# PENGO-E005 — Objective Reduction Planner, executable protocol v1

New configuration after E020 terminates; no retuning or resumption of E003,
E009 or E020. Target is ACT3 objectives6→0 with3lives and no detected death.
Score is recorded, never ranked. Category: TRAJECTORY SEARCH / PRIVILEGED.
No autonomous-agent or full-game record claim. 304000 is a numerical target
for later work, not an objective here; diamond bonus is not used.

Root: original retained E003 chosen reset tape clipped at frame7049, the first
full known aligned board after the formerly proposed6876root. The latter has
two UNKNOWN cells and is rejected. This measured root change is frozen before
new simulations. Offline seed gate checks healthy retained prefix, ACT3,
counter6,3lives, aligned player, all195cells known and reconciled eggs. Two
fresh native reset replays must match the exact observed prefix and each other
before search. Those replays are charged separately and do not change the root.

Gates: native E015 first30crush events (23complete,7death-censored); E016 two
blocked egg geometries with neutral controls and startup hatch (10native
traces); E003/E009 transport and counter acknowledgement reconciliation;
E018v2 nine-case admitted controller; E004 frozen summary; source contracts.
All gates retain source hashes and limitations. Unsupported simultaneous
events or attribution are UNKNOWN. These samples do not prove all mechanics.

New controller: opt-in E018v2 anticipatory release/settle, scoped nine cases.
New representation: per-ACT egg reader plus prefix-only causal events. Crush
contact is logged at raw10, confirmed at raw11 only with observed ice context.
Confirmed event gives one credit; official decrement acknowledges it and gives
no second credit. Direct egg destruction requires applied push, reachable
adjacent aligned stance, blocked ice and same-frame entry/counter removal.
Slide and hatch alone give zero credits. Event identity includes observed
prefix, actor generation and contact. Observed counter remains authoritative.

Policy ranks completion first, then unique confirmed causal credits since root,
then reachable attack setup quality minus measured danger and revisits. Setup
uses empty-cell BFS to a stance behind ice; clear push lane with a present
enemy, or a blocked egg block. Future enemy alignment is explicitly uncertain
and evaluated by simulation. Exact feature weights/code are frozen. Unknown
board cells are barriers for reachability; unknown eggs do not invent targets.
All legal MOVE/PUSH/WAIT macros are still simulated. Ranking never sees score.

One attempt,20workers,width6,depth6,max600decisions,80000000physical frame
evaluations,3600seconds search wall budget, progress watchdog40committed
decisions without a new unique causal credit. No reset on delayed counter ack.
Each worker maximum prefix+160frames, no save/load/RAMwrites/ROMchanges.
Deadline07:00UTC; start only before05:30UTC so registered verification fits.
Search never overlaps E020 or Rack diagnostic workers. The ordinary bootstrap
remains MAME0.289/pengoja,DIPs176/204. No Rack-derived tape can seed this trial.

Each macro scans all newly observed frames for death/phase/life fault. Branches
with unknown attribution or incompatible state/controller result are rejected.
Retained causal events store board at decision, player, block, stance, actions
and all contact/confirmation/score/disappearance/retirement/ack clocks. Stop at
first clean counter0 and retain only that prefix, charging unused worker suffix.
Success remains CANDIDATE until fixed1800neutral frames show ACT4entry and two
fresh native reset reproductions verify exact outcome and event clocks. No ACT4
planning in this configuration; completion reading is separate from ACT4 play.

Progress timeout, no safe branch or normal compute limit is NUL. Genuine
infrastructure failure is INCOMPLETE and preserves partial cost. Human stop is
separate. No automatic restart or extra attempt; later new configurations need
new preregistration. A failure here does not cancel the user's overnight work.
Public artifacts omit ROM, model weights and raw RAM. Publish null and success
outcomes with scoped trace identity; complete-machine identity remains UNKNOWN.