Pengo M03B: 216,480 verified / 200k checkpoint

216,480 alive / 216,480 at tape end. This passes the separately authorized 200,000 publication checkpoint, M03B. It inserts a checkpoint between frozen M03/100k and M04/300k; existing experiment IDs and outcomes are unchanged.

TRAJECTORY SEARCH / PRIVILEGED / REAL TIME REPLAY VERIFICATION. Native proof wall time: 1,892.75507 / 1,892.75494 seconds for A / B; nominal game duration 1,892.76450 seconds. The bridge measures machine.time zero to the terminal callback; Python setup and postprocessing are excluded. Clock scope and pinned receipts. The result is a trajectory prefix. It establishes neither a full-game completion nor an autonomous-agent result.

This is one canonical reset and one deterministic trajectory; robustness across different starts is not established by this milestone.

What passed

Two NEW complete native reset processes replay all 114,713 physical frames without LOAD or RAM writes. Actual inputs, the entire declared prefix trace and sampled rendered pixels/exposed devices agree. An independent terminal-pixel decoder reads 216,480 in both proofs and agrees with the result; corrupted glyphs are rejected. The pixel promotion receipt is bound to the final result hash.

An independent chronological audit checks both native proofs. ACT labels wrap after 16, so the count is two complete cycles plus ACT 1–4: 36 completed rounds, followed by unfinished ACT 5 of cycle 3. The later ACT 6 seen in the E061 neutral diagnostic is outside this tape and excluded.

For this publication, a third NEW whole native reset records the same tape. Its complete trace, actual inputs and sampled pixels/devices match both proofs. Every second rendered frame is presented at 30.303030 fps, with the final odd-frame snapshot adding 0.0165 seconds. Full duration: 1892.781 seconds, with no intentional cuts or speed-up of the native emulated cadence. The encoded video passes full decode, text-bound checks and visual inspection.

The complete capture/comparison operation took 1992.359 wall-clock seconds, longer than the nominal emulated duration. This measures the whole recording job, rather than isolated frame pacing. Native mode is requested and verified in the process manifest; it does not by itself prove an uninterrupted wall-clock rate of 60.606 Hz under screenshot I/O. This recording is separate from the two final native proof processes.

Search outcome and mechanism

A08’s preregistered objective was 300,000. It stopped for stagnation and remains NUL. Its independently verified lower prefix reaches 216,480; Stefano explicitly requested this 200k publication checkpoint. The new publication does not relabel A08 as a successful 300k experiment. Subsequent A09 work is separate and is not part of this milestone.

Search uses validated healthy-PLAY save-state restore, local CPU beam search, logical-cell macros and witnessed backtracking. Intermediate whole-reset checks may be accelerated. Final verification and this recording are complete native resets without restore. No model decision calls are used. Restore during raw17 departure remains forbidden.

The native ACT 5 diagnosis and separately frozen lifecycle/history/clone/negative-context admission explain the prior boundary uncertainty. They admit a scoped boundary reader to A09; the full offline generic causal ledger is not silently adopted.

Evidence and exact cost scopes

These scopes exclude earlier campaign work and ongoing A09, rather than treating them as zero. Historical and negative attempts remain in the campaign ledger and hourly SAL; no search or proof cost is counted twice here.

The separately dated campaign snapshot at 7 October 2026, 11:18:03 UTC retains 107 registered identifiers and a known lower bound of 99,339,861 physical frame evaluations / 144,200 owned launches. A09 is still running in that snapshot, so the campaign total is provisional. Per-identifier rows include reused corpora and historical categories and must not be summed directly; missing costs remain unavailable. All identifiers, outcomes and accounting scope.

The historical round audit retains public_report_allowed: false: it prohibits automatic publication of a search result. Stefano independently authorized this lower 200k checkpoint, and the separate recording/pixel/publication gates passed. The original audit and its flag remain unchanged; publication protocol v1.1 records this authority and distinction.

Native verification · Pixel gate · Chronological round audit · New recording evidence · Video metadata · Visual and decode checks · Publication protocol.

Input SHA-256: d33a5a90dd90ed20c4d7e55d4db5e64ca5cf11cbac828a988ef37fc2ec908dd4.

Observed trace SHA-256: 0790d606d4d3aeb1e0e41271cc0b092981aeeaa32ddb1953fb4393a2a1551981.

Video SHA-256: 1b8455417eaf9d8b99c1df1d592c17f530db2fc08e20709cdc663d47b302d801.

The observed trace hash covers the PENGO01 header, physical frame indices, 4096 bytes of RAM 0x8000–0x8fff, 16 sprite coordinates and five driver flags. Actual inputs, sampled pixels and exposed device state are checked separately. CPU/scheduler/timer and other device state are outside this hash; complete-machine identity remains UNKNOWN. Raw RAM, ROM, save-states, game-code bytes, input tapes and model weights are not distributed by this publication.

Video phase annotations

Boot, credit/start, round setup and egg reveal, bonus tally, intermission, diamond bonus, egg-destruction points and Sno-Bee stun are explicitly named. Both ACT 16 bonus-table/intermission animations are inspected in the new native capture. These are exact-tape editorial labels, not a new generic runtime phase reader or a claim that every possible context is understood. Trace-bound phase annotations.

Second-cycle score and pace

Whole-tape yield: 114.372 points per native game-second (216,480 / 1,892.7645 s), including boot, round setup, tally and intermission. This is not a GAME TIME gameplay-only metric.

Both native proofs agree on every interval. The interval starts at ACT entry and ends on the frame before the next ACT; it includes bonus tally and intermission. This measures actual game-score gain, not an estimated reward. All 16 second-cycle ACTs are complete; partial third-cycle ACT 5 is excluded from completed means.

Mean second-cycle gain: 5,450.62 points per completed ACT. The earlier approximate 6,700-per-ACT extrapolation is not assumed. Full paired interval evidence.

Settings, comparison and next step

MAME 0.289, pengoja, Medium, Upright, three initial lives, bonus 30,000, Rack Test OFF, DSW1=176, DSW2=204. Rack Test diagnostics supply no scoring seed. The native game rate is approximately 60.606 Hz. Four lives remain at tape end after the observed extra life.

The human comparison benchmark is 1,492,220 cumulative points, including two rollovers of 655,360 and a final HUD score of 181,500, as declared in the Paul Hornitzky / OOO submission, read directly on 7 October 2026. It is listed for OOO by Twin Galaxies. Its Points rules specify five initial lives; this laboratory uses three. The legacy ROM naming maps to pengoja in recent MAME. Human competition and privileged trajectory search are separate categories. This numerical comparison establishes no human rank or world record. This verified tape is valid with the laboratory’s three initial lives. The same inputs replayed with five initial lives diverge in sprite coordinates at frame 1,669 (E063v1.2, offline classification E068). They are not demonstrated valid with the TG settings. Stefano will choose the benchmark settings after the 300,000 milestone; the laboratory DIP remains unchanged.

Benchmark and set/settings provenance. Next verified milestone: 300,000. The broader ambition is to exceed the supplied benchmark; one million is intermediate. External human evidence and the static 16-bit score-word model agree on a 655,360-point rollover. The laboratory has not yet observed a legal crossing. Before a target above 500k, each rollover requires before/after word and rendered-pixel witnesses in both new native proofs, modular award handling and a small-score pixel gate. Any certified cumulative total will be labeled separately from the native HUD. The diagnostic RAM-write probe is withdrawn. Cumulative-score contract v1.1.

Project Pengo · M01 /21,210 · M02 /31,040 · M03 /100,400 · Complete YouTube replay. Historical Pengo and Pleiads evidence remains preserved.

Review T4 update, 2026-10-07T13:17:16.026158+00:00: updated comparison provenance. The dedicated five-life replay E063v1.2 first differs in unmasked raw life-related state at frame 181 and in sprite coordinates at frame 1,669, as classified offline in E068. Its final active-game score is 87,240 with 13 completed rounds, versus 216,480 and 36 for the verified three-life tape. E063 remains NUL for equivalence under TG settings; the laboratory DIP remains unchanged.