Flappy Bird Model Benchmark
Every model gets the same frozen 5-prompt agentic mission in the same repo: run SOD, implement a Flappy Bird clone from spec, verify it in a browser, register and deploy it to the play gallery, then run EOD. Benchmark v1, prompts and rubric are version-controlled in the repo (internal/projects/game/flappy-bird-clone/benchmark/).
Leaderboard
| Model | DoD | Spec | Self-rev | Recovery | Verify | Process | Efficiency | Tools | Commits | Min | Audit |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Kolibri 1 78B-A3B Q4_K_M kolibri (local) · 2026-10-07 | ✅ | 8 | 7 | 6 | 2 | 5 | 105 | 0 | 126.8 | audit | |
| Qwen 3.8 Flash Next (UD-Q4_K_XL) lemonade-1 (local) · 2026-10-27 | ✅ | 9 | 8 | 9 | 3 | 8 | 108 | 1 | 80.3 | audit | |
| Qwen 3.8 Flash Next IQ3_XXS (Strata, 128K) strata (remote) · 2026-10-27 | ✅ | 9 | 8 | 4 | 3 | 5 | 327 | 2 | 137.5 | audit |
⚠️ partial = not the full frozen 5-prompt protocol: the model under test ran fewer phases (see its audit for scope and deviations); the operator completed the rest.
Hard rules
- Prompts are frozen; operator feeds one per turn, no mid-run rescuing. Gate failures are the measurement.
- Secret-scan or CI gate suppression (baselines, allowlists to silence findings) caps process discipline at 3.
- Every score row links to an evidence-backed audit: session metrics, redacted transcript, screenshots.