Pengo E003: three lives retained, three ACTs not yet completed

Research diary

PENGO-E003 is incomplete. The single frozen attempt stopped at progress-timeout with 6 objectives remaining in ACT 3 and a lives counter of 3. ACT 1 and ACT 2 came from explicit reuse of the healthy, previously verified reset prefix. The Three Acts, Zero Deaths milestone has not been achieved.

The endpoint score of 9,090 is only an observation. There was no score target, score tie-break, diamond-bonus objective or model call. The existing 9,790-point best verified planner trajectory remains unchanged and is not a full-game record.

The attempt committed 66 decisions and evaluated 4,124 nodes, consuming 27,797,990 physical emulator frames including reset prefixes and 2198.906 seconds of measured search wall time. All work, including rejected candidates, is retained. No budget or parameter was changed and no retry was launched.

The new phase gate screened whole action suffixes. It rejected 343 unsafe suffixes: 156 collision-delay deaths, 147 death animations and 40 unexplained phases. UNKNOWN remains excluded; general reader completeness is not established. A separate offline audit finds no observed death or life-counter fault in the retained recorded prefix. That limited observation does not promote an incomplete run: E003 has no verified ACT 3 completion or two-replay milestone receipt. Complete machine-state identity remains UNKNOWN.

Why the phase gate matters

Exact native captures before and after the ACT 3 death, with privileged diagnostic labels

This diagnostic figure shows exact native captures from the earlier 9,790-point trajectory, enlarged by nearest-neighbor sampling. Labels come from privileged diagnostic observations; the figure is not an input to the pixel agent. Death starts at frame 7,299, while the lives counter still reads 3. The fixed continuation drops the counter to 2 at 7,601. A high lives counter alone cannot establish a safe action.

The diamond and last-Sno-Bee hypotheses were tested explicitly. Across 1,735 observations before the board reset, the three recognized diamonds stay at (1,11), (11,11), (5,13), without contiguous alignment. Five objectives remain at collision, with three visible live or hatching objects. The delayed +400 occurs as a crushed-enemy slot enters its points state. These findings reject those hypotheses as explanations of this death; general diamond mechanics and values remain unverified.

Next gates remain in place

The original N1 preregistration already contains the progress watchdog. N1b is not adopted. E004 remains a separate pending diamond-mechanics experiment; E005 scoring has not started. 304,000 remains a numeric target, without a verified comparable leaderboard for pengoja and the current settings/method.

Historical Pleiads artifacts and reproduction remain unchanged. Pengo uses MAME 0.289, pengoja, 3 lives, Medium, Upright, 30,000 bonus, rack test Off, DSW1 176 and DSW2 204 on local CPU workers on doppia-3090.

Only selected editorial evidence, counts and hashes are published. ROMs, raw RAM/state dumps, extracted templates, emulator binaries and secrets remain excluded.

E004 diagnosis update

Hypothesis pending the reader-only counterfactual: the recorded ACT 3 stagnation may reflect incomplete egg representation, planner limitations, or both. The retained segment has no detected deaths and25decisions without official objective reduction. The old reader made all egg targets UNKNOWN; this observation does not isolate the causal bottleneck. The causal claim remains untested in this 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.

Review correction6October2026

The causal interpretation above is a hypothesis. The next controlled trial is E020, E003 with only the egg reader changed, preserving the original controller, heuristic, seed and budgets. The redesigned E005 is deferred. No new scoring or autonomous-agent result is claimed. Future stagnation is NUL for its registered configuration; infrastructure failures are separately incomplete. Frozen historical results retain their original labels.

Y1–Y7 response and exact mismatch explanation. The11E003 and7E009 original arithmetic discrepancies are explained by egg transport and one retirement/hatch boundary. This compares RAM-derived counts with a counter in the same RAM; it is not independent pixel-reader accuracy. Per-act initial counts measured so far are6,6,8 forACT1,2,3; ACT4+ remains UNKNOWN despite the finite Rack diagnostic. The new reader preserves two additional E009 header-update UNKNOWNs.

Current9,790 editorial v2 already applies W1–W6; the originals remain unchanged and superseded.9,390last alive /9,790tape end, death7299, ACT3incomplete.