Every game stat site shows you a number. Almost none show you where it came from, and that is the whole question. A figure read out of the game's own files, a figure a patch note announced, a figure someone timed with a stopwatch, and a figure a wiki editor typed from memory can all be "40". They are not the same kind of true, and when a patch moves the number they go stale at different speeds.
When we built the Knowledge API, the first decision was that every fact would carry a tier saying which kind it is. Seven tiers, best first. Here they are with a real fact from the corpus for each, so the ladder is concrete rather than a taxonomy.
1. Game files
The developer's own data, read from the files the game ships or from a wiki that mirrors them mechanically. This is the top of the ladder because there is no human between the value and you.
The Vandal's body damage of 40 at 0–50 m comes from Riot's shipped weapon data via valorant-api.com. The AK-47's cycle time of 0.1 seconds and armor ratio of 1.55 come from a locally installed Counter-Strike 2, read out of scripts/weapons.vdata_c at build 24916958. Minecraft's item and enchantment figures come from Mojang's own data generator, run against a server jar whose checksum matches the one Mojang publishes. In each case the source string in the fact is a file path or a URL to the mirror, and you can go and look.
2. Patch note
Confirmed by a specific note, and dated to it. This is where the patch on a value comes from: Clove's Ruse lasts 6 seconds when cast while dead, per the 12.05 note, which states the change: "Smoke duration 14s >> 6s when casted while dead." The old 14 stays in the file with patchTo: 12.05.
A patch note is a promise from the developer about what changed. It is very good evidence for the change and slightly weaker evidence for the exact value, because notes round, and because they describe intent while the files describe the build.
3. In game
Observable in the client without any tooling: the number printed on the ability card, the price on the buy menu. Ruse costs 150 credits and has two charges (one while dead). Anyone can check it by opening the game, which is the point of the tier.
4. Measured
Derived by Ascend from replay files, with the sample size and the spread stated. Astra's Gravity Well: median 9.99 seconds across 12 replay instances, interquartile range 0.01. The corpus records the method (timed actor lifetime), the count, and its own confidence in the extraction.
Measurement earns its place on the ladder when it can check the tiers above it. Ruse's alive duration is 14 seconds per the 12.05 note, which says that includes half a second of formation and half of expiration. Measured across 349 replay instances, the median is 15, interquartile range 0. The fact carries both in an agreement block: official 14, measured 15, delta 1, tolerance 1.4, status agree. Nobody has to pick a winner; the reader gets the note's number, the replay's number, and the reason they differ.
5. Community
Adopted from a third-party dataset, attributed. The Fandom wikis, PokéAPI. Fortnite's Pump Shotgun does 84 damage at Common and 116 at Legendary per the Fortnite Wiki, tagged to patch band 42.00. Epic runs no wiki of its own, so this is the best public source and the corpus says so, with the licence.
Community is a ceiling, not a judgement. For Marvel Rivals, NetEase does publish per-hero figures; the tier records that the first-party value has not been taken yet, not that none exists. For PUBG the wiki disclaims its own numbers and cites a datamine, and the tier records that too.
6. Unconfirmed
The source itself flags the figure as an estimate, and the corpus carries it at half confidence rather than dropping it. Astra's Cosmic Divide: a 21-second total effective duration and a 2.5-second deformation time, both from the wiki, both marked. A reader who wants a number gets one; a reader who wants a proof knows not to cite it as one.
7. Inferred
Arithmetic on stated values, not itself stated anywhere. The AK-47 fires 10 shots per second. Valve's file does not say that; it says the cycle time is 0.1 seconds, and one over 0.1 is ten. Same for its 90 rounds of reserve ammo. The corpus computes these because they are what a calculator wants, and marks them so nobody quotes Valve for a number Valve never wrote.
Why the tier matters more than the number
Two things follow once every value has a tier.
First, disagreement stops being an error. A note and a measurement that differ by a second are both kept, both dated, and the reader decides which fits the question. A bot can say "14 per the patch note, 15 measured" instead of picking one and sounding sure.
Second, an absent number stays absent. If no source at any tier states a figure, the field is null. Nothing is inferred from a similar weapon, averaged from two wikis, or remembered from an old patch. The index says it in its own words: an absent figure is absent on purpose. That rule is the difference between a corpus and a rumour with a JSON schema.
What the licences let you do
Provenance also settles what you may do with the numbers. Each game's attribution line, stated in the index and printed on the docs, names the source and its terms:
- Riot's official wikis (VALORANT, League, TFT, 2XKO, Runeterra): CC BY-SA 3.0. Attribute, share alike.
- Fandom community wikis (Fortnite, Rainbow Six, Marvel Rivals): CC-BY-SA, version unstated by the wiki, so unstated by the corpus.
- Overwatch 2 and PUBG: CC BY-NC-SA. NonCommercial. Not for a paid product without the authors' permission.
- Destiny 2: Bungie's own manifest, under a published policy that treats soliciting money as commercial and enforces against it. Non-commercial.
- PokéAPI: BSD-3-Clause, the most permissive source in the corpus, with the copyright notice reproduced as the licence requires.
- Valve and Mojang's own data (Counter-Strike 2, Minecraft): the figures are facts extracted from the publisher's files. Nothing of theirs is republished: not the files, not Mojang's generated JSON.
- Bungie's manifest (Destiny 2): read over Bungie's own API with the User-Agent its README asks for; the figures are Bungie's data, and the policy above applies.
None of the publishers is affiliated with or endorses this work; the trademarks are theirs and every attribution line says so. That sentence is part of the provenance too.
The ladder, the shapes, and the attribution lines are on the docs page, and every fact the API returns carries its rung. If you have ever quoted a wiki and been corrected by someone who read the files, that is the difference the tier was built to show.
Questions
Which tier should I trust?
Game files first, then a patch note, then in-game, then measured, in that order. Community, unconfirmed and inferred are usable with the tier shown. The corpus's resolved[] already picks the best-attested current value per field; the tier tells you what you got.
How is 'measured' actually measured?
From replay files, by timing an ability's actor from spawn to removal across many instances and taking the median. The fact states the count and the interquartile range, and the corpus states its own confidence in the extraction. Where an official figure exists for the same field, the two are compared with a stated tolerance.
Why keep unconfirmed values at all?
Because a marked estimate is more useful than a blank, and less dangerous than an unmarked one. The tier and the half confidence travel with the value, so a calculator can exclude it and a bot can say 'the wiki estimates'.
Can I build a paid product on this?
On most games, yes, with attribution. Not on Overwatch 2, PUBG or Destiny 2, whose sources are non-commercial; the docs page names them and the index entry for each game carries its licence line.
