Research timeline
Validated checkpoint: 2026-10-05. Data last synchronized: .
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.
Pengo M05: 500,350 verified / 500k PASS
500,350 alive / 500,350 at tape end; 4 lives remain, zero detected deaths. 86 chronological rounds completed; ACT 7 of cycle 6 is unfinished with 2 objectives. 0 MODEL CALLS; two NEW native reset proofs and independent pixel/round gates.
TRAJECTORY SEARCHPRIVILEGED0 MODEL CALLSNATIVE RESET + PIXELS VERIFIEDPengo M04: 301,030 verified / 300k PASS
301,030 alive / 301,030 at tape end; four lives remain, zero detected deaths. 52 chronological rounds completed; ACT 5 of cycle 4 remains unfinished with four objectives.0 MODEL CALLS; two NEW native reset proofs and pixel/round gates.
TRAJECTORY SEARCHPRIVILEGED0 MODEL CALLSNATIVE RESET + PIXELS VERIFIEDPengo M03B: 216,480 verified / 200k checkpoint
216,480 alive / 216,480 at tape end; four lives remain, zero detected deaths. 36 chronological rounds completed; ACT 5 of the third cycle remains unfinished with one objective. Two NEW native reset proofs and independent pixel-score gate. A08 goal300k remains NUL.
TRAJECTORY SEARCHPRIVILEGEDNATIVE RESET + PIXELS VERIFIEDPengo M03: 100,400 verified trajectory points
Privileged trajectory search. Two new complete native reset proofs,0LOAD/0RAMwrites. Report and video linked separately; no autonomous-agent/world-record claim.
TRAJECTORY SEARCHPRIVILEGEDNATIVE RESET VERIFIEDPengo M02: 31,040 verified trajectory points
Privileged trajectory search. Two new complete native reset proofs,0LOAD/0RAMwrites. Report and video linked separately; no autonomous-agent/world-record claim.
TRAJECTORY SEARCHPRIVILEGEDNATIVE RESET VERIFIEDPengo M01: 21,210 verified trajectory points
Privileged trajectory search. Two new complete native reset proofs,0LOAD/0RAMwrites. Report and video linked separately; no autonomous-agent/world-record claim.
TRAJECTORY SEARCHPRIVILEGEDNATIVE RESET VERIFIED5,750 points / verified real-time Search milestone
Native5 Search; zero model calls; exact Lua-input replay and 2,239 / 2,239 captured frames matched. Article and reviewed media: /lab-notes/pleiads-search-5750-20261005-131958-e6ceb4/
SEARCH / 0 MODEL CALLSREAL TIMEREPRODUCED5,600 points / verified real-time Search milestone
Native5 Search; zero model calls; exact Lua-input replay and 2,168 / 2,168 captured frames matched. Article and reviewed media: /lab-notes/pleiads-search-5600-20261005-124119-f2b26d/
SEARCH / 0 MODEL CALLSREAL TIMEREPRODUCED5,540 points / verified real-time Search milestone
Native5 Search; zero model calls; exact Lua-input replay and 2,156 / 2,156 captured frames matched. Article and reviewed media: /lab-notes/pleiads-search-5540-20261005-120227-03ddb1/
SEARCH / 0 MODEL CALLSREAL TIMEREPRODUCED5,510 / Diversity Search milestone
Real-time native5 Search; zero model calls; exact replay and 2,091 / 2,091 captured screenshots matched. Original run evidence is ingested.
SEARCH — 0 MODEL CALLSREAL TIMEREPRODUCEDE006: Diversity Search begins
The method changes: broaden behaviours, canonicalise effective policies and balance exploration with exploitation. No new results are claimed.
POLICY SEARCHACTIVEA plateau reveals ineffective switches
At the frozen checkpoint, 1,980 E005 runs and 12 overnight batches yield no score above 5,050. In 1,052 policies, planned switches occur after the final death.
OBSERVED5,050 points — verified Search milestone
Real-time execution and exact replay: all 1,898 captured screenshots match. There are zero model calls during the run.
REAL TIMEREPLAY VERIFIED4,450 points in real time
A further verified Search milestone reaches 4,450.
REAL TIMEPOLICY SEARCH3,590 points in real time
Search reaches 3,590 in real time with exact replay.
REAL TIMEREPRODUCEDSearch begins outperforming direct model control
Automated policy search explores controller policies. These results belong to Search, not to Qwen-controlled runs.
POLICY SEARCHE002: a frozen perception benchmark
The 200-frame benchmark completes 4,800 inferences. Qwen 3.5 9B with nearest4 and history3 reaches 90% Exact and 1.5% opposite-side error.
QWENOBSERVEDE001: the first automated cohorts
Vision, frame history, telemetry and episodic memory are exercised. Deterministic replay is demonstrated.
QWENREPLAY VERIFIEDLab infrastructure comes online
A Windows workstation with three RTX 3090 GPUs, 96 GB RAM and local Ollama connects to the MAME bridge. Input recording, restarts, video rendering and experiment logging are available.
OBSERVED