Game Review

Judge game work against what actually plays, not what the prose claims. A review names the claim under review, checks the evidence behind it first, then applies the craft lenses the claim touches, and returns one sign-off with the artifacts it checked.

Evidence Before Opinion

Start every review by locating the claim and its evidence:

1. State the claim under review in one sentence (readable feel, taught level,

covered branches, playable build, release-ready).

1. Check what exists: build, capture, playtest note, or a check you can run

yourself. Read references/evidence-and-release.md for evidence categories and the verification flow.

1. Separate missing evidence from a failed check — they demand different

sign-offs and different next steps.

Mechanical evidence (files exist, project loads, tests pass) never stands in for playable build (a player performs the action and sees the consequence). For engine build claims, require a real run — launch, scene load, smoke check, capture — not file presence; the game-production skill carries engine-specific check ladders if installed.

Record evidence provenance as direct play, inspected video, inspected screenshot, user-reported description, or mechanical check. A description of a capture is reported evidence, not a capture you inspected. Timing, control, camera motion, and audio claims need direct play or inspected video; a screenshot can support only static layout or state.

Choose Lenses By Claim

Read only the references the claim touches:

Independent lenses may run as parallel workers when the harness supports it; running them sequentially in one pass is equally valid. Either way, keep each lens's findings distinct. Give each worker the same claim, evidence provenance, no-edit boundary, required checked artifact, finding and confidence fields, and stop condition; synthesize their findings instead of concatenating reports.

Sign-off

Return exactly one of:

• READY — the checked evidence supports advancement; the review does not

advance the project stage.

• CONCERNS — it can advance only if the listed risks are accepted or fixed.
• NOT_READY — advancing would create false confidence, hide a missing core

loop, or expand a public promise past the build.

• USERDECISIONNEEDED — the review itself cannot finish without a

human-owned taste, concept, direction, or public-promise decision. A separate human stage decision after a supported READY sign-off stays in the output field below; it does not replace the evidence sign-off.

Do not average lenses or simulate consensus: preserve dissent, and let the strictest finding that is supported by checked evidence set the sign-off. Separate observation, quote, inference, severity, and confidence in qualitative findings; style notes must not bury blockers.

Authority

A review or diagnosis authorizes inspection and non-destructive checks only. Do not edit the game, advance its stage, or change a public claim. If the user also asks for fixes, return the sign-off first and hand the accepted findings to game-production when it is available. Otherwise stop with a bounded fix handoff; review alone never becomes implementation authority.

Output

Claim under review:
Sign-off: READY | CONCERNS | NOT_READY | USER_DECISION_NEEDED
Evidence checked:
Evidence provenance:
Missing evidence / failed check:
Findings (per lens, strictest first):
Checks failed, skipped, or unverified:
Residual risk:
Required next build:
Human decision needed:

Gotchas

• Polish can hide missing recovery: juice and visual quality do not prove the

player can read failure and get back to play.

• A level that tests before it teaches fails review even when it is beatable.
• Branch and quest claims need coverage and collapse rules, not sample

transcripts.

• Screenshots do not prove gameplay; passing tests do not prove feel; a filled

template or confident prose is not quality.

• Do not approve a vertical slice before the core loop is proven repeatable

with readable success, failure, and recovery.

• Playtest notes without a build identifier are weak evidence; say so rather

than treating them as ground truth.