Routed

Routed answers three questions for an independent artist's tour: which city, which neighbourhood, which room. Name an artist, a starting city and the usual crowd. An agent reads Qloo's taste graph for the cities where that artist's fans over-index, the blocks inside each city where they live, the rooms they go to and the artists who share their audience. It routes and dates the tour by rules a booker would sign, sizes every room to the crowd, and drafts a hold request for every venue with each figure traced to its source. On 13 tours announced after the model's release, the city ranking Routed ships (Qloo + population) found 13.2 pts more of the real cities than population alone, better for 11 artists and worse for none (results). Run without Qloo, the same model sent Sierra Ferrell to Toronto (#202 on her fan map), Atlanta (#90) and Chicago (#50), averaging #50; Routed's cities averaged #10, with Asheville, her #1 fan city and only the 238th biggest, on the tour.

It is for the people who route tours without a data team: independent artists, their managers, and the small agencies and promoters who book them.

What a judge can check in a minute

• Press "Route the tour" on the home page (the form is pre-filled). A tour takes about a minute and streams its call sheet as it works: every Qloo question in plain words, with the exact request and what came back, and every rule that changed the agent's answer.

• Open any day sheet. The room holds a number with a source (Wikidata or a named web page) and fits the usual crowd, or it's a two-night stand, or the sheet says plainly what to ask the room. One support act plays every date. Under the date, a street-level map of the city: each dot a Qloo heatmap tile about 150 m across, the booked room marked, and the neighbourhood where the fans over-index most named from Qloo's own places there (or from OpenStreetMap when they name none). The hold request is written for the talent buyer: the dates and hold, the routing, the room against the crowd, one line of Qloo evidence, the support act, and the offer fields to fill in (ticket price, deal, set lengths, last play, radius).

• Scroll to "The same model, without Qloo". The same model routed the same shows from memory before the agent looked anything up; both tours sit on one rank axis, scored on the artist's own heatmap.

• /proof tests Qloo on rulers it didn't make, each with its protocol committed before it ran: 13 tours announced after the model's release, then 49 older real tours (30 of them held out), then the same comparison on twenty artists' Qloo fan maps. The losses are there too.

• Read "Stops Qloo changed" on any tour book: the same rules re-planned on city size alone, with no Qloo or model call, and the cities that differ, with Qloo's evidence for each.

• Add to calendar gives the tour as an .ics file with every show, both nights of a two-night stand, travel days and days off; Open in email opens a pitch in your mail client; Print the tour book prints it. The tour book also says what the run cost: Qloo requests, model calls and tokens.

Related MCP server: BandSaaS MCP Server

Why it needs Qloo

A manager deciding where to play has a streaming dashboard that counts plays, not people who'd buy a ticket. A language model can recite past tours; it can't tell you where this artist's fans over-index today, which rooms they use, which blocks they live on, or who they'd also turn up for. That is what Qloo's graph holds.

On the eight poster-wall tours (six routed Oct 25, two Oct 30 re-runs), a tour made 81 to 104 Qloo requests (median 87) and 5 or 6 model calls; since Oct 30 the agent may take a third re-booking round, so a run can make up to 9. Every Qloo call is listed, with its plain-words question, the exact request (never the key) and the answer, at the end of its tour book.

Does Qloo help? Three rulers

Every number below is on /proof, and each test's protocol was committed before it ran. The fairest ruler comes first.

1. Tours the model can't have read (Oct 25)

Nemotron 3 Ultra was released on June 4, 2026. We walked all 185 pages in Wikipedia's 2026 and 2027 concert-tour categories and kept the 13 tours announced after that date (protocol: data/outside2/PROTOCOL.md, commit bcaa653; results: data/outside2/RESULTS.md). As in the head-to-head below, each ranking picks as many cities as the tour announced.

On older tours the model's lead over Qloo was 12.7 pts; on tours it can't have read, it is 4.6 and not significant. Thirteen tours is a small sample.

The same round tested rooms and openers, and there the model alone wins by a wide margin. In 102 cities of 34 real tours, the room the artist really played was in the top 3 of Qloo's room ranking (as Routed asks for it) in 29 cities, and of the model alone's in 75. Qloo's support pick matched a real opener for 1 of 22 tours; the model's matched 6. These are 2023–2025 tours the model may have read about, and it also knows each city's main rooms. Qloo's ranking favours rooms where the fans over-index over the room a tour books. Qloo's own ranking with the artist signal also scored below the same request sorted by popularity alone (28.4% against 34.3% in the top 3), which is why Routed measures every room against the crowd before booking it, and we don't claim Qloo picks the room or the opener better. Round 4 (Oct 30, pre-registered) scored the room Routed actually books, after the agent names rooms, Wikidata's ladder and the size rule: the real venue in 46 of 102 cities, against 16 for Qloo's first room and 55 for the model's first guess (data/outside2/RESULTS.md).

2. The head-to-head: 49 older real tours (Oct 18)

The benchmark below scores with Qloo's own heatmap, and the first outside check came out a tie. Neither says what Qloo's data changes. So on Oct 18 we ran a test on a ruler Qloo didn't make, with the protocol committed before it ran (the header of scripts/head-to-head.ts; two later changes are recorded there: one cut the model budget before the ablation ran, one fixed a name-matching gap in the model's favour).

• Ground truth: the cities each artist actually played on one recent headline tour (2024 or 2025, so the model may have read the announcements). 49 artists: the 19 benchmark artists with a tour on file, and 30 held-out artists nobody here picked: the first 30 single-artist tours in Wikipedia's "2024 concert tours" and "2025 concert tours" categories, taken in SHA-256 order of the page title, with eight or more touring cities in one region, read off the article's date tables by a fixed rule (lib/wikitours.ts; festivals, cancelled and postponed shows left out). Every skipped page and why is in data/head-to-head/held-out.json.

• Test 1, four rankings of every touring city in the region: population (biggest metro first), the model alone (Nemotron 3 Ultra's own top 50 for the artist, no Qloo), Qloo (the artist's heatmap ranking, as Routed computes it) and Qloo + population (a city's two places averaged; no weight to tune). Each picks as many cities as the artist played (R); the score is how many were right (R-precision).

• Test 2, the agent with Qloo taken out: Routed's own shortlist step and city rules on the 19 benchmark tours, run with Qloo's table and with a table of the region's biggest cities instead, two runs each; the score is the share of the eight cities on the real tour.

• Per-artist differences, wins and losses, an exact sign test and a 95% bootstrap interval. Results: data/head-to-head/results.json, all on /proof.

Split by set, Qloo + population beats population on the benchmark artists by 4.5 points (11–3–5, p = 0.06) and on the held-out artists by 9.3 (25–4–1, p < 0.001).

What it shows, plainly:

• Qloo's taste data carries real information about where an artist tours that population doesn't. Added to population it finds 7.4 more points of the real tour's cities, for 36 of 49 artists, on a ruler Qloo had no part in: +9.3 pts held-out, p < 0.001; +4.5 pts benchmark, p = 0.06. Routed now reads cities in this Qloo + population order.

• On these older tours the model alone scores higher. Asked from memory, Nemotron found 58% of the real cities, ahead of Qloo's ranking and of Qloo plus population, and averaging Qloo into its ranking made no measurable difference. It has likely read these tours' announcements: on the 13 tours announced after its release (above), its lead over Qloo falls to 4.6 pts, not significant.

• Inside the agent, eight picks can't tell Qloo from population (+2.6 points, not significant), just as the first outside check couldn't tell Routed from the model. This ran before the Qloo + population order shipped.

Limits: one tour per artist; a past tour is one booker's answer, made with ticket data neither side has; the model may have seen these tours in training (which favours it), and Qloo's signals may include check-ins at these shows (which favours Qloo); Wikipedia's tour articles lean to arena acts, where population is hard to beat. Cost: 61 Qloo requests (31 artist searches, 30 heatmaps), 125 model calls, about 156,000 tokens, about $0.30.

3. On Qloo's own fan map (Oct 15)

Every tour runs twice, with and without Qloo. Twenty artists, eight cities each, run on the live site on Oct 15, 2026 (first date May 6, 2027). The same model (Nemotron 3 Ultra), starting city and crowd size on both sides; the model alone gets the artist's name, Routed's agent gets Qloo. Each number is the average rank of a tour's cities among every touring city in the region on that artist's heatmap (lower is closer to the fans).

Routed's cities averaged #8.3. The same model without Qloo averaged #26.2.

On this ruler the gap is widest where an artist's audience doesn't follow the population: the model alone sent Zach Bryan to Los Angeles (#244 on his map) and Tyler Childers to Chicago (#161), while Qloo put their strongest fans in Nashville and Asheville. The gap is smallest in the UK and Ireland, where touring cities are few and the biggest ones are also where the fans are. For Clairo the model alone did slightly better.

Read this with Qloo's own ruler in mind: Routed chooses from the heatmap ranking it is then scored on (these tours predate the Qloo + population order), so the number shows how far a model routing from memory strays from where the taste graph puts the fans, not that the graph is right. The benchmark page says so, gives the median stop and the share of stops in each artist's top ten, shows the exact baseline prompt, and names where the model alone keeps up. A model-alone city is placed by its name, then by the named room's coordinates on Qloo, then by Qloo's locality search; a stop that resolves to a town with no touring city near it counts last, and one that can't be placed at all is left out rather than counted against the model.

The benchmark tours were routed on the build of Oct 15 (commit d81a398); their rooms were then re-sized under the stricter capacity reader and the bar, pub and orchestra-hall rules that landed the same day (/api/admin/resize, which changes rooms and nights, never cities). Fan-map ranks are unchanged by that.

The first outside check (Oct 15, kept as history)

A ruler Qloo didn't make. For 19 of the artists (no recent UK headline tour could be found for Jamie xx), data/real-tours.json holds the cities of one recent headline tour they announced and played, from Wikipedia tour articles, press listings and label releases, every source linked; festivals, support slots and in-stores left out. Suburbs and venue towns count as the touring city within 60 km (Morrison is Denver; Wolverhampton is Birmingham). How many of the cities each tour booked did the artist really play?

49% of Routed's cities were on the artist's real tour (74 of 152); 48% of the model alone's were (73 of 152). A tie. (Eight picks against a twenty- to forty-city tour mostly measure whether both sides named the big cities; the head-to-head above fixes that.)

What it means: Routed matches real touring as well as a model that may have read these very announcements, while booking cities about three times closer to where Qloo puts the fans. What it can't mean: a past tour is one booker's answer, made with ticket data neither side has, and a big act's tour plays twenty to forty cities, so eight good guesses can all be on it.

The booking rules

Each rule is deterministic, tested, and logged on the call sheet when it changes the agent's answer.

• Radius clause, by region: no two shows closer than 150 km in North America, 120 km in mainland Europe or 60 km in the UK and Ireland, where a real tour plays London and Brighton. Never between two major markets of 1.5 million or more: New York and Philadelphia each carry a show.

• Major markets stay: a city of 1.5 million or more in the artist's top ten is always on the tour when the route can reach it, and the smaller city inside its radius gives way. No UK tour without London.

• No leg over 1,800 km between two shows (ferries and borders counted in hours, at van speed). A tour split across two coasts keeps the group with the artist's strongest cities and replaces the rest with cities that join it; a long leg left over is broken by the swap that gives up the least fan strength; one that no city with rooms can close stays, and its day sheet says so.

• Rooms sized to the crowd: a room holding 60–250% of the usual crowd for one night, a close fit (85–130%) first, then one the fans sell out (60–85%), then a loose fit (130–250%), since selling out a slightly small room beats half-filling a big one; failing that, 50–60% for two nights (both dates go on the calendar, and the two nights hold everyone); then up to four times the crowd in a reduced configuration; then a room whose capacity couldn't be found (the sheet says to ask). A room known to miss by more is booked only when nothing else is left. A day sheet says a room "fits" only at 85–130% of the crowd; otherwise it says what the room holds against the crowd. No stadium below a 30,000 crowd, never a park or playground, no housing, no gallery or art centre for 700+, no pub, bar or food-cart yard for 1,000+, no café or club for 3,000+, no orchestra's own hall for a rock, pop or hip-hop act.

• Capacities, one number per room: Wikidata's maximum-capacity property first (the item must be the room itself, not a hall inside it: Nordstrom Recital Hall is not Benaroya Hall), then a web page that names the room in whole words, names the city, and puts the number next to the name. Answers live in one versioned index, keyed by Qloo's room id, so a room never shows two numbers. Social posts, LinkedIn, YouTube and articles about planned buildings are not sources.

• One support act for the whole run, as on a real tour: the agent's most frequent pick. Never the headliner itself, a shared name, the headliner's own member or a member of another band (by Qloo's description: Grian Chatten fronts Fontaines D.C.; Kirk Hammett plays in Metallica), well below the headliner on Qloo's popularity (at least 2.5 times further from the top: Lukas Nelson & Promise of the Real, nearly Sierra Ferrell's size, gave way to Kelsey Waldon), and sharing a specific genre or two umbrella ones (no bluegrass under post-punk). The other artists each city suggested are listed as local options.

• The calendar: the shortest drive (every order searched exactly up to 12 stops, then 2-opt), at most 8 hours of driving on a show day, travel days for longer jumps, a day off after five shows.

• Clean names: venue names cleaned for the poster ("Think Tank?" was one), two cities with one name printed with their state, and census-place market names replaced by the city a booker writes (Ashland, not Meads; Palm Springs, not Indio; Margate, not Saint Peters).

Inside each city

For every stop Routed asks Qloo for the heatmap of that one city: a few hundred tiles about 150 m across. A rule finds the hottest tile, names its neighbourhood by a vote of the places within 600 m of it, each filed by Qloo under a neighbourhood (with how Qloo characterises it: a creative hub, a historic district), measures the booked room against it, and anchors the poster run on it, so posters go up where the fans live. A room "sits in" the fans' neighbourhood only when Qloo files the room under that same name: in Portland, Sierra Ferrell's hottest blocks are Downtown and so is the Crystal Ballroom, so the sheet says "McMenamins Crystal Ballroom sits in Downtown, 0.6 km from the hottest blocks"; a room filed elsewhere gets its distance instead ("1.6 km from the fans' hottest blocks (Chinatown); Qloo files it under Northern Liberties"). Where Qloo files no neighbourhood at all (Nashville's places carry none), the neighbourhood OpenStreetMap puts the hottest tile in is named instead and credited (Downtown Nashville; Newport, across the river from Cincinnati); where neither names one (Asheville), the sheet says so and points at the nearest place. The map sits under the date on every day sheet.

Use it from your own agent

Routed is also an MCP server (Streamable HTTP). Add https://routed-tours.vercel.app/api/mcp to any assistant that supports remote MCP servers and it gets five tools:

How it works

flowchart LR
  A[Artist + start city + crowd] --> S[Qloo search]
  S --> H[Heatmap: one polygon per region, tiles to ranked cities]
  S --> D[Audience: ages, sound, mood, podcasts, brands, film, TV, books]
  S -.-> G[Control: same model, no Qloo]
  H --> L[1 Shortlist: model]
  L --> C[2 Scout: rooms by size tier + support acts per city]
  C --> B[3 Book: model picks rooms + one support act]
  B --> K[check_plan: rules object, the model re-books, up to 2 rounds]
  K --> R[Rules: radius, major markets, legs, room sizes, nights, support act, route, dates]
  R --> N[Per stop: city heatmap, neighbourhood, poster run, aftershow, shared tastes]
  N --> W[Pitches: model writes, every figure checked]
  G --> X[Score both tours on Qloo evidence]
  R --> X
  X --> O[Outside check: real past tours]

• The agent (NVIDIA Nemotron 3 Ultra on Nebius Token Factory) shortlists cities with a few alternates, Routed scouts each one through Qloo, and the model books a room per city and one support act for the run. Then check_plan runs the rules over the booking and sends every objection back; the model re-books what the objections name, up to three rounds. A re-booking replaces the booking only when it leaves fewer objections; a worse one is set aside and the agent is told why, and each objection names the cities or swaps that would clear it; a round that is no better as a whole keeps the agent's changes that help on their own. Measured on /how (data/agent-loop/runs.json): on the eight tours of the Oct 25 wall as first routed, check_plan raised 15 objections on the first bookings and after two rounds 13 were open, some of them new (Alvvays went from 1 to 4). Oct 30 loops: the agent cleared 6 of 9 objections itself across 4 re-runs (leg cap 4 of 5 gone, radius clause 1 of 2, room 1 of 1, major market 1 of 1, with one new fan-city objection); the one run on the final code, the flagship artist, cleared 0 of 2, settled by rule (/tour/d53j244bwp): it asked to scout Lexington and Knoxville and booked neither, so Louisville stayed inside Cincinnati's radius and Portland 1,895 km from Denver. Three runs a version is a small sample, published as it came out. Cities are read in Qloo + population order (the ranking that tested best on real tours); Qloo's top three fan cities always stay, and only its top 25 are candidates unless a long leg needs closing. Every call returns strict JSON with reasoning off.

• Rules decide what must be exact (above). When a rule changes the room the model chose, the model's sentence about the stop is replaced by a rule's, so a wrong name never stays on the page.

• Pitches are checked: Nemotron writes each hold request from that stop's evidence lines only; every number in it must appear in the evidence (a two-night stand names both dates) or it is struck through on the page. A stop's one-line reason is checked against what the booking agent was shown for that city.

• City strength is a tile's Qloo affinity times its popularity, both from one heatmap call per region. Affinity and popularity are percentiles within the response they came in (measured: both run evenly from 0 to 1 in every response), so two countries asked separately would each have a "1.0" tile; one polygon, one scale.

• The artist someone typed wins over Qloo's first guess ("Wednesday" is not "Wednesday Campanella"), and the other matches are offered on the tour book.

• Fallbacks: with no model key, a spent daily budget or a model error, the same steps run as a fixed plan with template pitches, and the tour book says which ran. When the Qloo allowance is spent, the visitor gets the stored tour for that artist, or the nearest one on the poster wall, instead of an error. A run survives a closed tab; the record is saved as it goes.

Run it

Node 22.

git clone https://github.com/RohanGlitched/routed && cd routed
npm install
cp .env.example .env.local   # then fill in the keys
npm run dev                  # http://localhost:3000
npm test                     # 231 unit tests: every booking rule, scoring, capacity parsing, evidence checks (both nights of a two-night stand), the outside check, the head-to-head's scoring and tour-table reader, the city map, the calendar file

No keys? npx tsx scripts/dev-fixture.ts writes a sample tour to .data/ (every number in it is made up), then npm run dev opens its tour book so you can see every page.

scripts/head-to-head.ts runs the head-to-head (sample, run, score; its header is the protocol). node scripts/probe-qloo.mjs "Artist" measures every Qloo call Routed depends on and saves the raw responses under the gitignored .probe/. The current tree holds no raw Qloo response: the evaluation files in data/ keep only values derived from them (city, room and artist names, ranks and hit counts). Three earlier commits (bcaa653, 5d7c7c2, 0c98714) had Qloo entity ids and returned room lists in data/outside2; they were stripped in 35f45c6 (scripts/strip-qloo-data.mjs). The history is kept as it is because bcaa653 and 096d898 are the pre-registration proof the results cite. Round 4's room lists (data/outside2/rooms4.json, commit 73f7861) carried the room lists Qloo returned per city until Oct 30, when the strip script removed them (ranks, hits, the booked room and the model's and Wikidata's lists stay; history kept). The tour API carries no Qloo ids, image links, raw scores or raw street-level tiles, and none of Qloo's descriptive text (descriptions, place tags, addresses, websites): lib/publictour.ts. node scripts/bench.mjs <url> runs the benchmark; npx tsx scripts/bench-table.ts <url> <ids> prints the two tables above from the finished tours; scripts/measure-rooms.ts measures rooms a live run couldn't (Wikidata rate-limits shared serverless addresses) and adds them to the live index.

What Routed doesn't know

• Qloo affinity says where fans over-index, not how many tickets will sell. It's the first question a booker asks, not the last. The tour book's comparison measures with Qloo's own ruler and says so. Against real tours, Qloo adds to population (+13.2 pts on tours announced after the model's release, +7.4 on older ones), but the model alone predicts older tours better than Qloo does, and inside the agent the two couldn't be told apart on eight picks.

• Qloo's room and support-act rankings are weak predictors of the room and opener a real tour books: the model alone found the real venue in its top 3 in 75 of 102 cities against Qloo's 29, and a real opener for 6 of 22 tours against Qloo's 1.

• 1 of the 64 booked rooms on the poster wall has no confirmed capacity (The Broadway, Brooklyn); the sheet says to ask the room. 33 of the wall's capacities were checked by hand on Oct 30 against Wikipedia, the venue or its operator.

• Availability, holds, offers, guarantees and ticket prices come from the rooms; the pitch asks for them. The radius clause is distance, not time, and festival weeks aren't checked.

• The figure check asks whether each number in a pitch appears in the stop's evidence. A true figure written another way is struck; a wrong figure that happens to match another number passes.

• Momentum (/v2/trending) is shown only when the series moves.

• Nothing here has been validated with a working booker yet.

• Routed sends Qloo only public names (artists, cities, venues). It collects no personal data.

Credits

Taste data from Qloo. Agent and pitches by NVIDIA Nemotron on Nebius Token Factory. Capacities via Wikidata and Tavily. Cities from GeoNames (CC BY 4.0); coastlines from Natural Earth. Real tour dates from the Wikipedia articles and press listings linked in data/real-tours.json. Type: Big Shoulders and Schibsted Grotesk.

MIT licence.