Pengo20k milestone:21,210 verified

20,000-point milestone PASS:21,210 alive and21,210 at tape end. Three lives remain and no death is detected in the retained9,184-frame trajectory. ACT1 and ACT2 complete; ACT3 remains unfinished with3 objectives. Category: TRAJECTORY SEARCH / PRIVILEGED / REAL TIME REPLAY VERIFICATION. This is the best verified planner trajectory; it is not a full-game/world record, a general robustness result or an autonomous pixel agent.

Independent promotion from a shorter healthy tape

The successful E035v1.3 configuration aligns three interior diamonds at frame8,615. Its original diagnostic tape adds1,800 neutral frames and later detects death at9,589. Those later frames are retained as diagnostic evidence and excluded from this shorter result. The milestone candidate is retrospectively selected at the first healthy post-bonus endpoint9,184, then frozen before two NEW independent native reset replays. Inputs after the final push at8,590 are explicitly neutral and are part of this new candidate tape.

Both replicas match the complete declared prefix trace, exact latched input tape and each other, with sampled rendered pixels and exposed CPU/driver registers also equal. Each has0 LOAD and0 RAM writes. At9,184 the reader reports controllable play,3 lives, score21,210 and objective agreement3=3. Each complete tape contains no detected death. A bonus phase initially labelled UNKNOWN was separately diagnosed; UNKNOWN bonus states never became restore search seeds. Preregistered verification, source registration, sanitized proof, input tape.

Native gameplay timing is60.606Hz (0.0165seconds/frame). ReplicaA takes151.529799seconds. The30.303fps video samples every2 emulator frames and lasts151.536seconds. Video SHA256 62921c2f66debae9315f2efd9844249dbc8f0e8f6b71ddd5b90f1bf95d2f9abf. Both alive and tape-end scores, lives, ACT status and the fixed cumulative work totals appear from the beginning. Video metadata, visual validation.

Measured diamond mechanics, with a narrow scope

The aligned and matched control probes are each reproduced in two new native reset processes. In the actual variant, the game pays 100 increments of100 points every2 frames, from8,919 through9,117:10,000 points total. In the control without the final alignment push, the same count-up window gains0. The three diamond positions persist; no new moving score actor is created, and the player/three surviving actors are frozen during the count-up. Healthy controllable play returns at9,184. The claim is limited to this ACT3 interior alignment; wall variants, other acts and repeated rewards remain unverified.

The original E039v1 criterion assumed one distinct≥1,000-point increment and an ordinary PLAY phase at the award. It was NUL. E039v1.1 is an explicitly post-hoc count-up analysis of those already recorded probes, not a new preregistered discovery. It admits this scoped mechanic. E035v1.2 also remains NUL: its route omitted the push that moved the middle diamond from column5 to2. An initial preflight failure and E035v1.1 infrastructure failure are retained. No failed configuration is overwritten or silently retried.

What made the search practical

The corrected objective reader uses a per-act census: observed capacities are6 forACT1–2,8 forACT3–7,10 forACT8–11 and12 forACT12–16. Startup, egg transport and overlapping hatch acknowledgements are explicitly modelled. Two fresh Rack diagnostics per act validate active lifecycle accounting; Rack Test is diagnostic only, never a score result or source prefix. Later scoped reader admission covers88 retained traces and152,355 active objective observations with0 additional objective faults.96 legacy UNKNOWN occurrences remain visible. RAM-derived accounting compared with the game counter is not independent pixel-reader accuracy.

Save-state use is conditional and scoped. E026v1.1 passes2,500 fidelity cases; E027v2 passes172 chain/input cases plus6 native fresh-reset canaries; E028v1.1 passes27 cell-transport cases. Earlier failures expose host timers, stale immediate pixels and COIN impulse state. Production restore is restricted to healthy initialized PLAY nodes with no COIN/START suffix; its disposable priming frame occurs before LOAD and is charged outside the tape. Negative/dead leaves cannot seed search. Complete-machine identity remains UNKNOWN. Backtracking uses only previously witnessed living branches and identifies them using both trace and actual-input hashes.

The observed trace SHA256 is 43b7a0c98b8984b31efe98ae35d1fd48ba05596ee6a1b9887a87e72cdbab520f. Its coverage is the PENGO01 header, physical frame indices,4,096 bytes of program RAM0x8000–0x8fff,16 sprite-coordinate bytes and5 driver flags. It excludes CPU, scheduler, timers, other devices, inputs and pixels. Inputs, sampled pixels and exposed CPU/driver registers are tested separately; an equal observed trace is not complete-machine identity.

Cost, baseline and next target

The immutable milestone cost snapshot contains 83,040,579 known physical frame evaluations and14,919 owned Pengo MAME launches, including reader/fidelity/mechanics diagnostics, failed/null configurations and independent verification. Reused evidence is not charged twice; Pleiads work is excluded. E035v1.3 itself uses568 search nodes and62,871 physical frames including4 native diagnostics. Independent M01 promotion adds18,368 frames and2 processes. Cost snapshot. The protocol/script syntax preflight for M01v1 used0 MAME; v1.1 changes only that syntax and namespace.

Baseline: pengoja, MAME0.289,3 lives, Medium, Upright, bonus life30,000, RackOff, DSW1=176 andDSW2=204. ROM archive SHA256 c5e17040c5cdb63e1bbc7e015bd00fb1eedbe7f0bf36bdccdac65104c0c596d8; the ROM itself is never distributed. No model weights, raw RAM/state dump, save-state or secrets are included in these public artifacts.

The next separately frozen milestone is30,000, then100,000 and300,000. Stagnation is NUL for its configuration, with diagnosis and a new counted registration before another attempt. Native fresh-reset final verification remains mandatory. Models/student/PPO/new games are suspended. Earlier9,790 and10,780 trajectories remain historical; frozen PleiadsE007/E008 are untouched. Comparable leaderboard provenance remains unverified.