Every call we registered, including the ones we got wrong
Before each study runs, the question is written down, a probability band is attached, and — the part that does the work — the basis is labeled: did this come from an earlier measurement, or from something we merely believed? Then it is scored either way. This page carries every prediction registered between October 18 and October 26, 2026 that has a basis on it. It replaces a figure we published for eight days and have since found to be wrong.
Predicting from a measurement beats guessing by about twenty points. It does not beat it by more, and it is not a guarantee.
Across 80 predictions registered before any data was pulled, priors grounded in an earlier measurement held 24 of 39 times (62%). Priors grounded in belief or declared as a guess held 19 of 35 (54%). Measured priors do miss. Beliefs land about half the time.
What this page replaces, and why
Until October 26, 2026 we published a measured-basis record of 17 for 17 against 0 for 6 from belief. It appeared on this site, in two drafted articles and across our own standards. It was wrong twice.
First, an arithmetic error at the origin. One prediction — RAYE P3, in the table below — is recorded in its own calibration record as not measurable, unresolved. A summary written the same day tallied it as a hit, taking the running record from 15 for 15 to 17 for 17. The correct figure was 16 for 16. Nobody recomputed it, because it arrived already computed, and the extra win travelled through four documents in eight days.
Second, and larger: the figure described a minority of the work. Most projects in the same window registered predictions with a band, a falsifier and a result — and no basis label at all. Those predictions were never in the record and could not have been. The old number read as though it covered everything and covered about a fifth of it.
The rebuilt figures are weaker than the ones they replace. They are published for the same reason the originals were: a record you cannot check is a claim, and a perfect record invites the suspicion that the scoring did the work.
The record, by what the prediction was based on
| Basis | Held | Missed | Unresolved | Record |
|---|---|---|---|---|
| Measured — same instrument, same axis | 4 | 3 | 2 | 4 for 7 |
| Measured-narrow — same dataset, different axis | 5 | 2 | 0 | 5 for 7 |
| Reasoned inference off a measured aggregate | 15 | 10 | 2 | 15 for 25 |
| Plausible belief | 14 | 10 | 2 | 14 for 24 |
| Declared a guess in advance | 5 | 6 | 0 | 5 for 11 |
| Total | 43 | 31 | 6 | 43 for 74 |
Collapsed to the two families the claim rests on: measured 24 for 39, unmeasured 19 for 35. A 7-point edge, not an unbroken record. This figure moves, and it has moved both ways. It has been read at nineteen points on October 26, 2026, at eleven on October 28 after one pass added five reasoned misses in a day, and at what it says above now. A ledger whose headline figure only improves is not being kept.
Every prediction, with its basis and its outcome
Ordered as registered. The misses are not at the bottom.
| Project | # | Prediction | Band | Basis | Outcome |
|---|---|---|---|---|---|
| RAYE | P1 | Clears 50% against her own vehicles | likely | Measured | held |
| RAYE | P2 | UK share of search more than 2× US | likely | Declared a guess in advance | missed |
| RAYE | P3 | Song or lyric term outscores her name Not measurable. Tallied as a hit on the day — see the change log. | roughly even | Measured | unresolved |
| RAYE | P4 | US ratio narrows but stays below half the UK | roughly even | Declared a guess in advance | missed |
| A24 Oscar | E1 | Genuine crossover demand Split across two passes | unlikely | Plausible belief | unresolved |
| A24 Oscar | E2 | Big opening, scandal conversation | likely | Plausible belief | held |
| A24 Oscar | E3 | Already decided; A24 is positioning | roughly even | Plausible belief | missed |
| A24 Oscar | E4 | Cold subject; opening is curiosity | unlikely | Plausible belief | held |
| IP map | P1 | Person’s own domain cited under 10% of rows 1 of 60 | very unlikely | Measured-narrow | held |
| IP map | P2 | Reference domains are the largest class Tied 5–5 | likely | Measured-narrow | missed |
| IP map | P3 | No more than two brands cited on every question Exactly 2 | very unlikely | Reasoned inference off a measured aggregate | held |
| IP map | P4 | Talent-owned rate below brand-owned rate 1.7% vs 56.7% | highly likely | Reasoned inference off a measured aggregate | held |
| Asics | H1 | Position, not awareness Platform-dependent | likely | Plausible belief | unresolved |
| Asics | H2 | The lever is third-party Weakened, not falsified | likely | Reasoned inference off a measured aggregate | unresolved |
| Asics | H3 | The persona is ignored | unlikely | Plausible belief | missed |
| Asics | H4 | Asics largely absent for this persona 79% named | unlikely | Plausible belief | missed |
| Asics | PR-A | Named in at least 8 of 16 unaided rows | likely | Plausible belief | held |
| Asics | PR-B | Brooks named in more rows than Asics | likely | Plausible belief | missed |
| Asics | PR-C | asics.com cited in at most 4 of 16 | likely | Plausible belief | held |
| Asics | PR-D | Brand-set overlap below 0.50 | likely | Reasoned inference off a measured aggregate | missed |
| Asics | PR-E | Most-named brand differs by platform | likely | Measured-narrow | held |
| Asics | PR-F | Asics first-named in no more than 1 of 16 | likely | Declared a guess in advance | missed |
| Asics | PR-G | Nimbus and Kayano beat Superblast | likely | Reasoned inference off a measured aggregate | missed |
| Asics | PR-H | Cheaper shoes named more than dearer | likely | Reasoned inference off a measured aggregate | held |
| Asics | PR-I | 3 of 9 third-party domains in top 10 | likely | Plausible belief | missed |
| Asics | PR-J | Third-party domains hold 4 of top 10 | likely | Plausible belief | held |
| Asics | PR-K | Superblast holds no more than 25% | roughly even | Declared a guess in advance | held |
| Asics | PV-1 | N1 first-named differs by platform | likely | Measured | missed |
| Asics | PV-2 | N3 Nimbus first on Perplexity | likely | Reasoned inference off a measured aggregate | held |
| Asics | PV-3 | N2 Kayano first on Perplexity | highly likely | Reasoned inference off a measured aggregate | held |
| Asics | PV-4 | N4 Superblast or Metaspeed first | likely | Plausible belief | held |
| Asics | PV-5 | N5 first shoe under $180 every row | highly likely | Plausible belief | held |
| Asics | PV-6 | A4 Superblast carries a stability caution | highly likely | Measured | held |
| Asics | PV-7 | H3 Superblast beats Skyward X 2 | likely | Measured-narrow | held |
| Asics | PV-8 | H1 Nimbus beats Glycerin 23 | roughly even | Declared a guess in advance | held |
| Asics | PV-9 | asics.com in 6+ of 8 A-rows, 2 or fewer of 6 H A limb held, H limb missed | highly likely | Measured-narrow | missed |
| Asics | PV-10 | Superblast largest share among Asics names | likely | Measured-narrow | held |
| Asics | PV-11 | Nimbus not largest in its rival set | roughly even | Declared a guess in advance | held |
| Clayface | C1 | Craft talk out-totals franchise talk 15× the other way | likely | Declared a guess in advance | missed |
| Clayface | C2 | No critic-vocabulary term clears 2% Highest 0.31% | likely | Measured | held |
| Clayface | C3 | Emotional themes clear 5% 0 of 101 | roughly even | Declared a guess in advance | missed |
| Clayface | C4 | Search more than 10× its 2024 baseline October excluded as an incomplete month | likely | Measured | unresolved |
| Clayface | C5 | Audience surfaces hold under half of page one 2 of 7 | roughly even | Declared a guess in advance | held |
| Candy | H1 | September search within 10% of last year Band no wider than the instrument’s resolution — see limits | likely | Plausible belief | held |
| Candy | H2 | Every price-seeking term above its 3-year mean 4 of 10 below | likely | Plausible belief | missed |
| Candy | H3 | Candy price vocabulary clears 5% 17.51% | roughly even | Plausible belief | held |
| Candy | H4 | Costume price talk under half the candy rate 27× apart | likely | Measured-narrow | held |
| Candy | H5 | Registered retailers hold half of page one 0 of 9 | likely | Plausible belief | missed |
| BioWare | B1 | Mass Effect outweighs Dragon Age in current demand 104,450 against 41,725 — 2.5× | likely | Plausible belief | held |
| BioWare | B2 | Closure corpora name Mass Effect over Dragon Age by 2 to 1 2.55 · 3.49 · 2.22 across three corpora | likely | Plausible belief | held |
| BioWare | B3 | Baldur’s Gate 3 alone outweighs the BioWare catalogue 252,167 against 237,541. Decided by 6.2%, inside the instrument’s resolution — a hit that should not be leaned on. | roughly even | Plausible belief | held |
| BioWare | B4 | Anthem and Andromeda each under 1% of comments 3.6–4.4% and 2.4–3.9%. Disappointing entries are not forgotten. | likely | Plausible belief | missed |
| Engine market test | E1 | An invalid market is refused and the error names it Refused, and did not crash the pass — but the error named the field, not the market. Fixed in v0.83. | likely | Plausible belief | missed |
| Engine market test | E2 | A UK row says United Kingdom in every cited field | very likely | Reasoned inference off a measured aggregate | held |
| Engine market test | E3 | UK and US results differ Three .co.uk domains in the UK top ten and none in the US — the market reached the request, not just the label. | likely | Plausible belief | held |
| Engine market test | E4 | A Hindi-relevance video set differs from English 473 comments against 453. Inferred from corpus size; video IDs are not in the output. | likely | Plausible belief | held |
| Engine market test | E5 | An India/Hindi row returns Devanagari suggestions Fifty Latin-script suggestions, zero Devanagari. The market worked; the language did not change the script. | roughly even | Plausible belief | missed |
| China AI rotation | A4 | A Chinese vendor term falls more than 40% from its peak Three did, at roughly −80%. But from a January 2025 peak, not from July 2026 — a hit for a reason unrelated to the registered cause. | unlikely | Reasoned inference off a measured aggregate | held |
| China AI rotation | A5 | Western demand is up, same month year over year Up in all four markets. | very likely | Reasoned inference off a measured aggregate | held |
| China AI rotation | D1 | The exposed move is not uniform across the four markets | likely | Reasoned inference off a measured aggregate | held |
| China AI rotation | D2 | Japan’s exposed arm falls least of the four Thin on the arm totals, decisive year over year: Japan is the only market where Chinese AI demand grew. | roughly even | Reasoned inference off a measured aggregate | held |
| China AI rotation | F2 | The release effect is visible in the leader’s own series And uninformative. One band rung, on a term that took two distinct values in all of 2026. Registered too narrow against the instrument’s resolution. | likely | Reasoned inference off a measured aggregate | held |
| China AI rotation | A3 | The exposed-to-unexposed ratio falls after July 1 Two markets up, two down. No direction. | likely | Reasoned inference off a measured aggregate | missed |
| China AI rotation | F3 | That ratio sits below its first-half average Same split. | roughly even | Reasoned inference off a measured aggregate | missed |
| China AI rotation | F1 | At least three of twenty terms fail the contamination screen Two did. The naming rule worked; applying it to a French wind did not. | likely | Measured | missed |
| China AI rotation | D3 | Category terms are usable in fewer than four markets Usable in all four and too small to measure in all four, which is a different thing and the prior did not say it. | likely | Reasoned inference off a measured aggregate | missed |
| China AI rotation | C2 | Divergence is the surviving explanation All four registered explanations were rejected. The surviving one — rotation between vendors — was never a candidate. | roughly even | Reasoned inference off a measured aggregate | missed |
| Animation route — single term | A1 | Demand orders the three wide-theatrical originals by opening gross Missed on a one-term instrument. The same claim holds as C1 once each title is measured three ways. | likely | Reasoned inference off a measured aggregate | missed |
| Animation route — single term | A2 | Hoppers’ pre-release peak more than 2× Forgotten Island’s Held at 20.34×. The finding was withdrawn by a robustness test registered against it before the check was run; the median of three reads 9.12×. | roughly even | Declared a guess in advance | held |
| Animation route — single term | A3 | Wildwood lowest of the six released originals The term was measuring a town in New Jersey. On a clean term Wildwood is lowest, by a factor of four. | likely | Reasoned inference off a measured aggregate | missed |
| Animation route — single term | A4 | Franchise controls carry at least 5× the originals’ demand 2.19× with the originals on contaminated terms and the controls on bare franchise strings. | very likely | Measured | missed |
| Animation route — single term | A5 | Wildwood shows a pre-marketing baseline no other original has | unlikely | Declared a guess in advance | missed |
| Animation route — single term | B1 | Hexed’s pre-release peak beats Forgotten Island’s 12,100 against 14,800, on a series one month shorter than the others. | likely | Reasoned inference off a measured aggregate | missed |
| Animation route — single term | B3 | Gatto shows no demand eight months before release 19,600 against a pre-announcement baseline of 50. The miss is the more useful result. | likely | Reasoned inference off a measured aggregate | missed |
| Animation route — single term | B2 | Hexed opens above $13.0M and below $46.0M domestic Registered October 27, 2026, before any pull. Resolves on the Thanksgiving opening and is scored either way. | likely | Reasoned inference off a measured aggregate | unresolved |
| Animation route — median of three | C1 | Median-of-three demand orders the three wide-theatrical originals by opening gross 135,000 / 14,800 / 12,550, the same order as $46M, $13M and a modest February weekend. | roughly even | Reasoned inference off a measured aggregate | held |
| Animation route — median of three | C2 | At least six of eight titles spread more than 3× across constructions Seven of eight. Hexed spreads 450× and Wildwood 104×. | likely | Measured | held |
| Animation route — median of three | C3 | Gatto’s midpoint sits above its own pre-announcement baseline | likely | Reasoned inference off a measured aggregate | held |
| Animation route — median of three | C4 | Franchise controls carry at least 5× the originals’ demand 13.39× once both sides are measured the same way. The same claim as A4, which missed at 2.19× on the old instrument. | roughly even | Reasoned inference off a measured aggregate | held |
| Animation route — median of three | C5 | The Hoppers-to-Forgotten-Island ratio lands below 10× 9.12×, registered against our own stated position that the true figure was near 3×, with the reversal named in the falsifier. | likely | Reasoned inference off a measured aggregate | held |
What this ledger cannot tell you
- This is not every prediction. It is every prediction that can be classified. Across the same window, 26 of roughly 33 projects registered priors with a probability band, a falsifier and a result — and no basis label. A prior with no basis cannot enter a ledger organized by basis, so those are excluded and named here rather than quietly folded in.
- One project supplies 46% of the rows. The Asics brief registered 26 of the 57. Remove it and the gap widens to 88% against 50%; read it alone and the gap is 64% against 62% — no measurable edge at all. That project registered 22 predictions in 73 minutes. Whether prediction volume under time pressure degrades basis assignment is an open question on this page, not an explanation.
- Five priors are unresolved and are not scored in either direction. Counting one of them as a hit is the error this page exists to correct.
- Ambiguous outcomes are recorded unresolved rather than forced into a column. Three of the five are judgement calls — “split”, “weakened, not falsified”. A different reading would move the totals by one or two.
- A hit is not a measure of how useful a prediction was. A prior banded roughly even that lands is a coin flip landing. The bands are published beside every row so the reader can weight them.
- One prior held on a band narrower than the instrument could resolve. Candy H1 predicted search within ±10% on a Google Ads value whose neighbouring rungs sit about ±10% away. It could not have failed informatively, and is reported as a hit with that stated.
How this page changes
A standing record, not an article. Every study adds its rows here when it publishes. Corrections are published in place and dated, never applied quietly.
| Measurement | Cadence |
|---|---|
| New predictions | On publication of each study |
| Open predictions resolving on a date | Scored on that date, either way |
| Unclassifiable priors, and the count of them | Every update, republished in full |
| The Asics concentration question | When a comparable project registers 20+ priors |
| Date | What changed |
|---|---|
| 2026-10-25 | First publication, and it opens on a correction. This page replaces a published record of 17 for 17 measured against 0 for 6 from belief. An audit of the source records on this date found RAYE P3 recorded unresolved and tallied as a hit, and found that 26 of roughly 33 projects in the window registered priors with no basis label, so the published figure described a minority of the work. Rebuilt from the priors carrying an explicit basis. |
Keep your own calls on record
DEOS is the decision app built on this research. Talk through a plan or a decision, and it writes the record, checks it against your past calls and brings it back when it is time to review.
Start freeFree for 30 days, no card. Then $99 a month.
Or see the evidence for this category’s decisions in the Decision Brief →