# PENGO-E020 — E003 reader-only counterfactual, protocol v1.1

Preregister before the first new simulation. This implements review Y2, after
the sealed Y3 explanation and reader admission. One new trial; neither resume
nor retune E003. Category: TRAJECTORY SEARCH / READER-ONLY COUNTERFACTUAL.

Hypothesis: E003's missing egg observation contributed to ACT3 stagnation.
Success: ACT1/2 frozen prefix followed by ACT3 completion, three lives and no
detected death on the entire retained reset trajectory. A new stagnation is
NUL for this configuration: correction of the reader alone was insufficient.
Neither outcome establishes a unique cause or robustness over new mazes.

Seed is E003's exact E006 ready tape ending at physical frame6017,6018bytes,
SHA256 abb14d7ddc2a1094f97909116bd70c589a6c940ec99019fe70438a88034c862f.
Its canonical COIN/START, ROM pengoja, MAME0.289 and baseline DIP remain fixed.
This is prefix reuse, not a new solution of ACT1/2. No9790tape or Rack tape.

Keep the old E003 controller oracle_v2.lua, E012 phase/death reader, all
candidate full-suffix safety checks, BeamSearch20workers/width4/depth4,
legal-action order, original E009 danger/distance/tactical/visit heuristic
and E003 removal of score tie-break. All source files remain immutable.
The new runner retains identical Policy AST and evaluate AST after experiment
label normalization; input/worker names and forward outcome metadata differ.

The only behavioral intervention is state.eggs via the opt-in E020 reader.
Capacity is derived from each act's native initial header, full coordinate
list, player initializer state1 and enemy initializer states0/14, requiring
two identical consecutive observations. The measured baseline capacities
are ACT1=6,ACT2=6,ACT3=8, inferred from the reset prefix before6017. There is
no global capacity8. Later acts are not admitted. Parser bound32 is a safety
ceiling, not a claim about game capacity. A missing/malformed census remains
UNKNOWN. Frozen prefix memory preserves an egg in native ice transport;
unexplained mismatches remain UNKNOWN, including E0096501 and6895.

Admission v1.1 has2570known egg observations out of2572, all1290E003 frames,
1280of1282E009 frames. The first stricter negative admission is retained.
All non-egg state fields are unchanged. Labels use privileged RAM/object
observations. There is no claim of a pixel agent or runtime pixel validation.

Hard bounds are identical to E003:50,000,000 physical frame evaluations
including all reset prefixes,600committed decisions,2400seconds search wall,
25consecutive playing decisions without official counter reduction, one
attempt, no retry/escalation. Wall admission and already submitted bounded
workers retain the original scheduling semantics; report actual elapsed
time, including any in-flight completion. Each worker has its inherited
180second infrastructure timeout. Root W0 and its cost remain included.
No safe branch, unsupported representation and normal budget exhaustion are
NUL. A genuine worker/transport/parity/storage failure is
INFRASTRUCTURE_INCOMPLETE; human stop is separately STOPPED_BY_USER.

Contact/crush/counter clocks will be retained in retrospective event logs.
Reference event onset6778, confirmed crush6783 and acknowledgement6895 are
distinct. Contact is the causal onset only once its kill is confirmed; the
official decrement gives no second credit. Do not change this trial's rank
or counter watchdog to causal-event progress: that would confound Y2.
E018 controller and redesigned E005 causal policy are excluded.

Stop at the first healthy ACT3 official counter0, truncate the unused tape
suffix as E003 does, and retain full paid worker cost. No ACT4 planning.
Success requires the original separate fixed1800neutral-frame transition
probe, observed ACT4 entry with3lives/no death, followed by two fresh native
replays of the trajectory ending at first ACT4 entry. Compare declared trace,
sampled pixels and exposed devices and audit death frame-by-frame. Complete
machine identity remains UNKNOWN. A candidate pending this verification is
not success. Failure to establish the transition within the fixed probe is
NUL, with no retrospective extension. Verification costs are separate from
the search budget. Later ACT egg capacity may remain UNKNOWN at entry.

Before running, prepare writes a source-hash manifest with all inherited
sources, new runner/reader/protocol, seed tape/observed prefix, all three
historical contracts and validation receipts. Run rechecks it before the
first simulation. Output creation is exclusive; duplicate launch fails.
Do not start after7October06:15UTC, preserving a window before the authorized
07:00UTC work deadline. Freeze is not a game outcome. Preserve all negative
results and publish them. No score/diamond objective, models, loads, RAM
writes, cheats, cloud, new infrastructure or historical source edits.

Y4 later-act Rack validation remains an explicitly open diagnostic item:
E019v2 reached only ACT3, never ACT4. It neither seeds this trial nor proves
later-act capacity. Decision-model v2 remains stopped on its prerequisites.

## Pre-emulation serialization repair

Version1 froze successfully but its launch admission compared integer keys
in the freshly inferred capacities with JSON string keys in the frozen file.
It stopped before creating any search folder or MAME process. Preserve that
manifest and zero-cost failure. Version1.1 normalizes both sides through JSON
for comparison and adds a read-only check CLI. No reader, policy, root, action,
controller, budget or game-result criterion changes. This is still one game
attempt; the unexecuted version1 consumed zero game attempts.