PaperlayerView exact revision on arXiv

Source-linked AI summary

RSIGame: Autonomous Agentic Game Development with Recursive Self-improvement

Wenyi Wu, Minghao Fu, Jieyu You, Kun Zhou, Siqi Liu, Aayush Salvi, Yiheng Lin, Ce Zhang, Xiaohan Lan, Jiahui Zhu, Yujie Zhong, Qi She, Biwei Huang

arXiv:2609.39045v1cs.CLcs.GTcs.LGcs.MA

TL;DR

Iterative game refinement can overfit limited tests, leaving defects and weak generalization to broader player interactions. RSIGAME combines evidence-grounded local revision, global checkpoint and progress control, and training-based experience internalization. Across 140 tasks, two engines, and five generators, it improves quality under matched budgets, with internalized Qwen3.8-27B reaching 61.38 on Godot versus 50.26 for Codex+GPT-5.5 one-shot.

  • Problem

    Iterative game refinement can overfit a small set of tests, leaving bugs, missing behaviors, and corner cases unresolved and weakening generalization to broader player interactions.

    Source: §1 · p. 2

  • Method

    RSIGAME uses local explore-diagnose-improve and global quality-control loops, then fine-tunes the generator on successful planning, generation, and improvement experience.

    Source: §1 · p. 2, §3 · p. 3, §3.2 · p. 5

  • Results

    Across 140 GameCraft-Bench tasks, two engines, and five generators, RSIGAME improves quality under matched budgets; Qwen3.8-27B reaches 61.38 on Godot versus 50.26 for Codex+GPT-5.5 one-shot.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Internalized development experience can strengthen future game generation beyond test-time refinement.

    Source: §1 · p. 2, §7 · p. 9

    The study focuses on general-purpose game-generation and development tasks and excludes unsafe or sensitive artifacts from evaluation and data curation.

    Source: ETHICS STATEMENT · p. 10

Abstract

from arXiv · show

Recent advances in large language models have made automatic game generation increasingly feasible, yet reliably improving generated games beyond a playable version remains challenging. Naive iterative refinement can easily overfit a small set of test cases, producing fragile games with unresolved bugs, missing behaviors, and poor generalization to broader player interactions. We introduce RSIGame, an autonomous agentic game development framework with recursive self-improvement. RSIGame organizes development into complementary local and global loops. Concretely, a local explore-diagnose-improve loop broadly explores the executable game, diagnoses and prioritizes discovered issues, and performs evidence-grounded revision, where an evolving checklist continually accumulates new testing and improvement guidance. A global loop tracks overall quality, preserves the best checkpoint, and detects saturation or regression over long-horizon development. Beyond test-time improvement, RSIGame further internalizes successful development experience into the generator through training. Across 140 GameCraft-Bench tasks, two game engines, and five generators, RSIGame consistently improves game quality under matched development budgets. Notably, experience internalization enables Qwen3.8-27B to reach 61.38 on Godot and 58.53 on Phaser, exceeding GPT-5.5 one-shot scores while reducing Qwen's generation tokens by 11 times.

1 INTRODUCTION

Iterative game development can improve an initial game, but narrow testing can leave bugs and missing behaviors unresolved. RSIGAME broadens testing and organizes revision around accumulated evidence, while global monitoring and experience internalization support long-horizon improvement.

1 INTRODUCTION

  • Iterative refinement can overfit a small test set, leaving bugs, missing behaviors, and corner cases that impair generalization to broader player interactions.RSIGAME addresses this by broadly exploring the executable game, diagnosing and prioritizing issues, and carrying new testing and revision guidance in an evolving checklist.Source: §1 · p. 2
  • RSIGAME combines a local explore-diagnose-improve loop with global quality monitoring, best-checkpoint preservation, saturation and regression detection, and training-based experience internalization.The local loop revises from accumulated evidence; training converts successful trajectories, diagnostic decisions, and verified revisions into generator supervision.Source: §1 · p. 2
  • Across 140 GameCraft-Bench tasks, two engines, and five generators, RSIGAME improves game quality under matched budgets; Qwen3.8-27B reaches 61.38 on Godot versus 50.26 for Codex+GPT-5.5 one-shot.The passage also reports an 11-times reduction in Qwen's generation tokens.Source: §1 · p. 2

2 PRELIMINARIES

Agentic game development turns a natural-language specification into an executable project whose code, assets, and configuration must work together. RSIGAME applies autonomous iterative testing and revision to that process.

2 PRELIMINARIES

  • Given a natural-language game specification, an agent aims to produce a complete, playable game that follows the intended design and executes correctly.The project includes source code, visual and textual assets, and configuration files, which must be integrated into a coherent interactive system.Source: §2 · pp. 2–3
  • Game quality is assessed through runtime behavior and player experience, using replayable demonstration traces scored against a hidden task-specific rubric.The traces are replayed during evaluation.Source: §2 · p. 3
  • RSIGAME replaces human-driven testing and revision with an autonomous cycle that executes the current project, diagnoses observed failures, and revises it.The paper frames this as iterative development from an initial project through successive versions.Source: §2 · p. 3, Figure 2 · p. 3

3 RSIGAME

RSIGAME combines training-free iterative game development with training-based internalization of development experience. Its local loop uses observed gameplay and an evolving checklist to guide revisions, while its global loop preserves the best game state and controls continued development.

3 RSIGAME

  • RSIGAME combines training-free local and global loops with training-based internalization of accumulated development experience into the backbone game-generation model.The framework treats iterative development and generator training as complementary strategies.Source: §3 · p. 3

3.1 TRAINING-FREE RECURSIVE SELF-IMPROVEMENT

  • The training-free strategy pairs a local explore-diagnose-improve loop with a global loop that monitors long-horizon progress, preserves the best state, and detects saturation.The local loop explores the executable game, diagnoses issues, and refines it across iterations.Source: §3.1 · p. 3

3.1.1 LOCAL LOOP: EXPLORE-DIAGNOSE-IMPROVE

  • The local loop uses a controller to choose an exploration direction and an explorer to gather gameplay evidence, then an editor proposes changes and a verifier checks fixes and side effects.An evolving checklist records issues, priorities, and verification outcomes so subsequent rounds build on prior evidence.Source: §3.1.1 · p. 4

3.1.2 GLOBAL LOOP: PROGRESS MONITORING AND CONTROL

  • The global loop tracks whole-game quality and preserves the strongest checkpoint; when progress stalls, development can stop or continue under new high-level guidance.The quality monitor is distinct from the local verifier, which checks whether a particular edit resolves its target issue.Source: §3.1.2 · pp. 4–5

3.2 TRAINING-BASED KNOWLEDGE INTERNALIZATION

  • RSIGAME internalizes planning, generation, and verified improvement experience through supervised fine-tuning, transferring intermediate development decisions into model parameters.The curated data includes executable generations, successful planning traces, and improvements verified through post-edit interaction.Source: §3.2 · p. 5

4 EXPERIMENTS

Across matched-budget evaluations, RSIGAME improves game quality on Godot and Phaser, with experience internalization further strengthening generation. The experiments use 140 GameCraft-Bench tasks and report development-time scaling and global-control results alongside the engine-specific comparisons.

4.1 EXPERIMENTAL SETUP

  • Evaluation covers 140 GameCraft-Bench tasks in Godot and Phaser, with replayable traces scored on Mechanics, Depth, Visuals, and Art against a hidden rubric.Methods start from identical cloned base projects and receive matched per-round tool budgets; Table 1 reports results averaged over the tasks.Source: §4.1 · pp. 5–6, Table 1 · p. 6

4.2 GODOT GAME RESULTS

  • Across the five Godot settings, RSIGAME improves frozen initial games by 10.7–20.3 Overall points and outperforms Play2Code by 7.2–13.8 Overall points under matched budgets.For Qwen3.8-27B, experience internalization raises the one-shot score from 37.07 to 48.22, then RSIGAME raises it to 61.38.Source: Table 1 · p. 6, §4.2 · pp. 6–7
  • As development budgets grow, RSIGAME sustains improvement across engines and initialization strengths, while Play2Code can plateau or regress; retaining the best checkpoint stabilizes progress.Saturation-aware stopping achieves comparable quality to longer fixed-budget runs with fewer rounds.Source: Figure 3 · p. 7, Figure 4 · p. 7

4.3 PHASER GAME RESULTS

  • Across Phaser's three generators, RSIGAME improves frozen bases by 8.8–13.4 Overall points and achieves the strongest final quality in every setting.The passage attributes Play2Code's stronger Phaser results relative to Godot to more readily improvable functional headroom.Source: §4.3 · p. 7

5 FURTHER ANALYSIS

Further analysis finds that global monitoring stabilizes development-time gains, while adaptive effort allocation and agentic verification improve local development. Training on planning and improvement experience also produces gains across game-quality dimensions before test-time development.

Can Global Quality Monitoring Improve Development-Time Scaling?

  • RSIGAME consistently converts additional development rounds into higher quality across both engines and initialization strengths, while Play2Code often plateaus or regresses.The Global Quality Monitor stabilizes the volatile gains of the local loop by retaining the best state reached so far.Source: §5 · p. 7
  • A retained checkpoint tracks the oracle best within each budget, and saturation-aware stopping matches longer fixed-budget runs with fewer rounds.Together, these results indicate more reliable and compute-efficient development-time scaling.Source: §5 · p. 8

Does Adaptive Evolution Matter?

  • Adaptive development is preferred in 59% of blind pairwise comparisons versus 37% for round-robin, with advantages across all four quality dimensions.It shifts effort toward improvement after build failures and toward art when visual quality lags.Source: §5 · p. 8, Figure 5 · p. 8

Does Agentic Verification Make Local Improvement More Reliable?

  • Replay-based post-edit verification detects 76.2% of unsuccessful improvements and reaches 84.4% balanced accuracy, while build-only checking detects none of these behavioral failures.Before editing, evidence-grounded verification raises grounded precision from 58.6% to 72.3% and reduces unsupported targets from 1.93 to 0.50 per round.Source: §5 · p. 8, Table 2 · p. 8

Can Development Experience Transfer to Future Generation?

  • Training that incorporates planning and verified improvement experience yields consistent gains across Mechanics, Depth, Visuals, and Art before test-time development.The reported gains include Mechanics +14.9 and Depth +14.1; the supplied passage truncates the subsequent overall-score comparison.Source: §5 · pp. 8–9

6 RELATED WORK

Related benchmarks score complete executable games but generally evaluate a single submitted build, leaving iterative improvement across development rounds open. RSIGAME builds on iterative development while emphasizing reliable long-horizon improvement, independent verification, and transfer of development experience to future generation.

Game Generation and Development Benchmarks

  • Game-generation benchmarks evaluate complete executable games, but the cited benchmarks score a single submitted build and leave improvement across development rounds open.OpenGame-Bench covers web games, GameCraft-Bench engine-based games, and GameDevBench larger codebases and assets.Source: §6 · p. 9

Iterative Agentic Game Development

  • Play2Code, VibeGame, and OpenGame pursue iterative development, while RSIGAME focuses on sustaining it reliably over long horizons rather than introducing iteration itself.RSIGAME combines adaptive, evidence-grounded local improvement with global best-checkpoint tracking and saturation detection; high-level guidance can reopen development after progress stalls.Source: §6 · p. 9

Verification and Learning from Development Experience

  • RSIGAME independently replays updated games to verify intended improvements and detect regressions, then writes verified outcomes into shared development state.It jointly internalizes planning, generation, and verified improvement experience to strengthen future game generation and development.Source: §6 · p. 9

7 CONCLUSION

The conclusion presents RSIGAME as a recursive self-improvement framework combining evidence-grounded local revision, global progress monitoring, and training-based experience internalization. The appendix overview and reproducibility statement point readers to implementation, protocol, and evaluation artifacts, while specifying study scope and model-use disclosures.

7 CONCLUSION

  • RSIGAME combines broad exploration, structured diagnosis, prioritized improvement, an evolving checklist, evidence-grounded local revision, and global monitoring that preserves checkpoints and detects saturation or regression.Training-based experience internalization transfers successful development trajectories into the generator for future game creation beyond test-time refinement.Source: §7 · p. 9
  • Across 140 GameCraft-Bench tasks, two engines, and five generators, RSIGAME consistently improves quality under matched development budgets, enabling smaller open models to reach or surpass stronger generators after iterative development..Source: §7 · p. 9
  • The authors disclose GPT-5.5, GLM-5.3-Flash, and Qwen3.8-27B roles in generation, improvement, evaluation, and internalization, plus Claude Opus 5 for an independent verification audit.They also report using LLM assistants for editing, presentation refinement, and programming assistance, and state that authors reviewed artifacts and verified reported claims.Source: AI USE STATEMENT · p. 10
  • The study focuses on general-purpose game-generation and development tasks, excludes unsafe or sensitive artifacts, and uses no private user data, personal information, or user-generated content.Human involvement is limited to evaluation and high-level development guidance; participant attributes are not studied.Source: ETHICS STATEMENT · p. 10
  • The reproducibility statement says results are traceable to retained artifacts and documented protocols, with Table 3 indexing artifact contents, release status, and protocol locations.The implementation and evaluation pipeline are available, while some project and training artifacts await approval for release; evaluation artifacts include 51,644 scoring files.Source: REPRODUCIBILITY STATEMENT · p. 10
  • Controlled comparisons freeze the initial project and give each method an identical clone, improvement backbone, and per-round tool budget; each resulting game is replayed and scored three times.Development agents do not see benchmark rubrics, scores, or judge outputs.Source: REPRODUCIBILITY STATEMENT · p. 10

Appendix Overview

  • The appendix provides implementation details, experimental protocols, training specifications, and extended analyses, with overview entries pointing to pipeline scope, evaluated dimensions, and project size..Source: Appendix Overview · p. 14

A REPRODUCIBILITY INDEX

Table 3 serves as a reproducibility index, mapping artifacts to their contents, release status, and the locations of related implementation or protocol details. Released files are screened to remove identifying information and are re-parsed and re-scanned.

A REPRODUCIBILITY INDEX

  • Table 3 indexes artifacts by contents, release status, and locations of corresponding implementation or protocol details.It is the single index referenced by the Reproducibility Statement.Source: A REPRODUCIBILITY INDEX · p. 15, Table 3 · p. 15
  • Before release, machine-specific paths and identifying terms are removed, and exported files are re-parsed and re-scanned..Source: A REPRODUCIBILITY INDEX · p. 15

B LIMITATIONS

The study evaluates a directed extension on a limited set of multi-stage cases, uses four game-quality dimensions that omit several higher-level properties, and focuses on compact games within practical agent budgets.

B LIMITATIONS

  • The directed extension is evaluated on a limited set of multi-stage cases; broader questions about when and how much guidance helps, and how human and model guidance differ, remain future work.The core pipeline remains fully autonomous; the limitation concerns the optional directed extension.Source: B LIMITATIONS · p. 15
  • Evaluation covers Mechanics, Depth, Visuals, and Art, but does not fully capture originality, narrative quality, long-term engagement, or subjective enjoyment.Source: B LIMITATIONS · p. 15
  • Experiments focus on relatively compact games that fit practical agent budgets; extending recursive development to larger projects with longer horizons, richer assets, and more complex dependencies remains future work.Source: B LIMITATIONS · pp. 15–16

C RSIGAME SYSTEM DESIGN AND IMPLEMENTATION

This section describes RSIGame’s recursive development and quality-monitoring design, including director briefs, the shared review interface, and the separation of agent decisions. It also explains how the verifier and global monitor assess defects and build changes, and how evaluation and token costs are accounted for.

C.1 RECURSIVE DEVELOPMENT PROCEDURE

  • Development restarts each guided stage from the best champion, and ends only when a stage saturates without further guidance; the paper uses E = 3 and K = 3.A champion promotion resets the idle-checkpoint count; saturation advances the stage when guidance is available.Source: §C.1 · p. 16

C.2 GLOBAL QUALITY MONITORING

  • At checkpoints, the Monitor compares champion and candidate on shared replayed demonstrations, scores only jointly observed criteria, and adopts a candidate for a 0.05 brief-score gain if task quality drops by no more than 0.05.The rubric reports weighted brief and task values separately, and the Monitor is isolated from the benchmark’s held-out rubric and scores.Source: §C.2 · pp. 16–17
  • The Monitor checks every three rounds and stops after K = 3 unchanged checkpoints; Figure 4 reports that this setting sits at the knee between early stopping and fixed-budget performance.Smaller K stops sooner and delivers less, while larger K approaches the fixed-budget score with fewer rounds.Source: §C.2 · p. 17

C.3.1 WHAT THE DIRECTOR IS GIVEN

  • The director reviews a live build alongside its task specification and a thin development summary, but receives no scores, source, logs, transcripts, or other director’s review.The interface is intended to elicit what the game needs next rather than repeat the loop’s score-based assessment of the last round.Source: §C.3.1 · p. 17

C.3.2 ONE INSTRUMENT, ONE BUDGET

  • Human and model directors use the same review session and budget counters; six converged game builds were independently reviewed by both, with each brief tested against a no-brief branch.A director receives 60 actions and two resets; human directors use a browser, while models use a command-line client connected to the same game session.Source: §C.3.2 · pp. 17–18, §C.3.5 · p. 20

C.3.3 HOW THE MODEL DIRECTS

  • The model director uses a frozen prompt, must base its review only on the session, and submits a four-field brief whose criteria guide planning, stage ranking, prompts, and Monitor weighting.The fields are a stage objective, why now, one to three directional priorities, and zero to three items to preserve.Source: §C.3.3 · p. 18, §C.3.4 · pp. 18–19

C.3.4 WHAT A BRIEF IS

  • From the same Alley Brawlers checkpoint, human- and model-directed branches reached 62.0 and 61.5, respectively, versus 60.0 for a no-brief control.The directors identified different needs—presentation and onboarding versus mechanics—showing that briefs supplied directions not produced by the loop during the prior eighteen rounds.Source: Table 4 · p. 19, §C.3.4 · p. 19

C.4 AGENT PROMPTS AND OUTPUT CONTRACTS

  • The prompts separate agent decisions: the Controller chooses what to find out, the Explorer selects an actionable probe, the Editor repairs an established problem, and the Verifier checks visible goal achievement and breakage.The Verifier can mark achievement unclear when the demos do not reach the relevant game state, and distinguishes missing features from broken inputs.Source: §C.4.1 · p. 20, §C.4.2 · p. 21, §C.4.3 · p. 22

C.4.4 VERIFIER

  • The verifier reports an input as broken only when it changes nothing, and directs the repair agent to fix the worst issue first.A feature the game never had is a task requirement, not a broken feature; visible responses that need more scripted steps are not faults.Source: §C.4.4 · p. 23

C.4.5 GLOBAL QUALITY MONITOR

  • The Global Quality Monitor compares champion and candidate builds under the same fixed script, treating only persistent changes visible through the final frame as build differences rather than recorder timing.It compares relative changes without assuming the newer build is better, marks unexercised behavior unobserved, and reports regressions separately; it does not score either build or propose repairs.Source: §C.4.5 · pp. 23–24

D.1 EXPERIMENTAL CONFIGURATION

  • The appendix identifies model roles and run parameters, and scores each method using the Qwen3.8-27B judge’s mean over 140 tasks and three independent replay-and-score runs.Base rows use frozen P0, Play2Code rows use the version after thirty rounds, and RSIGAME rows use the monitor-held version at saturation stop; tasks without a project or unfinished development score 0.Source: §D.1 · p. 24, Table 6 · p. 24, Table 7 · p. 25

D.2 TOKEN AND COST ACCOUNTING

  • Token and cost accounting separates generation from development and excludes evaluation-judge scoring; cache reads are excluded from billable tokens but included in cost.Development cost sums usage reported for agent calls, while generation is reported separately.Source: §D.2 · pp. 24–25
  • A factor of eight separates GPT-5.5 generation bills under Codex and OpenGame, with the difference attributed to the harness rather than the model.The reported cache rates are 94% under Codex and 42% under OpenGame.Source: §D.2 · p. 25

D.3 REPLAY AND SCORING STABILITY

Replay-and-score evaluations show that variability depends on when a task’s rubric evidence arrives, while RSIGAME’s relative gains persist under independent play and a second judge. The section reports three-run scoring stability alongside these robustness checks.

D.3 REPLAY AND SCORING STABILITY

  • Across three independent runs, the median score range is 0.73 points among 529 artifacts; 225 score identically in all three runs.Improved artifacts have a median range of 0.87, compared with 0.36 for frozen artifacts.Source: §D.3 · p. 26, Table 10 · p. 26
  • Replay variance tracks how late rubric evidence arrives: sports, horror, and openworld vary most, while rhythm, racing, and strategy have zero median range.The late-deciding criteria include match outcomes, haunting resolution, and world traversal; rhythm and racing criteria are resolved by replayable frames.Source: §D.3 · p. 27, Table 11 · p. 27
  • RSIGAME ranks above Play2Code and the frozen base under both judges; its gains over the base are +14.04 points with Qwen3.8-27B and +12.04 with GPT-5.5.The judges agree on the RSIGAME-over-base direction for 37 of 40 tasks, despite only moderate absolute-score correlation (ρ = 0.64).Source: §D.3 · p. 27, Table 12 · p. 27

D.4 FREE-PLAY EVALUATION

  • RSIGAME is preferred over the frozen base on 17 of 20 free-play tasks and over Play2Code on 18 of 20, with benchmark ordering matching 45 of 59 decided comparisons.Blind agents chose their own actions and trajectories, rather than using the demonstration scripts.Source: §D.4 · p. 28, Table 13 · p. 28

D.5 STATISTICAL RELIABILITY OF MAIN RESULTS

The appendix evaluates statistical reliability, adaptive development policies, verification, and multi-stage guidance with paired or blind comparisons. It reports significant gains over frozen starts across settings, while identifying an unresolved Qwen Phaser comparison against Play2Code.

D.5 STATISTICAL RELIABILITY OF MAIN RESULTS

  • RSIGAME gains +8.8 to +14.3 Overall points over frozen starts across included engine–generator settings; its Qwen3.8-27B Phaser comparison with Play2Code remains unresolved (+1.55, p = 0.394).Against Play2Code, gains are significant for both Godot generators and GPT-5.5 on Phaser.Source: §D.5 · p. 28, Table 14 · p. 29

D.6 ADAPTIVE DIRECTION POLICY EVALUATION

  • Adaptive and round-robin policies are compared on 16 paired games over 30 rounds, with development focus as their only difference and outcomes analyzed using signals withheld during development.A held-out diagnostic rubric identifies the lagging quality dimension only for post-hoc state-conditioned analysis.Source: §D.6 · p. 29

D.6.1 PAIRWISE JUDGING CRITERIA

  • Blind pairwise judging compares builds on visible mechanics, content depth, functional visuals, presentation, and overall preference, with broken builds losing dimensions they fail to show.The criteria require reasons to identify visible evidence and reject extra assets or effects as sufficient evidence of quality by themselves.Source: §D.6.1 · pp. 29–30

D.7 AGENTIC VERIFICATION AUDIT

  • The verification audit tests whether pre-edit targets are evidence-grounded and whether post-edit verification detects unsuccessful changes or regressions, against free-form-critic and build-only baselines.The audit uses development trajectories from 40 GPT-5.5-initialized games and 48 adjudicated rounds for post-improvement detection.Source: §D.7 · pp. 30–31

D.8 MULTI-STAGE GUIDANCE EVALUATION

  • Later development stages are consistently preferred over earlier champions in within-game blind comparisons, assessing game preference rather than benchmark score.The comparisons cover stage 2 versus 1, stage 3 versus 2, and stage 3 versus 1; the figure reports later-stage blind pairwise win rates with ±1 standard error.Source: §D.8 · p. 31, Figure 9 · p. 31

E.1 TRAINING CORPUS CONSTRUCTION AND CURATION

The appendix documents how development trajectories and generated games are curated, serialized, and used for experience internalization, then reports a comparison with VibeGame and qualitative evolution cases. It flags incomplete recordings in parts of the released training data.

E.1 TRAINING CORPUS CONSTRUCTION AND CURATION

  • The corpus includes 1,998 de-duplicated games and 910 complete run trees; examples are filtered for executable artifacts, valid plans, and successful sessions, with improvement rounds positively labeled only after independent verification.The corpus also includes generation, planning, and improvement supervision; failed or empty improvement rounds are dropped.Source: §E.1 · pp. 31–32, Table 15 · p. 32
  • All 284 Phaser generation trajectories and one Godot improvement arm contain flagged recording gaps; a clean subset can be formed by filtering the released rows.The affected rows identify summary-only and incomplete-trajectory conditions explicitly.Source: §E.1 · p. 32

E.2 FINE-TUNING CONFIGURATION

  • Qwen3.8-27B is fine-tuned on long agentic trajectories with loss applied only to assistant outputs; the full model combines planning, generation, and independently verified improvement examples.The compared training variants use the same configuration and differ in corpus composition.Source: §E.2 · p. 33

E.3 TRAINING-DATA ABLATION

  • The training-data ablation compares one-shot Godot generation variants over all 140 tasks; Figure 4c draws the +GEN and full-model variants.BASE and the full model correspond to the Qwen3.8-27B and Qwen3.8-27B (SFT) rows in Table 1.Source: §E.3 · p. 33, Table 18 · p. 34

F.1 COMPARISON WITH A MULTI-AGENT GENERATE-AND-VERIFY SYSTEM

  • On VibeGame’s twelve shared Phaser tasks, VibeGame scores 23.3, the frozen base 50.1, and RSIGAME 62.4; VibeGame scores 30.5 against 46.0 for one-shot generation on its own tasks.The comparison uses one VibeGame trial per task, with GPT-5.5 assigned to all eight roles.Source: §F.1 · p. 34, Table 19 · p. 35

F.2 QUALITATIVE EVOLUTION CASE STUDIES

  • Three qualitative cases trace each game’s official scores, retained champion, saturation stop, later guidance passes, edits, and judged frames through development.Builds are driven through the same script so their displayed panels are comparable.Source: §F.2 · p. 35

F.2.1 HOLDING THE BEST VERSION THROUGH REGRESSIONS

The case studies show how the global monitor preserves the best checkpoint through regressions, while outside briefs can supply direction when the loop’s evidence runs out. Outcomes vary with the starting game: briefs help expose and address an unwinnable match, but do not improve an already functional Block Cascade score.

F.2.1 HOLDING THE BEST VERSION THROUGH REGRESSIONS

  • R12’s 86.1 champion, up from 61.7, survived regressions and was delivered at round 21 after saturation.Art and feedback rounds lowered scores, so the monitor kept the stronger checkpoint; a verification replay could also miss a real fix, and the official judge could reward art covering the play field.Source: Figure 10 · p. 35, §F.2.1 · p. 37
  • Lawn Guardians had working rules but poor presentation: only 4 of 46 shipped art files were referenced.After three stable checkpoints, an outside stage brief supplied new direction, followed by passes for controls, animation, set dressing, and a written soundtrack.Source: §F.2.1 · p. 36, Figure 11 · p. 36, Figure 12 · p. 37

F.2.2 TWO BRIEFS ON THE SAME CHECKPOINT

  • Director briefs identified that nobody could win and that fighters overlapped; brief-guided branches scored 61.5 and 62.0 versus 60.0 for the control.A second brief made exchanges more readable rather than raising the score.Source: §F.2.2 · p. 38
  • Thirty autonomous rounds improved Alley Brawlers’ scenery while its fighters became torn smears, leaving the delivered champion at the base game’s score.The saturation stop delivered that champion; the score gains from the visual improvements were canceled by the degraded fighters.Source: Figure 13 · p. 38, §F.2.2 · p. 38

F.2.3 A GAME THE LOOP CANNOT IMPROVE

  • Nine checkpoints never improved Block Cascade’s already functional generated build; brief-guided and control branches all reached 50.0, with briefs reaching it at round 6 rather than 12.The second brief added a named mode and live clock, but introduced clashes with existing readouts.Source: §F.2.3 · pp. 40–41, Figure 16 · p. 41

F.3 PER-FAMILY RESULTS

  • Per-family tables cover Godot and Phaser generator groups, but their captions alone provide no family-level scores.The captions identify individual generator groups, including Codex + GPT-5.5 on Godot and OpenGame + GPT-5.5 on Phaser.Source: Table 20 · p. 43, Table 25 · p. 46
Loading 2609.39045v1…