# PENGO-E003 — Three Acts, Zero Deaths

Frozen after E012 diagnosis/reader validation and original N1 recovery, before
execution. This is a progression experiment, not score search. Success:
ACT1 complete, ACT2 complete, ACT3 complete,3 lives remaining and no detected
death anywhere along the full reset trajectory. No304000 objective, diamond
alignment/bonus objective, score tie-break, model calls, saves/loads or RAM write.

Reuse the immutable reset prefix to ACT3 ready6017: E002 ACT1, verified E006
ACT2 and its separately measured intermission. Audit it with the opt-in E012
reader for deaths before search. This is explicitly prefix reuse, not a new
independent solution of the first two ACTs. The9790 tape is not a seed because
it already contains a death. ROM/set/DIP/input clock remain unchanged.

Same opaque generic beam engine, width4/depth4/20 CPU workers. Reuse the E009
progression/distance/danger heuristic while removing its score tie-break.
Every candidate's entire applied suffix must have no death latch, death phase,
life decrement or unexplained game phase, not merely a safe endpoint. The new
reader identifies the collision freeze before the lives counter changes.

Hard limits:50,000,000 physical frames including reset prefixes,600 committed
decisions,40-minute wall limit,25 playing decisions without objective progress,
no safe branch, reader disagreement or human stop. One attempt. All costs and
rejected nodes remain recorded. No silent parameter changes/retry/escalation.

Stop a candidate at the first healthy ACT3 objective count zero, truncating
any unused macro suffix explicitly and retaining its paid full worker cost.
Do not search the unseen transition. Measure it with a separate fixed1800
neutral-frame reset replay. Only observed ACT4 entry with3 lives and no death
allows an ACT3-complete candidate; retain the prefix ending at first ACT4 entry,
before subsequent play. If the transition is not established within that fixed
probe, report incomplete and do not enlarge it retrospectively.

Promotion requires two fresh native-speed reset replays of the retained
candidate, all declared trace bytes/sample pixels/exposed devices identical,
and a full per-frame zero-death audit. Complete machine-state identity stays
UNKNOWN. E004 verifies diamond mechanics separately; E005 scoring waits for
those verified mechanics and this methodological milestone. N1b is not adopted.