The proof model
How the verdict is proven
For a promoter: this page explains why you can trust the break-even number before you sign — the formula, and how every Qloo claim carries a receipt you can re-run.
Skim the headings; the detail below is for anyone who wants to check the work.
Static page. It links the live ledger and the recorded trials — it never restates numbers that can drift.
1. The formula
Break-even tickets = ceil((guarantee + entered fixed costs − entered sponsor offset) ÷ net per ticket), where net per ticket = ticket price − variable cost per ticket. Optional fixed costs are venue cost, production cost, marketing cost, and other fixed costs. A sponsor offset is subtracted only when entered. A blank field is not entered — it is left out of the sum, never treated as zero. An explicit 0 is entered zero. No tax or fee rate is assumed. With every fixed-cost field blank, the count is ceil(guarantee ÷ net per ticket). A blank variable cost means the net is before variable costs — never net of zero costs. An explicit variable cost of 0 is cost 0. The worked example and the line-by-line breakdown on every verdict show the exact numbers used.
When capacity is known (user-supplied or Ticketmaster venue, source always labelled), the verdict also shows break-even as a percentage of room capacity, and when break-even exceeds capacity it says plainly: “Cannot break even in this room at this guarantee.” When capacity is unknown, no percentage is shown at all — only tickets needed, plus the verbatim unknown-capacity sentence.
2. The receipt model
Every Qloo-backed claim on a verdict traces to a receipt (arithmetic lines carry no receipt by design): the claim, the operation (MCP tool name or REST path), a SHA-256 request hash, the upstream requestId (null when the upstream returns none — never invented), a result count, and, where the leg holds one, an ordered ID list. In a verdict, receipts exist only for legs that returned ok. Every receipt in the drawer carries a Replay button that re-runs the exact call live; a replay matches only when the replayed ID list equals the original ID list in order — counts alone never prove a match.
Keys never appear in receipts. The REST client sends the key only as an X-Api-Key header, and MCP arguments never contain keys.
3. The fail-closed rule
When a required evidence leg fails, the verdict is withheld and no numbers render. The states:
- 503 / Qloo not configured — visible error, form disabled. Nothing is guessed. Other missing keys do not produce a 503: a missing Ticketmaster key renders the verdict with capacity unknown, and missing LLM providers render the verdict with the arms unavailable.
- Verdict withheld — missing legs named, refusal kind named (OUTAGE when a leg errored, THIN DATA when legs returned empty or skipped). No numbers shown.
- Needs input — an ambiguous headliner name returns a picker, never a guess.
- Price unknown — no break-even number without a price; partial evidence only.
4. The miss ledger
Every verdict records the Gut-vs-Graph race: the LLM-only pick, the Qloo-grounded pick, and whether they differ. A miss names the LLM pick and the contradicting Qloo pick. The recorded trials are in the benchmark ledger; the frozen, checksummed copy is the offline replay pack (recorded, not live).
5. The trial suite
The replay pack pins the 3 dated trials (inputs, break-evens, recorded Qloo picks, checksums). To verify offline: recompute ceil(guarantee ÷ ticket price) for each trial and compare the recorded Qloo pick — no fresh Qloo call is needed, and no fresh call is made.
6. The confidence label
The HIGH / MEDIUM / LOW label next to a verdict is computed by code from leg coverage — never narrated by an LLM — and it measures evidence strength, not probability:
- HIGH — all 7 core Qloo legs ok (find tags, describe, recommend, rank, where popular, demographics, entity tags) and no leg in error.
- MEDIUM — core evidence present but some leg empty, errored, or skipped.
- LOW — thin core: no ranked shortlist AND (no audience areas OR no shared tags).
Two disclosures always accompany the label. Tie check: Qloo rank returns names only, no scores — a near-tie cannot be ruled out from a verdict. Unranked shortlist: when the ranked list is empty and candidates are present, the Graph pick is the first candidate returned, not a ranked winner.