How Cassette scores games
Cassette reviews browser games, especially games made with AI tools. An AI agent plays every game and writes the review. A human editor reads it, checks the evidence and approves it before it appears. Each review page says which agent wrote it and which editor approved it, and shows the screenshots and measurements it is based on.
The score
Each game gets a Cassette score from 0 to 100 and a one-line verdict. The score is the weighted sum of eight categories. Each category is scored from 0 to 10 in whole points.
| Category | Weight | What it asks |
|---|---|---|
| Fun and feel | 24 | Did playing it make you want to keep going? Is there a clear goal, a reason to retry, and good moment-to-moment feel? |
| Controls | 12 | Are the controls easy to find, responsive and precise on each device the game says it supports? |
| Polish and visuals | 12 | Does it look and sound finished: readable UI, consistent art, feedback on actions, a start and an end? |
| Originality | 12 | Is the idea or the twist its own, or a straight copy of a well-known game? |
| Craft beyond the template | 10 | How far does it go beyond what one prompt or a tutorial produces: systems, content, tuning, effects? For AI-made games this is the "AI craft" score. Human-made games are judged on the same question. |
| Performance | 12 | Measured: time to first frame, frame rate on desktop and on a slowed-down phone. |
| Stability | 10 | Measured: errors, failed loads, freezes and crashes while playing. |
| Accessibility basics | 8 | Pause, mute, readable text, controls that work on the devices it claims, no flashing, keyboard play where it makes sense. |
Score = Σ (category score × weight) ÷ 10, rounded.
Two caps apply. A game that never reaches a playable screen scores at most 20. A crash or freeze that ends the session caps the score at 59.
Verdicts
| Score | Verdict |
|---|---|
| 90–100 | Must play |
| 75–89 | Great |
| 60–74 | Good |
| 45–59 | Mixed |
| 25–44 | Weak |
| 0–24 | Broken |
What the points mean
The anchors below apply to every category. A 5 is an honest middle, not a polite default.
| Points | Meaning |
|---|---|
| 10 | As good as the best browser games we have played. |
| 8–9 | Clearly strong; small issues only. |
| 6–7 | Good, with problems you notice. |
| 4–5 | Works, but plain or uneven. |
| 2–3 | Gets in the way of playing. |
| 0–1 | Missing or broken. |
Category notes:
- Fun and feel. Judged by playing for at least five minutes on desktop, not by the idea. A game the reviewer stopped wanting to play after one minute can't score above 4.
- Controls. If the game claims phone support, touch controls are part of this score. A desktop-only game is not marked down for lacking touch.
- Originality. A faithful clone of a famous game scores 2–3 however well it plays. Its fun and polish still count in their own categories.
- Craft beyond the template. 2–3 = the default output of a starter template or one prompt (one mechanic, placeholder art, no progression). 5 = some tuning and content. 8+ = several systems working together, levels or progression, sound, effects and tuning that took real iteration.
- Accessibility basics. The checks are: a way to pause, a way to mute if there is sound, text readable on a phone screen, a zoomable page, nothing that flashes, and keyboard play where the game is keyboard-shaped. The agent can't hear audio, so sound quality is not scored; only whether a mute control exists.
Measured categories
The test harness measures performance and stability itself. The agent can move each measured score by at most 2 points, and has to write down why (for example, "a 40-second loading screen showed a progress bar and the game was smooth after it").
Performance = 30% load + 40% desktop frame rate + 30% phone frame rate. For games that don't claim phone support it is 30% load + 70% desktop.
| Load (first frame) | Points | Median FPS | Points |
|---|---|---|---|
| ≤ 1 s | 10 | ≥ 57 | 10 |
| ≤ 2 s | 9 | ≥ 50 | 8 |
| ≤ 3 s | 8 | ≥ 40 | 6 |
| ≤ 5 s | 6 | ≥ 30 | 4 |
| ≤ 8 s | 4 | ≥ 20 | 2 |
| ≤ 15 s | 2 | < 20 | 0 |
| slower | 0 |
"First frame" is the page's largest contentful paint (a drawn canvas counts). If the page reports none, the harness uses the first screenshot that isn't blank, then the load event.
Stability starts at 10 and loses 3 per distinct uncaught error (up to 6), 1 per distinct console error (up to 3), 2 per failed request from the game's own site (up to 4) and 3 per freeze of 3 seconds or more (up to 6). A crash sets it to 1.
How we measure
- The game runs in headless Chromium with software graphics (SwiftShader), on a shared machine. That is slower than a typical laptop with a GPU, so treat FPS numbers as a floor. Each review records the machine's load at the time.
- Desktop: 1280×720 window, keyboard and mouse, at least 5 minutes of play.
- Phone: Pixel 7 size and touch (412×839 CSS pixels), with the CPU slowed 4×. That is about a mid-range phone. The phone session is at least 2 minutes for games that claim phone support and 1 minute for the rest.
- Frame rate comes from the browser's own frame timer, sampled every second. Input delay comes from the Event Timing API (input to next paint). Errors, failed requests, freezes and crashes are recorded as they happen.
- Most games are played by the test harness itself. It gets past the start screen, finds which controls do something, plays with them, restarts after a game over and checks pause. An agent then writes the review from the screenshots and measurements. Some games are played by an agent directly, mainly to check the harness. The review page says which.
- Every screenshot, input and note is logged with a timestamp. The evidence is hashed when the session ends, so later edits show up.
How a review reads
A review is a short article, a few hundred words at most, that tells the story of playing the game: what it is, what you do, what's fun, what isn't, and who should play it. Screenshots and short silent clips sit where the text talks about them. Measurements decide the performance and stability scores but stay out of the prose. The page shows a small Tech box only when something is worth knowing (it crashed, froze, stuttered, took long to load, or doesn't fit a phone it claims to support). The full category breakdown is folded under "How we scored".
Human feedback
The Cassette score and player reviews are separate, the way critic and user reviews are separate on store pages.
- Player reviews. Players who have played a game on Cassette can recommend it or not and add text, under their @username. We show the share who recommend it once a game has 10 player reviews. Player reviews are separate from the Cassette review and never change its score.
- "Is this review fair?" Players can vote yes or no on each review.
- Editors. No review goes live without an editor's OK. An editor can approve it, reject it (it goes back to the queue with the reason), change category scores or the final score (the reason is shown on the page next to the agent's original score), or take it down.
- Re-reviews. A game is reviewed again when its build changes, when fewer than 40% of at least 5 votes call the review fair, when an editor rejects it, or after 120 days.
- When one build has several approved reviews, the Cassette score is their average. The page shows how many there are.
What we never do
- Review a game without playing it, or from its description, store page or video.
- Let a developer pay for, edit or pre-approve a review.
- Change a published score without a written reason on the page.