Current AI Analysis Method
This page documents the implemented inputs, formula, and pre-publication evidence checks. External data that was not retrieved is not treated as evidence.
The unusual-review signal is the average of five published rules over retrieved Steam reviews. Model-returned numeric judgments are not blended into it. AI input is limited to 40 representative reviews, and core claims are checked against exact source quotes.
โ๏ธ Current Implementation Summary
| Item | Current processing | Limit |
|---|---|---|
| AI input | Up to 40 representative reviews | Not the entire retrieved corpus |
| Core claims | Review IDs and exact quotes are checked | Failed results are not published |
| External comparison | Metacritic / OpenCritic | Used only after successful retrieval |
| Optional analysis | Accepted only with required text or metadata | Missing data is not filled by inference |
| Display languages | Japanese, English, Korean, Chinese | Cultural nuance can still be missed |
๐ก Data Sources
The primary input is review text that was successfully retrieved. External scores are used only when retrieval succeeds.
For Steam titles, retrieved review text, verdict, playtime, and helpful votes provide the primary evidence.
For non-Steam titles, successfully retrieved reviews form the primary corpus. The analysis is rejected when that corpus cannot be retrieved.
Used for comparison only when retrieval succeeds and usable values exist. They are not added to the unusual-posting signal formula.
๐ Unusual-Review Signal Calculation
The published number is the retrieved-corpus average of five rules. AI text analysis separately produces the rank, summary, and evidence and cannot overwrite this number.
Five rule indicators are applied to each review in the retrieved corpus, then averaged.
After removing duplicates and overly short text, up to 40 representative reviews are selected with positive/negative balance and sent to Gemini.
This is a 0โ100 heuristic score, not a percentage, probability of manipulation, or proof of fraud.
๐ 6-Axis Quality Assessment
The six 0โ100 axes are estimates from the representative review excerpts. An axis with little review coverage is more uncertain and is not a measurement.
Visual quality, art style, framerate stability
Controls, mechanics depth, replayability
Narrative quality, world-building, characters
Content amount vs price point
Playtime, replayability, endgame content
Bug frequency, crash rate, optimization
๐ Overall Rank System
Several estimated axes are high in the representative sample. Check cited evidence and personal fit before buying.
Most axes are high, but opinions outside the retrieved sample may be absent.
The sample contains both clear strengths and concerns worth checking.
Strengths and weaknesses are roughly balanced in the sample.
The representative reviews contain several material concerns. Check the cited reviews.
๐งฌ Special Analysis Modules
evaluation_gapEstimates differences within the retrieved sample only when enough Japanese and English reviews are present.
ai_generated_suspicionSummarizes repetition or formulaic wording in cited text. It does not classify authorship or generation method, and uncalibrated percentages or confidence labels are not published.
spec_analysisOrganizes concrete GPU, FPS, and fault reports only when they occur in the input reviews.
review_contradictionsShows verdict/text contradictions only when an exact supporting quote is available.
โ๏ธ AI Model & Technical Specs
โ ๏ธ Limitations & Disclaimers
โ Estimates only:The unusual-posting signal, hardware-related estimates, and six-axis scores are derived from retrieved data, rules, and/or AI. The signal is not a percentage or proof of wrongdoing, and none of these values is guaranteed accurate.
โก Review sampling:AI input is limited to 40 representative reviews. When the retrieved corpus is larger, omitted opinions may not be reflected in the result.
โข AI limitations:Gemini excels at context but may not fully understand sarcasm, jokes, or cultural nuances.
โฃ Point-in-time and external data:Analysis reflects retrieval and cache time. Metacritic / OpenCritic comparisons and optional modules appear only when their required data was retrieved.
โ FAQ
Five rules are averaged over the retrieved review corpus. Up to 40 representative reviews are sent to AI for the separate rank and summary, but model-supplied numbers are not added to this signal.
Overall reasoning, pros, and cons must bind to valid input review IDs and exact matching source quotes. Unknown IDs, quote mismatches, and insufficient evidence are rejected before publication.
Graphics, Gameplay, Story, Value, Volume, and Stability are estimated from representative reviews. Axes with sparse coverage are uncertain rather than measured facts.
No. Metacritic / OpenCritic comparisons run only after successful retrieval, and each optional module requires the relevant review text or metadata.
๐ Methodology Update Record
Changes that affect the mapping between public explanations and implementation are recorded here.
| Date | Change | Impact |
|---|---|---|
| 2026-07-20 | Removed unacquired sources and unsupported forecast fields from public explanations | Accuracy |
| 2026-07-20 | Documented the 40-review maximum and exact-quote claim validation | Transparency |
| 2026-07-20 | Limited external comparisons and optional modules to available required data | Grounding |