Home / Game Walkthrough / How Roblox Evaluates Build-Made Games: The Four-Signal Scoring System Explained
Game Walkthrough

How Roblox Evaluates Build-Made Games: The Four-Signal Scoring System Explained

Screenshots of three Roblox games created with the Build AI tool
Screenshots of three Roblox games created with the Build AI tool
Quick Summary

Roblox's October 8 engineering writeup reveals exactly how it scores games made with Build, its AI game-creation tool: four separate signals covering spec quality, engineering, design, playtesting and visual polish. We break down the framework and the real worked example Roblox published, a Build-made obby that scored 98/100 on engineering but just 31/100 on design.

On October 8, 2026, Roblox's engineering team published a detailed look at how it grades the games its AI tool Build makes for players. Build turns a text prompt into a playable Roblox experience, and since its New Zealand alpha launched, more than 100,000 players have tried it and roughly 9,000 Build-made games have been published — 71% of those creators had never opened Roblox Studio before. The writeup explains the scoring system Roblox built to judge whether those AI-generated games are actually any good, and it is far more specific than the usual "we're improving our AI" post: real rubrics, real weighted scores, and one game graded end to end in public.

This matters beyond the engineering blog. Every creator using Build — and every player who stumbles into a Build-made game — is interacting with the output of this exact grading system. Here is how it works.

1. Why Grading an AI-Made Game Is Hard

Roblox's own framing is that a game is not a static document. It is an executable, real-time system — qualities like responsiveness, pacing, and "game feel" only show up once a player is actually moving through it. You cannot judge whether a game is fun by reading its code, and you cannot judge whether its code is sound just by watching someone play it. Roblox's evaluators argue a Build-made game has to pass two separate bars: it has to be low-level correct (no broken systems, no networking bugs) and it has to feel finished rather than like an unpolished prototype.

Diagram of Roblox's four-stage Build evaluation pipeline
Roblox's four-stage evaluation pipeline for Build: a spec check at the planning stage, a static check once the place file exists, a playtest pass-rate check once it can be played, and an experiential check against captured gameplay video. Image: Roblox

2. The Four Signals Roblox Actually Scores

Build's output is measured against four separate signals, deliberately kept apart rather than blended into one number:

  • Game specification quality (0–100): Before a single asset is built, Roblox checks whether the generated plan makes the key design decisions explicit, preserves what the player actually asked for, is technically feasible inside Roblox's runtime, and doesn't quietly invent requirements nobody requested.
  • Static game quality (0–100): Once a place file exists, it's inspected in Studio's edit mode through two separate lenses — an engineering lens (are systems correctly wired, is the client/server boundary sound, does the UI actually work) and a design lens (how deep is the core loop, how much content variety exists, is there real progression and replay value, does any of it feel original).
  • Playtest pass rate (0–100%): An automated agent plays the finished game using player-like inputs — taps, clicks, keystrokes, navigation — against a per-game test plan. Each test item is a binary pass or fail, weighted by how important it is, and the whole playtest is repeated multiple times to check the results are consistent rather than lucky.
  • Visual and experiential quality (0–100): Captured gameplay footage, shot across several device screen sizes, is scored on asset craft, UI polish, motion, lighting, and audio.

Roblox also logs process metrics — how many tokens the AI used, how many tool calls it made, how long the build took — but these don't factor into the score itself. They're there to catch regressions, like a model update that suddenly becomes far more expensive to run without actually improving anything.

3. A Real Example: Grading "Skyforge Ascent"

Rather than keep this abstract, Roblox walked through a real Build-made game end to end: Skyforge Ascent, a third-person 3D obstacle course where the player runs and jumps across floating platforms, dodges moving hazards like rotating sweep bars and sliding blockers, and reaches checkpoints on the way to a finish arch.

Four screenshots of Skyforge Ascent showing its start screen, checkpoint 1, checkpoint 5, and the finish screen
Skyforge Ascent, the Build-made obby Roblox used as its public worked example — shown here at its start screen, checkpoint 1, checkpoint 5, and its "Ascent Complete" finish screen. Image: Roblox

The scores Roblox published for it were deliberately uneven, to make a point:

  • Spec score: 91/100 — docked points for missing numeric constants and a few acceptance criteria that weren't testable as written.
  • Engineering lens: 98/100 — the systems were correctly built and wired together.
  • Design lens: 31/100 — the core loop was judged simple and linear, with little progression, replay incentive, or original hook.
  • Playtest: 100/100 across nine test items, with zero runtime errors.
  • Visual quality: 86/100, scored on a phone-sized viewport.
Bar chart comparing Skyforge Ascent's 98.0 engineering score to its 31.5 content quality score
The gap Roblox's two-lens static check is designed to expose: Skyforge Ascent scored 98.0/100 on generic engineering correctness but only 31.5/100 on design and content depth. Image: Roblox

The build itself took about eight minutes with zero tool errors. Roblox's point in publishing this specific example is that a single blended score would have hidden the gap entirely — a 90-ish average would have looked fine, while the real story is a technically flawless game with a design that needs more work. "Code quality and content quality are fundamentally different dimensions," the engineering team wrote.

4. How the Scores Are Already Changing Build

Roblox says this framework isn't just a report card — it's been steering actual changes to Build since the New Zealand alpha:

  • Early alpha fix: testers complained Build sometimes produced a 2D game when they'd asked for 3D. Using the eval signals to isolate the problem, the team improved 3D intent understanding, object positioning, and lighting within two weeks.
  • Model benchmarking: every candidate model upgrade is run through offline A/B comparisons across all four signals before it ships. One base-model upgrade improved every signal at once. A separate attempt — a dual-agent setup using a lightweight sub-task model — kept content depth about the same but dropped playtest score by 11.2 points and visual score by 13.2 points, and was caught before shipping.
Table comparing a production model to a candidate model across game spec, static, playtest and visual judge scores
Roblox's internal offline comparison format: a candidate model update is run against the current production model across all four signals before it's allowed to ship. Image: Roblox

Roblox also shared where the numbers currently stand by game type. An early set of 2D games averaged a 75.9 playtest score, 67.1 visual score, and 19.2 content-depth score. A more recent set of 3D games averages 51.5 playtest, 66.3 visual, and 33.9 content — meaning 3D games are harder to make bug-free right now, even though their content depth has improved. Genre matters too: mechanically heavy genres like RPGs, shooters, and racing games tend to produce more playability issues, while simpler-logic genres like survival and tycoon games pass playtests more reliably, which shifts their main weakness toward visual polish and UI instead.

5. What This Means If You're Building (or Playing) With Build

For creators using Build — including the 71% who are trying game creation on Roblox for the first time through it — the practical takeaway is that the tool is being measured, and improved, on design depth specifically, not just whether the game runs without crashing. Roblox's own data shows engineering correctness is already strong; design and content depth is the part still catching up, and it's the part the team says it's actively working to raise.

For players, it explains something you may have already noticed: Build-made games tend to run cleanly, but can feel thin on replay value compared to games built by hand in Roblox's other creator tools, several of which were also covered in depth at RDC 2026. Roblox has said it plans to broaden this evaluation system to cover more genres, more complexity tiers, and multiplayer games, and to improve how it scores multi-turn editing as creators refine a Build project over several prompts rather than one.

6. Quick Recap

  • Build has produced roughly 9,000 published games since its alpha began, from over 100,000 players who've tried it.
  • Every Build game is scored on four separate signals: spec quality, static (engineering + design) quality, playtest pass rate, and visual/experiential quality — kept separate on purpose.
  • Roblox's own worked example, Skyforge Ascent, scored 98/100 on engineering but just 31/100 on design — illustrating exactly why the scores aren't blended into one number.
  • The eval system has already driven real fixes, including a 3D-intent fix shipped within two weeks of being identified in alpha.
  • 3D Build games currently average lower playtest scores than 2D ones; simpler genres like tycoon and survival currently score better on playability than mechanically complex ones like RPGs and shooters.

7. Official Resources

For more on Roblox's creator-side tools and recent platform changes, see our coverage of the Build, Scene Generator and 2D Engine update, the Creator Hub changes around game page videos, or browse our full walkthrough archive.

Ahmed Hassan
Lead guide writer at Walkthroughs.net. Publishes daily solutions for hidden-object adventures, escape rooms, and word puzzles.