Audit

Can the checker be fooled?

Every answer the site shows (36 cached on 2026-10-21) was copied and one fact was changed per copy: a score moved by 0.03, a rank or count moved by one, a brand swapped, a citation pointed at the wrong step. The checker re-checked each copy. 0 Qloo calls, 0 model calls. Today's rule rejects 219 of 230; each kind is below, with what got through. Any answer on the site has a “Try to fool the checker” box that runs the same code on your edit.

• 25 of 33 rejected: score nudged by 0.03.

  What got through (8)

  • Mt. Joy: “Outdoor Voices has the top score under Mt. Joy fans at 1.00, and Mt. Joy ranks 26 among its fans.”
  • Mt. Joy: “Outdoor Voices ranks 1 among Mt. Joy fans with the top score of 1.00.”
  • Mt. Joy: “Outdoor Voices ranks 1 among Mt. Joy fans with the top score of 1.00.”
  • Mt. Joy: “Outdoor Voices scores highest under Mt. Joy fans at 1.00, and Mt. Joy ranks 26 among its fans.”
  • Playstation: “Twitch has the highest score in the top group at 1.00 among Playstation fans.”
  • Mt. Joy: “Outdoor Voices has the highest score under Mt. Joy fans at 1.00 and ranks 1 among them.”
  • Patagonia: “Backcountry scores 0.98 under Patagonia fans and ranks 3 among them.”
  • Taylor Swift: “Forever 21 has the top score under Taylor Swift fans at 1.00 and ranks 1st among them.”

• 29 of 32 rejected: count or rank changed by 1.

  What got through (3)

  • Playstation: “The split left only 4 names, so the candidates came from the whole top 50 instead of the Playstation side of the split.”
  • Noah Kahan: “It ranks 2 among Noah Kahan fans with affinity 0.98, inside a single top group.”
  • Chipotle: “It scores highest under Chipotle fans at 0.95 and ranks 2 among them.”

• 35 of 35 rejected: bracketed brand swapped for an invented one.

• 18 of 18 rejected: bracketed brand swapped for a real one the cited steps don't hold.

• 33 of 33 rejected: unbracketed real brand the cited steps don't hold, added.

• 0 of 36 rejected: unbracketed invented brand, added (a known limit, not counted above).

• 36 of 36 rejected: citation pointed at a step that doesn't exist.

• 33 of 33 rejected: citation pointed at the resolve step only.

• 10 of 10 rejected: first call renamed to an uncited brand.

What got through, read by hand: seven of the eight nudged scores became 1.00, which the checker reads as the brand's own rank of 1 (it compares decimals at two places, so 1.00 and 1 are the same number to it); the eighth landed on another of Backcountry's own scores. The three ranks that got through sit in sentences that open with “It” or no brand at all, so there is no brand to bind them to. The known limit: a brand that no step ever returned, typed without the agent's brackets, can't be matched by name. The agent brackets every name it writes, and the box on the answer page brackets any new capitalised name you type, so swapping in “Nike” there is rejected.

One honest counterexample

Liquid Death and e.l.f. made Corpse Paint together. It sold out in 45 minutes and drew more than 250 million social impressions (Marketing Dive). On a Liquid Death shortlist run Oct 10, e.l.f. measured 0.62, 0.25 below the strongest name and in the lowest group, while being one of the best-known brands on the list. Taste is one input for deciding who to call first, not a verdict on a deal.

What the agent decided, counted

Across 39 cached asks the tools refused a call 0 times, the candidate list widened 4 times, and 3 stopped at the name. In 34 of the 36 that reached a decision, the first measurement tied (no clear three), and the agent named its calls anyway in 34 of them.

The second look, run on the 16 names with a current answer, reads that tie and may measure the other direction: it did in 11, then kept the calls in 3, changed them in 11 and declined in 2; 0 failed the checker. The checker sent 7 answers back once, and 7 passed the second time only by citing less. It passes its bar: 16 of 16 checked, needed 14.

audit/forks.json (Redis only, 0 calls); audit/second-look.json, run 2026-10-21 under audit/second-look-protocol.md. The second look isn't on the answer pages; the answer of record doesn't change.

What the ask agent may say

The agent on the front page chooses what to measure next and writes the answer, but it can only measure through Crossover's own tools, and its answer is shown only if every name and number in it appears in a step it cites. When a line doesn't match, the page shows the steps without the answer and says so. Each prepared ask's candidates (eight for Allbirds and Mt. Joy, six for Calm) tie within 0.05 by taste, because they come from the top 50 of the seed's own affinity list; the three shown are the agent's pick inside the tie, with its reason in the trace, and the overlap check is shown whether it finds a pair or not.

The checker has blocked 7 answers so far; blocked under the current rule: 3. Each, with the sentence that failed, the reason and what the current rule says:

The checker reads names and numbers, not meaning; a true sentence can be blocked, a misleading one with correct numbers can pass.

• Liquid Death, first live run (ask-v1, rule v1) Under rules v2 and v3: still blocked.

  “Whatnot scores highest in the top group at 0.86 and sits closer to Liquid Death (0.86) than to Death Wish Coffee Co. (0.75).” (cited step 3, 5): Liquid Death isn't in a cited step; Death Wish Coffee Co. isn't in a cited step; Whatnot isn't in a cited step.

  “Rose In Good Faith scores 0.85, and it is the only candidate whose fans also rank Liquid Death, at 11.” (cited step 5): Liquid Death isn't in a cited step; Rose In Good Faith isn't in a cited step.

  “The Berrics scores 0.84 and ranks 9 among Liquid Death fans.” (cited step 5): Liquid Death isn't in a cited step; The Berrics isn't in a cited step.

  “Only 12 brands sit closer to Liquid Death, while 39 sit closer to Death Wish Coffee Co..” (cited step 3): Liquid Death isn't in a cited step; Death Wish Coffee Co. isn't in a cited step.

  “Liquid Death's top markets are Los Angeles, San Diego, Denver, Portland and Austin, while the rival's are Buffalo, Boston, Pittsburgh, Philadelphia and Cleveland.” (cited step 3): Liquid Death isn't in a cited step.

• Olipop (ask-v1.2, rule v1) Under rules v2 and v3: still blocked.

  “There is no clear three: all four candidates sit in one group within 0.05 of each other.” (cited step 5): 0.05 isn't in a cited step.

• Mt. Joy (ask-v1.2, rule v1) Under rules v2 and v3: still blocked.

  “No same-field rival was found, so the candidates came from the fans' top 50.” (cited step 2, 3): 50 isn't in a cited step.

• Mt. Joy (ask-v1.4, rule v1): the score-binding rule was then narrowed: the number was the method's gap, not a brand's score Under rules v2 and v3: passes.

  “Outdoor Voices sits closer to Mt. Joy fans than the rest at 0.97, but the other seven are within 0.05 of it.” (cited step 4): 0.05 isn't Outdoor Voices's own number.

• Mt. Joy (ask-v1.5, rule v1): the score-binding rule was then narrowed: the sentence was a group range, not one brand's score Under rules v2 and v3: passes.

  “All eight candidates scored between 0.94 and 0.97, with Outdoor Voices at the top.” (cited step 4): 0.94 isn't Outdoor Voices's own number.

• Mt. Joy, re-warm after the popularity fix (real Oct 14) (ask-v1.6+t2, rule v1): the numbers after Outdoor Voices described the other seven; the binding rule reads them as Outdoor Voices' own Under rules v2 and v3: passes. re-checked against the same cached steps; the blocked run itself was replaced by the Oct 14 re-warm.

  “Outdoor Voices sits closest at 0.97, with seven others between 0.94 and 0.95.” (cited step 4): 0.94 isn't Outdoor Voices's own number; 0.95 isn't Outdoor Voices's own number.

• Noah Kahan, the judge-name probe (real Oct 15) (ask-v1.6+t2, rule v1): the numbers after 'against' described the other candidates; rule v1 read them as Little Words Project's own Under rules v2 and v3: passes.

  “Little Words Project sits in a lower fame band with popularity 0.69, against 0.95 to 0.99 for the others.” (cited step 4): 0.95 isn't Little Words Project's own number; 0.99 isn't Little Words Project's own number.

Until Oct 15 the fixes went into the tools and the prompt, never into the check: the tools now name both sides of the split and state the method's own numbers, and the score-binding rule was narrowed to the clause right after a brand. Recovered from the cache by scripts/blocked-extract.ts.

Rivals and the collision map

Measured Oct 11 on the hackathon API (probe/VERIFY-4.md and the build reports). These are limits of the front door, not held-out results.

• Chelsea vs Arsenal fails both pass marks. Of 60 brands in the two top-50 lists, 6 lean Chelsea, none lean Arsenal and 54 are shared, and the two maps share 9 of their top 10 US metros. To Qloo, two Premier League clubs read as one US audience, so the pair isn't offered. The other three test pairs passed both gates.
• The map compares percentiles, not raw ranks. Each side's heatmap rank is skewed by how many cells it carries (Olipop's matched cells sit 0.24 above Poppi's at the median), so a raw diff paints most of the country one colour. Each side is re-ranked within the cells both carry. A cell leans one way at a gap of 0.25 or more, only cells in at least one side's top half are drawn, and the page shows a lean per cell, never a number.
• The same-field cut has a ceiling. A brand sharing a second-level genre tag with either side (baseball, drinks) moves out of the columns into its own group. Tags are kept to 12 per brand, so a brand whose genre tag sits past the cut (Juventus) can slip into the columns. The top-level tag was tried first and was too broad: food and beverage took 57 of Olipop vs Poppi's 89 names and emptied Poppi's column.
• The rival finder picks for 7 of 9 test brands, 4 by genre tag and 3 by Qloo's product category: Olipop, Erewhon and A24 carry no brand genre tag in 98 to 114 tags each. A24 finds nothing in its top 50 and gets Qloo's competitor labels only. Artists and people get brand chips, never an automatic pair. Erewhon's pick (Clean) shares a supplements tag and is off-axis; Qloo's labels under it name the real grocers.
• Qloo has one Stanley brand, and it reads as tools. Typed “Stanley” asks brand or artist. The brand is the same entity the YETI vs Stanley pair uses, picked for its bottles and tumblers, but its competitor labels are Milwaukee Tool, DeWalt and Craftsman. That's Qloo's data; we don't override it.
• An age bracket doesn't split the map. Run real Oct 12, before anything was built (8 calls, scripts/age-probe.ts). At 24 and under, every drawn cell reads “both” (Olipop vs Poppi 1,500 of 1,500; Liquid Death vs Celsius 1,485). At 55 and over, a slight lean. No leaned cell flips between the brackets, against a pass mark of 15 in 100, and about 40 of the 1,500 hottest cells are shared between them. An age-bracketed heatmap shows where that bracket lives, not where the two brands part, so the map stays national.
• An audience segment doesn't reorder the eight. Run real Oct 14 (about 8 calls, scripts/audience-probe.ts), pass mark set first: for each seed it probed, some Qloo audience must change the top three and move a name by two ranks. Under “Parents With Young Children” and “Vintage Apparel” every one of Allbirds' and Mt. Joy's eight dropped by the same amount (0.06) and kept its order, and the two segments returned identical numbers. On this key the audience signal reads as a flat discount, not a lens, so no audience step was built.
• Where a candidate leads doesn't separate the eight. Run real Oct 15 (8 calls, scripts/markets-probe.ts), pass mark set first: two of Allbirds' eight must differ from Allbirds' top five metros in two or more, and one of the agent's three must have an Outdoor Voices city in its top five. It passed on paper (8 of 8; Warby Parker and ClassPass have Austin), but each brand's top five sits inside a tie, so the difference is tie-break order:

  Brand	Metros within a point of its fifth
  Allbirds	13
  Bombas	7
  Warby Parker	26
  ClassPass	8
  Parachute	14
  Tuft & Needle	11
  Felix Gray	7
  Quince	12
  Summersalt	8

  Warby Parker “leading in Austin” is one of 26 metros at the same score, and Allbirds scores Austin the same as its own top four. The split map's own diff was tried too, candidate against Allbirds at 0 calls (scripts/lean-probe.ts): 1,375 to 1,487 of the 1,500 drawn cells read “both”, and none of the three leans its way in a metro outside Allbirds' top eight except a single Dallas cell. So no market column and no rings for the first call.
• Six names a judge might type, run before you do. Real Oct 15 on the live site (scripts/judge-names.ts, about 92 calls), names and pass mark written first: four of six return a checked answer cold in under 45 seconds with none blocked, and each first call sits outside the seed's own field. It fails. Four came back checked in time, but one was blocked and two first calls sit in the seed's own field as a reader would put it.

  Typed	Rival	Seconds	Calls	Answer	The agent's three
  Los Angeles Dodgers	Oakland Athletics	32	19	checked	Las Vegas Raiders, San Francisco 49ers, Los Angeles Chargers
  Noah Kahan	none	24.6	17	blocked	Pura Vida Bracelets, Altar'd State, Love Your Melon
  Cedric the Entertainer	none	cached	0	checked	Cadillac, Ford Motor Company, Lamborghini
  PlayStation	Xbox	39.3	21	checked	Twitch, Electronic Arts, Razer
  Calm (asked which one; brand picked)	Headspace	4.4 + 36	21	checked	Peak, MasterClass, Joytunes
  Chipotle	none	26.3	14	checked	F'real Milkshakes, MOD Pizza, Thomas' Breakfast

  What it shows. The Dodgers' three are NFL teams and Chipotle's include MOD Pizza: the own-field cut is a shared genre tag, and baseball and football carry different ones, while Chipotle carries none, so neither cut caught a reader's idea of the same field. PlayStation's three (Twitch, Electronic Arts, Razer) sit outside its genre tag and inside gaming. Noah Kahan's answer was blocked by the checker on a sentence that was right (“Little Words Project sits in a lower fame band with popularity 0.69, against 0.95 to 0.99 for the others”): it bound a range to the first brand named. The rule was refined the same day and the answer now shows; see the blocked list above. Cedric the Entertainer replayed from an earlier cold run and his three are carmakers, the fame-saturated case. Calm asked which one first, as it should.
• When Qloo fails mid-ask. Drilled real Oct 15 on a local build with forced faults (lib/force.ts, development only). A Qloo 503 on the fourth call stopped Hoka after two steps; the page kept them, the unanswered step said Qloo didn't answer it, and the message said so. The retry reused the three answered calls and finished. A 12-second stall on Oatly's fourth call: the page kept its steps and kept measuring, and the answer arrived at 70 seconds. At the daily cap the page keeps its steps and says the allowance is used.
• Agent or pipeline: the routes, counted. Across 39 cached asks, counted Oct 15, the agent took 5 distinct routes; 3 stopped at the name, before a call (scripts/routes.ts reads them from the cache, 0 calls). The order of tools is set in its instructions, and the branches follow what Qloo returned: whether it knew the name, whether a rival came back, and whether the split left enough names. 18 went through the split, 13 had no rival and used the fans' top 50, 4 widened from a thin split, and 1 is the replay case. Each step on an ask now says where it forked.
• On two of the three asks we probed, the other direction separates the agent's three. Run real Oct 15 (18 calls, scripts/reverse-probe.ts), pass mark set first: the agent's three must sit 5 or more places apart on the seed's rank among each one's own fans' top 50, and that order must differ from the taste order. Two of three asks pass, and the page now shows the table under “Call first”; it never reorders the agent's calls. Across 23 real deals it mostly can't be measured; see the reverse replay.

  Ask	The agent's three rank the seed	The rest	Gate
  Allbirds	Bombas 1, Warby Parker 11, ClassPass 16	9, 12, 30, 31, out	passes
  Mt. Joy	Outdoor Voices 26, Allbirds 4, Alternative Apparel out	28, out, out, out, out	passes
  Coca-Cola (the Converse replay)	Converse out, Vans out	all out	fails, no table shown

• A typed aisle can't ship honestly. Probe run real Oct 12 (9 calls, scripts/aisle-probe.ts): Qloo's brand tag that best matches a typed category, then its top twelve brands. Running shoes, hard seltzer and skincare gave lists a partnerships lead would recognise. “Streaming services” matched “Financial Services” and returned banks, and “fast-casual burgers” matched no tag at all: 3 of 5 against a pass mark of 4. A word match that confidently shows the wrong aisle is the reason the three aisles stay prepared.
• A brand's footprint decides an aisle map, so the leaders map uses three. Run real Oct 12 on cached heatmaps (scripts/leaders-probe.ts), pass mark: no brand above its share limit, and most coloured cells agreeing with their neighbours. All nine energy brands, each ranked within its own cells: Red Bull leads 2,880 of 6,134 coloured cells, because its map has 11,060 cells and 5-hour Energy's 1,235, so a small map can't lead where it has no reading (4,134 of 5,923 cells agree with their neighbours; fails the share limit of two in five). Ranked within the 1,208 cells all nine carry: the largest share is Red Bull's 195 of 1,053 coloured cells, but only 491 of 870 agree with their neighbours. The top four by cell count (Red Bull, Monster, Bang, Rockstar; 7,018 shared cells): Bang leads 1,953 of 4,834 and 3,440 of 4,587 agree, but Celsius drops out. The camp heads (Red Bull, Monster, Celsius; 3,957 shared cells, all of Celsius's map): Celsius leads 1,241 of 2,878, and 2,123 of 2,650 agree. The shipped rule: three camp heads per aisle, ranked within the cells all three carry, a pass mark of at least 3,000 shared cells, no brand above half the coloured cells, and at least three in five agreeing with their neighbours. Soda passes with Olipop, LaCroix and Simply Pop (4,681 cells; LaCroix leads 1,205 of 2,922; 2,141 of 2,734 agree; Poppi shares only 1,961). Outdoor fails on regions (Patagonia, The North Face, REI: 11,240 cells; Patagonia leads 2,242 of 6,364; only 3,485 of 6,170 agree) and has no leaders map.
• A popularity cap keeps the fit, so the answer page shows smaller names. Real Oct 18 (12 of 54 calls that day). Four seeds (Liquid Death, Olipop, Celsius, Allbirds), Qloo's brand list for each read three ways: as is, and with filter.popularity.max at 0.7 and 0.5. Pass mark, written before the run: affinity has to be the same number for the same brand in every read, or no comparison across reads means anything (it was, for every brand that appeared in two reads); and for three of the four seeds, at least five brands under the 0.7 cap with the strongest of them scoring at least four fifths of the leader's affinity (all four did: every capped list had 20 brands, and the strongest kept 0.98 or more of the leader's score). Passed. The block on the answer page uses the 0.5 cap and the product's own 0.05 tie rule, both chosen after the run. Two things the run showed: Qloo gives some brands popularity 0, which is no measurement, so they pass any cap untested and are left out; and Olipop's capped list is mostly those unmeasured brands, so it shows little. The measured file stays private under Qloo's terms.
• Qloo's named audiences barely separate partners, so there is no audience lens. Real Oct 18 (about 40 of 54 calls that day). Four seeds (Liquid Death, Olipop, Celsius, Allbirds), each with its top three brand partners, scored under eight Qloo audiences (Single, Parents With Young Children, Technology Enthusiast, Discount Shoppers, Running, Travel, Foodie, Music Festivals) with signal.demographics.audiences and the four ids in filter.results.entities. Pass mark, written before the run: most of the twelve partners spread at least 0.10 between their highest and lowest audience, and not the same top audience everywhere. 3 of 12 spread 0.10 or more (the largest 0.111); Foodie tops 6 of 12; Olipop, its three partners, Celsius and Fallen Footwear return 0.76 under every audience, which reads as a value with nothing behind it. Failed, not built. The measured file stays private under Qloo's terms; the script is scripts/audience-probe.ts.
• Fresh deals, chosen by rule, didn't reach the floor. Real Oct 16 (34 calls, no affinity read), a third replay on deals chosen by rule after the protocol froze (9e3c0aa, Oct 13). The rule was committed before the search (30da077: Marketing Dive and Food Dive, Aug 17 to Oct 16, two brands per deal, a floor of 5 scored cases) and the list before any Qloo call (b718e83). 13 deals found; 3 excluded under the rule (Netflix and Wendy's were already used, Papa Johns' partner was a film). Of the 10 left, 5 names didn't confirm (BeatBox, Biscoff and Flamin' Hot matched more than one entity; BuzzBallz, Mike's Hot Honey and Pilgrim's matched none), and Realtree and Skippy carry no brand genre tag, so no decoys can be drawn. 3 built (Hoka x Strava, Shock Top x Spirit Halloween, e.l.f. Cosmetics x Zepeto) against a floor of 5, so nothing was scored and there is no third replay line. Most of the losses are small food and drink brands, as in the notes on the API below. Both sites' listing pages wouldn't paginate, so the search rests on 26 domain-restricted web searches plus the first page of each listing (log in audit/replay-v3-search-log.md). Files: audit/replay-v3-protocol.md, replay-v3-cases.json and replay-v3-search.json; the scored file stays private under Qloo's terms.

Notes on the API

• Ten small food, drink and beauty brands searched by name, Oct 8. None came back as itself; search returned another brand (Fruition Chocolate Works as Bath & Body Works). Shipped: every name is confirmed before a call, and an unknown name stops the ask. (probe run Oct 8, probe/out)
• Twenty-four typed names resolved, Oct 10. Corona, Mars, Ghost and Prime confirmed to an entity of another type that Qloo listed first. Shipped: an exact name that another type also carries asks which one. (scripts/audit-run.ts)
• Five category words matched to brand tags, Oct 12. “Streaming services” matched Financial Services and returned banks; “fast-casual burgers” matched no tag. Shipped: the three aisles stay prepared; no typed aisle. (scripts/aisle-probe.ts)
• Olipop scored against six brands in both directions. The reverse values were identical to the forward ones. Shipped: the other direction is read as a rank among each brand’s own fans’ top 50, never as a second score. (probe/VERIFY-3.md, scripts/reverse-probe.ts)
• Top five metros for each of Allbirds’ eight, Oct 15. Each top five sits inside a tie within a point of its fifth. Shipped: no card names a call and a place together. (scripts/markets-probe.ts)
• Genre tags for nine test brands, Oct 11. Olipop, Erewhon and A24 carry no brand genre tag among 98 to 114 tags each. Shipped: the rival finder falls back to Qloo’s product category. (lib/rival.ts, probe/VERIFY-4.md)