Pengo: a 9,790-point planner trajectory and the ACT 3 death it exposed

Research diary

Best verified planner trajectory: 9,790 at tape end; 9,390 alive. This is not a full-game record. ACT 1 and ACT 2 are complete; ACT 3 is reached but incomplete. The retained reset tape ends during a phase that the original reader labeled UNKNOWN. The new diagnostic identifies it as a death already in progress.

The complete retained prefix was verified in two fresh native-speed MAME reset replays. The new editorial video labels the lives value as a counter and explicitly marks the death already started in its final segment. Search used zero model calls and privileged observed state, with no save-state load, RAM write or cheat. It is not a Qwen or autonomous-agent result. Complete machine-state identity remains UNKNOWN.

What happened at frame 7,299

At frame 7,298 the player was ready. At 7,299, native pixels show contact with a live Sno-Bee, the player object changes to raw state 0 and a collision latch is set. The lives counter still shows 3. Exactly 600 neutral frames reproduce a death animation and the counter falling to 2 at frame 7,601.

Ten separate fresh reset probes test neutral, four directions, button alone and each direction with button. Every probe reproduces the identical prefix and the same counter decrement. Inputs after this collision cannot recover this tested state; earlier alternative trajectories remain a separate question.

Three diamonds remain at (1,11), (11,11), (5,13), with no contiguous alignment. Five objectives remain at collision and four after a pending enemy crush. The delayed +400 occurs in that crush/points sequence. Thus neither a diamond sequence nor last Sno-Bee escape explains this event. General diamond rules and values are still unverified.

Reader repair, kept separate from history

The opt-in phase reader recognizes the collision freeze using both the verified latch and player state. Raw state 0 alone remains UNKNOWN. Two further native-speed reset trajectories reproduce the diagnosis, including a distinct collision at 6,542 with the life decrement at 6,668. Per-frame checks cover 18,385 observations across these confirmations and frozen E002, without rejecting a historically playing frame. Three new focused reader tests pass locally and on the research machine.

Historical Pleiads and Pengo tapes, readers and verification receipts remain preserved. The repair does not complete ACT 3 or make the 9,790-point prefix death-free.

Next experiments, in order

PENGO-E003 — Three Acts, Zero Deaths targets ACT 1, ACT 2 and ACT 3 complete, with 3 lives remaining and no detected death. It reuses the healthy reset prefix to ACT 3 ready, explicitly documenting that reuse. It has no score target, score tie-break or diamond-bonus objective. Every candidate suffix is checked for death, rather than relying on its final lives counter.

E004 will separately verify current-ROM diamond mechanics and values. E005 scoring planning waits for those prerequisites. The original N1 preregistration already specified its 25-decision progress watchdog and classified an early watchdog stop as incomplete; N1b is not adopted.

304,000 remains a numeric target. No comparable leaderboard is established for pengoja, the current DIP/difficulty settings and assisted offline method. Reaching that number would not by itself establish a world record.

Reproduction and evidence

Set pengoja, Pengo (Japan, 315-5010 type, rev C), MAME 0.289; 3 lives, Medium difficulty, Upright, bonus life 30,000, rack test Off; DSW1 176, DSW2 204. Local MAME/Python CPU workers run on doppia-3090. Planner state and pixels-only agent state have distinct provenance; the pixel controller remains a guarded prototype.

These are selected laboratory observations and input records. ROMs, raw RAM, state dumps, extracted assets, emulator binaries, model weights and secrets are excluded. Gameplay video is editorial research evidence. No leaderboard homologation, full-game record or pixel-agent robustness is claimed.

E003 outcome update

The single preregistered Three Acts, Zero Deaths attempt is incomplete, stopped at progress-timeout with 6 ACT 3 objectives remaining. The retained recorded prefix has a lives counter of 3 and no observed death in an offline audit; ACT 3 completion is not verified. No score search or retry follows. Read the outcome and exact diagnostic captures.

E004 diagnosis update

E003 failed because the planner could survive the segment but could not reduce the six remaining ACT 3 objectives. The bottleneck shifted from survival to task progress. This is scoped to the recorded segment. E004 Objective Reduction Analysis identifies missing ACT 3 egg coverage: the inherited reader scans six entries, while the observed state uses entries 7 and 8 too. Three Sno-Bee and three eggs reconcile the endpoint count. One existing 6→5 ice-crush event is reproduced twice at native speed, with 117 frames from contact to count decrement. Read the diagnosis, geometry and new video. No new search follows. Diamond verification remains separate.

Editorial video review v2

The W1–W6 review is closed: fixed ACT 3: death at frame 7,299, 9,390 alive / 9,790 at tape end, fixed 13,211 TOTAL evaluations, EXPLORATORY PILOT title and explicit trace-hash region coverage. E009 v1/v1.1 are superseded and not publishable; E012 editorial v1 is unlisted and superseded. Original files remain unchanged. Read the review closure and revised video. No new search follows.