Definitions, calculation methods, and limitations for the main terms used by Game Evaluation Lab. Detailed analysis methods are also public.
A 0–100 warning metric averaging five published rules over retrieved Steam reviews. It is not a percentage, probability of manipulation, or proof of misconduct. Model-generated numeric judgments are not blended in, and a rule is skipped when its input is unavailable.
A separate corpus metric combining review quality (30%), inverse manipulation risk (30%), inverse review-bomb risk (20%), sentiment stability (10%), and source diversity (10%). It is not 100 minus the unusual-review signal and does not certify that a purchase is safe.
A band derived from the simple average of six scores estimated from representative reviews: S at 90+, down to F below 40. It is a sample-bound estimate, not measured product quality or a purchase guarantee. No rank is published when the six axes are incomplete.
Six 0–100 estimates—Graphics, Gameplay, Story, Value, Volume, and Stability—derived from representative review text. Axes with sparse mentions are more uncertain, and the values are not benchmarks or hands-on measurements.
A per-review value capped at 100: +30 for under one hour of playtime, +20 for text under 50 characters, +15 for fewer than five owned games, +25 for over 100 posted reviews, and +10 for zero helpful votes. The public unusual-review signal is its average over retrieved reviews.
A process that estimates six axes, reasoning, pros, and cons from up to 40 representative reviews after removing duplicates and extremely short text. Core claims are checked against input review IDs and exact quotes; failed outputs are not published. It does not determine whether an author is human.
A caution signal based on short-term negative-review concentration in retrieved data or changes between stored rating snapshots. It cannot automatically determine cause, coordination, or relevance to the game, and does not confirm a review-bomb event.