Experiments on Arcade.it AI Lab

E001 — Infrastructure & first gameplay

Sun, 04 Oct 2026 00:00:00 +0200

FROZENQWEN / COMPUTER VISIONOBSERVED

Validated checkpoint: 2026-10-05. Data last synchronized: 2026-10-05T13:10:46.576063+00:00.
Latest record verified from reviewed published run evidence. Registry counts retain the last full dataset checkpoint; record publication does not refresh the complete dataset. This is a snapshot, not a live counter.

Registry count checkpoint: 2026-10-05T08:28:16.420013+00:00. The latest reviewed record and its evidence update separately from full dataset totals.

Infrastructure and gameplay

A Windows research workstation with three RTX 3090 GPUs and 96 GB RAM runs local Ollama. A MAME bridge observes the emulated game and sends real joystick/fire inputs. Input recording, automatic restarts, video rendering and experiment logging form the initial laboratory loop.

E002 — Perception benchmark

Sun, 04 Oct 2026 00:00:00 +0200

Summary / the selected 9B baseline

9B + history3 + nearest4 → 90% Exact / 1.5% Opposite. The same model with a single 4× frame measured 80% Exact and 12% Opposite. These are horizontal player-localization results, not gameplay scores.

FROZEN v1QWENOBSERVED

A balanced, independently labelled benchmark tests horizontal player localization under different visual conditions.

200 game frames · 24 conditions · 4,800 completed inferences

Selected E002 v1 horizontal localization conditions
Model / input	Exact / 95% CI	Adjacent	Opposite	Median latency
Qwen 3.5 9B 4× nearest-neighbour + 3-frame history	90% 85.9–94.1%	8.5%	1.5%	1.984 s
Qwen 3.5 27B 4× nearest-neighbour + 3-frame history	91% 86.2–95.2%	7.5%	1.5%	4.390 s
Qwen 3.5 35B-A3B 4× nearest-neighbour + 3-frame history	86% 80.3–91.5%	9%	5%	4.031 s
Qwen 3.5 9B 4× nearest-neighbour, single frame	80% 73.7–86.1%	8%	12%	0.938 s

E003 — Controllers & Search

Sun, 04 Oct 2026 00:00:00 +0200

FROZENCOMPUTER VISION / POLICY SEARCHOBSERVED

Registry count checkpoint: 2026-10-05T08:28:16.420013+00:00. The latest reviewed record and its evidence update separately from full dataset totals.

Controllers and valid inputs

The controller branch uses fast computer vision and deterministic policies to produce legal movement and firing inputs. Timing is recorded in emulator frames so a policy can be replayed without relying on wall-clock scheduling.

E004 — Qwen selection & timing

Sun, 04 Oct 2026 00:00:00 +0200

FROZENQWENOBSERVED

Games: 30
Overall verified maximum around this phase: 2,020
Qwen selector maximum, approximately: 1,740

Registry count checkpoint: 2026-10-05T08:28:16.420013+00:00. The latest reviewed record and its evidence update separately from full dataset totals.

Higher-level selection

E004 tested Qwen as a policy selector instead of requesting every joystick action directly. It examined whether selecting a controller and its timing could improve gameplay.

E005 — Policy Search milestone

Sun, 04 Oct 2026 00:00:00 +0200

Search means zero model calls

Automated policy search explores movement and firing schedules. No Qwen or other model is called during a Search run. Candidate policies can be discovered in accelerated research; milestone confirmation runs in real time.

FROZENPOLICY SEARCHREPRODUCED

Verified real-time milestones: 3,590 → 4,450 → 5,050.

The plateau was useful evidence

Final E005 registry: 2,002 games, comprising 1,996 accelerated research games and 6 real-time confirmations. In the earlier 1,980-run sample, 12 overnight batches did not exceed 5,050.

E006 — Diversity Search

Sun, 04 Oct 2026 00:00:00 +0200

ACTIVE at checkpointPOLICY SEARCHREPRODUCED

Diversity Search improved the published verified real-time milestone from 5,050 to 5,750 points. Search uses zero model calls.

A new verified milestone

VERIFIED FROM INGESTED ARTIFACTSAUTOMATED POLICY SEARCHMODEL CALLS: 0REAL-TIMEREPLAY VERIFIED

Run ID: pleiads-20261005-131958-e6ceb4
Experiment / controller: E006 / native5
Score: 5,750 points
Survival: 216.760 s (emulated game time)
Real-time / replay: Confirmed / exact recorded-input replay
Screenshot matches: 2,239 / 2,239 captured samples
Original manifest SHA-256: 9df9cf29b00e0b2ad6f2d209b462bff7f27495a0c01c38b14e83a7c013ec3943
Evidence files: Original manifest · Ordered actions CSV · Replay comparison · Screenshot hashes
Video: Record article and video

All captured samples were checked, at one capture every six emulated frames. Screenshot bytes and game software are excluded. CSV preserves the original TSV action order. This is a laboratory milestone, not a world-record claim.