The evidence
Every test below that has a bar had it written before it ran; evidence 1, 5 and 8 have none, and evidence 22 describes the code. The full record of each is in docs/evidence.md.
In fifteen seconds
- Shown: the agent drafted a full lineup from a town's own venues in 29 of 29 towns, and lists every Qloo call it made.
- Partly shown: venue fans like the acts festivals book. At 40 real festivals, the acts each festival booked showed more lift from the town's venues than tier-matched random acts at 27: 7 of 10 in the first round, one short of its bar of 8, and 20 of 30 in the second, the bar set beforehand. With another country's venues in place of the town's, they still did at 18 of those 30, so the test doesn't show the edge is the town's own, and the random acts start higher in the headliners' ranking, which leaves the booked acts more room to climb (evidence 34). The differences are small. Treefort 2026 used two of the three Boise venues tested.
- Not shown: that the town's crowd knows these acts, or that they would sell tickets, take the slot or fit a budget. Undercard drafts a shortlist with its evidence attached; the buyer makes every call.
Five terms used below. A climb is how many places an act rises in Qloo's ranking when a venue's fans are added to the headliners' fans. A lift, in the real-festival tests, is the most Qloo affinity an act gains from any one of the town's venues. Mid-size and opening are the two groups of slots, each with its own fee range. A fee range is a range of Qloo popularity the buyer sets, fixed to Sagebrush Sound's in this demo. It stands in for a fee because no fee data exists. Matched random acts are drawn at random from Qloo's ranking for the same headliners and matched on fee range, as docs/evidence.md describes.
Does the idea hold?
| Question | Result | Limit | Source |
|---|---|---|---|
| At 30 real festivals chosen before any run, did the acts they booked show more lift from the town's venues than tier-matched random acts? | 20 of 30, the bar set beforehand; 20 of the 22 that could be scored | No margin, and not the town's own: with another country's venues the booked acts still led at 18 of 30, and at all 22 scored they start lower in the headliners' ranking than the random acts (evidence 34). The 8 festivals with fewer than two acts in the fee ranges couldn't be scored and count against the bar. The differences are small, 0.003 to 0.059 in affinity. Most lineups don't name their headliners first. Research subagents gathered the lineups; 28 of 30 were checked against their pages. | evidence 26, evidence 34 |
| At 10 real festivals, did the acts they booked show more lift from the town's venues than tier-matched random acts? | 7 of 10, one short of the bar of 8 | Partial lineups from Wikipedia; small differences. The second round above used complete lineups because of this result. | evidence 18 |
| Without Qloo, do the model's own picks pass the tests every Qloo act passes? | 10 of 290 model picks pass all three; every Qloo act does | Qloo's own measures, which the model can't see. The two lineups share no act in 28 of the 29 towns, and one in the other. | evidence 31 |
| Does Boise's own festival book the venues Undercard reads? | Treefort 2026 used two of the three Boise venues tested | Shows the venues are used, not that their fans predict bookings | evidence 1 |
| Do Boise's venues pull Qloo's ranking toward Treefort's own acts? | No-go: 0 to 2 of 50 acts per list | Too few acts to show the signal either way. One festival can't test the idea; the 40-festival rows above are the broader version of this check. | evidence 2 |
| Can Qloo say directly whether a town's crowd knows an act? | No-go: a usable local signal for 7 of 10 acts, bar 8 | Not claimed anywhere in the product. Artists have no location of their own in Qloo, and its location data for them, tried three ways, gave a usable local reading for at most 7 of 10 acts, which is why venues are the route to a town's taste. | evidence 6 |
Does the product work?
| Question | Result | Limit | Source |
|---|---|---|---|
| Filling a lineup. Does the agent draft a full lineup in every checked town? | 29 of 29 | One fresh build each, on one day | evidence 13 |
| Does a US town a judge might type fill both groups of slots? | 8 of 10, measured twice: first with plain town names, then with towns resolved to their state | The first run was corrected by the second. A thin town is offered checked towns instead. | evidence 7, evidence 12 |
| Does the same brief give the same lineup when run again? | Every run filled both groups; three of four towns gave the same ten acts | On a second day, 3 of 4 towns gave their saved ten acts again; a second pair of headliners changed 6 of Boise's 10. | evidence 5, evidence 33 |
| Does a town name resolve to the right place? | Go: With at most one extra Qloo call | Relies on Qloo's geocode | evidence 11 |
| Reading venues. Can parks and arenas be kept out of the venue list? | No-go: They are labelled instead | A festival site can still anchor a lineup, and the label says so | evidence 10 |
| Can a festival filed as a place be told apart from a venue? | Go: Labelled, picks unchanged | The rule fits the one case seen | evidence 14 |
| Can a lineup be spread across more of a town's venues? | No-go: 0 of 4 towns | The venues that lift acts are already on the lineup | evidence 16 |
| Running order. Can Qloo score the shared taste of every pair of acts on the bill? | 180 of 180 pairs within 10 calls | A measure of taste, not of shared ticket buyers | evidence 15 |
| Does that score agree with a booker-style judgement of a shared crowd? | 5 of 5 same-crowd pairs scored above every different-crowd pair | The pairs were Claude's judgement, not a working buyer's | evidence 17 |
| Live runs. Does a live build survive Qloo's rate limit? | Retries with backoff inside a 45-second budget | A busy Qloo still ends a live run, with the replay offered | evidence 8 |
What the no-gos taught
- Qloo's location data gave a usable local reading for at most 7 of the 10 acts the venues picked (evidence 6), so a town's taste enters the ranking through its venues' fans; drafting from Qloo's location signal itself wasn't tested.
- One festival can't test the idea (evidence 2), so the test became 40 festivals with bars written first (evidence 18, evidence 26).
- Parks and arenas can't be filtered out (evidence 10), so they are labelled and the buyer sees them.
- Lineups can't be spread further (evidence 16): the venues that lift acts are already on the lineup.
What real people said, and what changed
- The first person to see an early design couldn't tell what it was for. The design was rebuilt; the current one is a chart of climbing acts, each leading with its climb and the venue behind it.
- [Still to come: this week's test with real people, against the bar written first (evidence 32): how many, what they said within sixty seconds, and what changed.]
Five objections, answered
Does the town know these acts?
Not shown
Undercard measures that Qloo ranks an act higher with a town venue's fans added to the headliners' than for the headliners' fans alone: its climb. A direct local measure was a no-go (evidence 6); the ten-festival test fell one short of its bar (evidence 18) and the thirty-festival test met it exactly (evidence 26). Both measure the venues' fans, not the town's crowd.
What does the AI model decide, and what does Qloo decide?
Shown
In every replay: the agent (Claude) plans the Qloo calls: which venues to read, in what order, within 12 calls. Qloo supplies every ranking, affinity and popularity. A fixed rule picks the acts: a climb of 5 places or more, inside the fee range, two genres shared with the headliners. The model doesn't choose the acts (evidence 22).
Why venues?
Partly shown
Artists have no location of their own in Qloo, and its location data for them, tried three ways, gave a usable local reading for at most 7 of 10 acts (evidence 6), so a venue's fans are how a town's taste gets into the ranking. A town's festival books its venues: Treefort 2026 used two of the three Boise venues tested (evidence 1).
What happens when a town is thin?
Shown
The town is resolved to its state or country first (evidence 11). With too few venues, the page says so in one sentence and offers checked towns in the same country; a group of slots that can't fill keeps open slots, never padded (evidence 12, evidence 22).
Why trust a pick?
Shown
Each act shows the venue whose fans lifted it, by how many places, with its Qloo affinity and popularity. Every Qloo call is in the trace, and a saved build replays call by call (evidence 13, evidence 19). The limits are below.
What these numbers don't establish
- That the acts would sell tickets, take the slot or fit the budget: the fee ranges are popularity ranges the buyer sets, not fees.
- That a town's crowd knows the undercard acts, or that a town's venues predict what its festivals book: the first real-festival test fell one short, the second met its bar exactly, with small differences, and both measure the venues' fans, not the town's crowd.
- That every venue behind a lineup is a club: a park or an arena can still anchor one, and is labelled.
- Stability beyond the runs shown: a few runs per town over two days, and one second pair of headliners in one town (evidence 5, evidence 33).
- That the lineup includes local acts: Treefort books many, and evidence 2 infers that most of its acts sit below the opening range's popularity floor, so Undercard drafts the touring slots and leaves local ones to the buyer.
- The control room is matched on Qloo popularity alone, not on size or style: one small club's control was an arena, another venue's a pool bar. A control that draws a different crowd makes the bar easier to clear. (Section 37, added 2026-10-28.)
- Every control is abroad, so a local lift can reflect the country as much as the town: one opening act clears the bar in all four US demo towns. (Section 37, added 2026-10-28.)
- That Undercard finds the acts festivals actually book: at 33 real festivals its shortlists didn't beat the model alone, or the same agent with another country's venues (section 38). Audited on 29 October by an agent that didn't write the script: every count reproduced from the saved answers.