Add Flappy Bird local model comparison demo - #10
Open
Alex-Wengg wants to merge 3 commits into
Open
Alex-Wengg wants to merge 3 commits into
Alex-Wengg wants to merge 3 commits into
Conversation
Native macOS Flappy Bird (60 Hz fixed step) where a Core ML decision model picks flap/coast every six frames. Selectable backends: GLiClass Edge Apps v2 LUT8, Laya multilingual E8, GLiNER 2.5 base W8, GLiNER 2.5 multilingual W8, plus manual and heuristic control. Physics keeps running while inference is in flight, so late replies are visible. The prompt is a compact text state with a 300 ms physics forecast per label; option order alternates. An optional safety guard overrides a choice when its forecast collides and the other action is safer; the app counts overrides so guarded play is not mistaken for model skill. FlappyBirdCheck runs a bounded headless comparison (seeds, raw vs guarded) at the 10 Hz decision cadence. On seeds 1-4 (M5 Pro), raw policies: GLiNER 2.5 multilingual 1 pipe / 10.98 s combined survival; Laya, GLiClass, GLiNER base 0 pipes / ~5.4-5.9 s; heuristic 40 pipes. No model is a reliable controller yet; this is the harness for comparing further models as they are integrated. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Wires the models merged in #9 into FlappyBirdPolicy: GLiNER 2.5 small, Verdict (native, L128, abstention dropped so one action is always chosen), and the PublishedCoreML bridge (Kev 0.5B/0.6B and Kai/Lex via System One choice, LFM2.5-350M-RLCD via a constrained enum field with the forecasts in context, Jeff via label classification, NanoJev from a local conversion via NANOJEV_MODEL_DIR/NANOJEV_PYTHON). Raw policy, seeds 1-4 (M5 Pro): none of the nine passes a pipe; combined survival 5.6-9.0 s vs GLiNER multilingual's 1 pipe / 10.98 s. LFM picks the second-listed option on every call (position bias) and, with NanoJev, exceeds the 100 ms decision period. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Alex-Wengg
added a commit
that referenced
this pull request
Sep 23, 2026
Flappy Bird was built and run in #10 against all 13 integrated decision models. None is a usable controller (best: GLiNER 2.5 multilingual, 1 pipe over seeds 1-4; heuristic 40). Moves it from "next demo" to tried in the ranking and demo tables and adds a results section. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Alex-Wengg
added a commit
that referenced
this pull request
Sep 23, 2026
The Flappy Bird trial (#10) showed per-tick control is a poor fit for label-choosing models. Removes Mario, the Subway Surfers-style runner, Frogger, Pac-Man, Space Invaders, Breakout/Pong, and Racer from the ranking, demo list, and reuse sources (runner repos and paragraph), and lists them under "Ruled out: real-time games" with reasons. Re-ranks the remaining turn-based games toward setups where each legal move can be described with its consequence (the Tetris harness pattern): Connect Four, chess position challenge, Codenames first. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Native macOS Flappy Bird demo + headless checker for comparing local decision models as real-time controllers. Covers every model on main, including the #9 runtimes; more can be slotted in as they land.
GLiClassFlappyDemo: 60 Hz game, model picks flap/coast every 6 frames while physics keeps running. Controllers: GLiClass, Laya, GLiNER 2.5 small/base/multilingual, Verdict, bridged Kev 0.5B/0.6B, Kai, Lex, LFM2.5-350M-RLCD, Jeff, NanoJev (local weights only), manual, heuristic. Optional safety guard (override count shown) and delay slider.FlappyBirdCheck: bounded headless runs by seed, raw vs guarded (--raw-only --seeds=1,2,3,4).FlappyBird/FlappyBirdPolicytargets + tests.Raw policy, seeds 1–4, M5 Pro (bridged times include the Python worker round trip):
No model plays reliably yet. LFM always picks the second-listed option, which suggests position bias. Guarded scores mostly reflect the deterministic guard. Details are in
Sources/GLiClassFlappyDemo/README.md.🤖 Generated with Claude Code