
Landed invariantUnanimous
Discount selection and application are two phases bound by a quote-ID + dual-revision CAS — a checkout can never apply a price the customer wasn't shown.
GO narrowed from fleet-wide to a flagged canary — the narrowing is itself the panel's judgment. The unresolved eligibility-base gap is legal only under this conditional verdict.
Condition: Full rollout contingent on the flagged canary holding the quote-ID invariant; violation flips the verdict to NO-GO and keeps the old engine.
05
Preserved dissent
Preserved
All three panelists: global-beats-value and none-suppression are unsigned product decisions being executed by code. Signed as documented-provisional; revisit post-canary.
06