Pengo E004 review: causal hypothesis and the reader-only test
6 October 2026. Original protocols, results, videos and source contracts are preserved. These corrections qualify the interpretation and govern new trials.
| Review | Response and evidence |
|---|---|
| Y1 | The claim that the bottleneck shifted from survival to planner strategy is a hypothesis. Observations establish zero detected death in the retained segment and25consecutive decisions without official reduction. They do not isolate algorithm quality from the invisible egg targets: eggs were UNKNOWN at all1290E003 ACT3 observations. |
| Y2 | PENGO-E020 is the reader-only counterfactual. Same6017reset prefix, original E003 planner/heuristic/action order, original oracle_v2 controller, width4/depth4/20workers and all original budgets/watchdog. Only egg observation changes. E018 and the causal heuristic are excluded. The redesigned E005 is deferred. |
| Y3 | All18 original arithmetic mismatches have concrete state explanations; definitions and frame sets are below. Their native contexts were reviewed. This is internal RAM consistency, not independent pixel-reader accuracy. |
| Y4 | Capacity is measured from each act’s initialized native header and full coordinate table. Baseline and Rack diagnostics agree ACT1=6,ACT2=6,ACT3=8. Rack v1 had zero admissions, v2 reproduced onlyACT1..3 within the fixed30000frames/replica. ACT4+ remains UNKNOWN; this later-act validation request is still open. No Rack tape can seed E020 or any game result. |
| Y5 | Causal candidate contact is6778; native confirmed-crush animation begins6783; official acknowledgement is6895. The event record retains all three, with contact as the onset. A contact alone is not automatically a confirmed kill. Ack gives no second credit. E020 retains the original counter-based watchdog to avoid a second intervention. |
| Y6 | Future no-progress stops are NUL for the registered configuration. Normal budget exhaustion, no-safe-branch and unsupported representation also fail that configuration. Infrastructure interruption is separately INCOMPLETE. Historical E003/E010 outcomes keep their original frozen labels with this forward classification stated. |
| Y7 | W1–W6 are already applied in the separate current editorial v2, public at z0bCW2dLh9w. The original MP4s are unchanged and nonpublishable. Current SHA256c8d0bdde79d0cb48c851a8cbc3f87dd115d554e4237203e93985f79e18ab4520; first frame and postdeath7324 were re-inspected for this review. |
What the original mismatch comparison measured
The original E004 v2 arithmetic was:
number of nonzero actor raw-state slots (four slots) + valid nonzero egg coordinates (eight observed ACT3 slots)
compared with native official objective counter at program-space0x8d98 in the same observation. It did not compare with pixels, independent object labels, the pixel-agent extractor or the egg header. Native replay equivalence supports the reliability of those recorded observations; it does not make the arithmetic an independent semantic ground truth.
| Corpus/frame | Original sum vs counter | Native event | Reconciliation |
|---|---|---|---|
| E0036753..6763,11frames | 3populated actor slots+3table entries=6 vs7 | Egg entry6 temporarily zero while its ice moves from(7,9) downward; moving ice position is independently consistent in object/sprite observations. Header stays4. Entry returns at(7,11) on6764. | Add the one transported egg, not a newly created egg or kill. |
| E0096468..6473,6frames | 3populated slots+3table entries=6 vs7 | Entry6 is absent while its ice travels from(7,9) upward; header stays4. It returns at(7,8) on6474. | Add the transported egg once. |
| E0096432,1frame | 3populated slots+5table entries=8 vs7 | An already crushed actor changes from points/raw12 to retirement/raw13. The following frame6433 starts a replacement hatch/raw15 at(3,13) and removes that egg coordinate. | Exclude the retired old generation at6432. A nonzero slot is not necessarily a remaining enemy. |
Thus all18 original discrepancies are explained at the observed update boundaries, with no new simulation. The eight native rendered examples show block motion, crush-points and hatch context. The egg inside an ordinary ice block is not directly visible in these images; its identity requires offline table history. This does not establish pixel-agent egg-memory performance.
Two additional header-update boundaries
The new strict reader admission found E0096501 and6895: the egg coordinate removal precedes native egg-header reduction by one frame. These are not the seven original arithmetic mismatches, because that comparison did not use the header. The opt-in reader retains UNKNOWN on those two frames and never zero. Its frozen validation admits1290/1290 known E003 and1280/1282 known E009 frames, including both transport sequences; all non-egg phase/state fields are unchanged. First overstrict validation failure and its separate corrected validator are retained. Later-act capacity is not inferred from this ACT3 test.
Interpretation of the counterfactual
If reader-only E020 completes, it supports a causal contribution from the egg representation under this one frozen start/configuration. If it stagnates again, the reader fix alone was insufficient. That does not prove the planner is the only cause: controller limits, residual update boundaries and finite search budgets remain possible contributors. These are predeclared interpretation limits, not a post-result reinterpretation or silent parameter change.
9,790 remains best verified planner trajectory at tape end,9,390last alive, death7299,ACT3incomplete.9,090 is the historical E003 observation.304,000 remains a numeric target; no new score search, diamond assumption or leaderboard claim.
Evidence: data/pengo/e004/review-y3-mismatches-v1; data/pengo/e019/rack-egg-capacity-v1
and v2; data/pengo/e020/egg-reader-admission-v1/failure.json and v1-1/result.json;
docs/pengo/PENGO-VIDEO-9790-REVIEW-v2.md and publication receipts.
Derived mismatch evidence · Reader admission · Retained review response · Current corrected video.