Skip to content
Editorial

Methodology: How We Verify Every Guide

How we source facts, score confidence, handle conflicting sources, and re-verify guides within 48 hours of every game patch.

Why guides go wrong

Most game guides are written once, in the launch window, and never touched again. When the developer patches difficulty, damage numbers, or parry timings, those guides silently become wrong — and nothing on the page tells you so.

That is the problem this site exists to solve. We treat a guide as a view over a database of individually sourced facts, not as a document. When a patch invalidates a fact, we can find every page that depends on it.

Where facts come from

Nothing enters our database without a source URL, a retrieval date, and the game version it was checked against. Source type sets a baseline confidence score:

Source typeBase confidenceIn use now
Official patch notes 0.95 2
Official website / dev blog 0.92 not used yet
Official wiki or in-game data export 0.90 not used yet
First-hand play testing (recorded) 0.90 not used yet
Community wiki with edit history 0.75 73
Press coverage of official notes 0.70 5
Reddit answer with 50+ upvotes 0.60 not used yet
Video guide captions 0.55 not used yet
Scattered forum discussion 0.40 not used yet

Right now 80 sources back the fact database. The counts above are generated from it — including the tiers we do not currently have any sources for.

Confidence rises when independent domains agree: two domains add +0.05, three or more add +0.10, capped at 0.98. Nothing is ever certain enough to score 1.00. A claim below 0.60 confidence cannot carry a guide's core conclusion on its own — it needs corroboration.

When sources disagree

We do not merge conflicting claims, and we do not let a model pick the more plausible one. Contradictions are recorded as disputed, both versions are kept, and the guide says so in plain language instead of stating a number we cannot stand behind.

Every guide shows its disputed-fact count in the Fact Integrity panel near the bottom of the page.

What happens when a patch lands

  1. Detect. We monitor official patch notes and log each release with the entities it touches.
  2. Invalidate. Every fact attached to an affected entity is marked for re-check, and every guide referencing those entities is flagged.
  3. Re-verify. Every fact attached to an affected entity is re-extracted from its source and re-dated. A value that now disagrees with what we had is recorded as disputed — both readings kept, neither silently overwritten.
  4. Publish. Guides get an updated version stamp, and the patch tracker records what changed and how many guides were revised.

Our target is 48 hours from patch to re-verified guide. Until a guide has been re-checked, it carries a visible warning stating the build it was written for — we would rather show you a stale label than a confident wrong answer.

How AI is used here

We use AI to draft prose from facts that are already in the database, and to help extract structured data from source material. It is not permitted to invent numbers: generation is constrained to facts that carry a source, and output is checked against the database before publishing.

The decisions that matter are not left to a model's judgement, because a model has no reliable way to report its own uncertainty. They are enforced as rules in code, before anything is published: a claim with no source is rejected outright; a claim below 0.60 confidence cannot carry a guide's conclusion on its own; two sources that disagree are both kept and marked disputed rather than reconciled; any name or number in the draft that is not traceable to the fact database blocks publication. A guide that fails these checks is not downgraded and shipped — it is not shipped.

What we do not do

We do not play these games. Nothing on this site is first-hand play testing, and we have no recorded footage of our own. Every claim is drawn from a published source, and every one of those sources is listed — per claim, with its date and the game build it applied to — at the bottom of the guide it supports.

We think saying so plainly is worth more than the alternative. A guide that quietly implies someone beat the boss themselves is asking for trust it cannot back up; one that shows you exactly which source each number came from lets you check it in a click. The second is the only kind we can honestly offer, so it is the one we built for.

Corrections

If a guide is wrong, we want to know. Every page lists its sources so you can check our work. Corrections are applied to the underlying fact, which means they propagate to every guide that depends on it, not just the page you were reading.