door-center-shiftのcross-wall層。良い候補はあるが選べない。door-center-shift cross-wall: a good candidate exists but is missed.
全10条件で+5ptの実効ゲートに未到達。None of 10 cells reaches the +5-point practical gate.
診断は成立したが、この線形介入は選択だけで回収できる余地を埋めない。The diagnosis holds; this linear intervention does not close the selection headroom.
世界モデル、潜在表現、rollout、候補バンク、checkpoint、成功と失敗を先に定義する。Define world model, latent, rollout, candidate bank, checkpoint, success, and failure first.
専門語を処理の順番に並べて読む。Read each term in processing order.凍結した世界モデルの想像軌道を座標へ復号し、既知の壁・扉との違反量を潜在コストへ足す。Decode imagined paths, check known geometry, and add risk to latent cost.
候補生成は変えず、選択だけを変える。Generation stays fixed; only selection changes.予測軌道と実行軌道を重ね、成功候補が在るのに別候補を選ぶ過程を確認する。Overlay predicted and executed paths to see how a good candidate is missed.
想像上の近さと実行結果がずれる。Imagined closeness diverges from execution.128万候補の実行結果を、条件・モデル・配置(同側/壁越し)・選択器ごとに集計する。Aggregate 1.28M executions by condition, model, topology, and selector.
最大58.8%が順位付け失敗。Ranking failure peaks at 58.8%.現在位置と壁を描いた開始画像、目標位置を描いた目標画像。A start image showing current state and walls, plus a goal image.
符号器は画像を384個の数値へ変換する。その数値列が潜在表現で、画像そのものではない。The encoder converts an image into 384 numbers. That vector is the latent representation, not an image.
現在の潜在表現と行動列から、将来の潜在表現を順に予測する。1候補を内部で想像する処理がrollout。Predicts future latents from the current latent and actions. A rollout is one internally imagined candidate future.
潜在表現から(x,y)座標を読む小さな変換器。200個の座標ラベルだけで学習する。A small decoder from latent to (x,y), trained with only 200 coordinate labels.
通常は目標に近い想像を選ぶ。PCCは壁・扉違反を加点し、候補を作り直さず順位だけを変える。Baseline selects imagined goal proximity. PCC adds wall/door risk and changes ranking, not generation.
通常の成功率が教えるのはここまで。失敗原因はまだ一つに決まらない。Ordinary task success stops here; the cause remains ambiguous.
候補集合に成功経路が無かったのか。あったのに世界モデルが別候補を選んだのか。Was no successful route generated, or was one present but left unchosen?
各候補の真の成否を得て、coverage と ranking failure を同じ候補銀行上で確定する。Ground-truth every outcome and identify coverage and ranking failure on one shared bank.
counterfactual_audit/PROTOCOL.md · exact identity: 1 − selected success = generation failure + ranking failure.環境は65×65で描画し、モデル入力時に224×224へリサイズ。開始画像と目標画像を符号化する。Rendered at 65×65, resized to 224×224; start and goal images are encoded.
1低水準行動で位置は 2a だけ移動する。壁に当たると環境側で投影される。One low-level action moves by 2a; wall contact is projected by the environment.
6マクロ時刻 × 5低水準行動 × 2軸 = 合計30ステップ。各話に最大256系列。Six macro steps × five 2-D low-level actions = 30 steps; up to 256 sequences per episode.
偶数話は同じ側、奇数話は反対側。壁越しでは扉を通る必要がある。Even episodes stay on one side; odd episodes cross and must use the door.
成功は終端位置だけで判定。衝突の有無は別の監査列として記録する。Success uses terminal position only; collision is logged separately.
PCC checker は (32,30,4) に固定。条件変更時もcheckerを変えず、ずれへの耐性を測る。The PCC checker stays at (32,30,4), even under shifted conditions.
wall_gym_wrap.py::DEFAULT_CFG · DotWall._generate_transition · run_official_audit.py::{make_env, execute_candidate}層のつなぎ方と計算手順。いわばモデルの「設計図」。Layer structure and computations: the model blueprint.
入力224×224、384次元、予測器種別などの選択値。Choices such as 224×224 input, 384-D features, and predictor type.
訓練で更新された数百万個の数値を保存した .pth.tar。A .pth.tar containing millions of numbers learned during training.
load_model() で組み合わせると、著者が学習した時点の世界モデルを再現できる。チェックポイント単体は実行プログラムでもデータセットでもない。load_model() combines all three to reproduce the authors' trained world model. A checkpoint alone is neither an executable program nor a dataset.jepa_wm_wall.pth.tar約202 MiB · SHA-256 8efb0623… · JEPA-WM用≈202 MiB · SHA-256 8efb0623… · for JEPA-WMdino_wm_wall.pth.tar約257 MiB · SHA-256 ff170be5… · DINO-WM用≈257 MiB · SHA-256 ff170be5… · for DINO-WMpred_type: AdaLN、行動をtoken条件付け、学習時2段rollout。pred_type: AdaLN, token action conditioning, two training rollouts.pred_type: dino_wm、行動をfeature条件付け、学習時1段rollout。pred_type: dino_wm, feature action conditioning, one training rollout.hubconf.py::_MODEL_CONFIGS · released Wall configs and checkpoints at pinned commit 13cf1d9c.Wall上の真の (x,y) と行動を生成元にする。True Wall (x,y) state and actions form the source system.
直交混合を加え、座標をそのまま使わない。Orthogonal nonlinear mixing prevents direct coordinate access.
8,000 offline遷移で、行動条件付きlatent遷移を学習。Learns action-conditioned latent transitions from 8,000 offline samples.
想像軌跡を座標へ戻し、壁外横断を選択コストへ加える。Decode imagined paths and add outside-door crossings to selection cost.
EXPERIMENT_CARD.md · run_experiment.mjs · controlled result only; claim boundary retained.V-JEPA 2-ACとDINO-WMは、想像終端の目標近接度で候補を順位付けする。V-JEPA 2-AC and DINO-WM rank candidates by imagined terminal proximity.
RC-aux・FF-JEPA・Value-Guidedは、到達可能性・階層サブゴール・価値整形で潜在距離の弱さを補う。RC-aux, FF-JEPA, and Value-Guided address latent-distance limits through reachability, hierarchy, or value shaping.
iKCEは運動学的一貫性を測る。本研究は既知の壁幾何との不一致を候補ごとに測り、選択失敗を監査する。iKCE checks kinematic consistency; we check each candidate against known wall geometry and audit selection failure.
65×65で描画し224×224へ拡大したWall画像と目標を入力。Input Wall images rendered at 65×65 and resized to 224×224, plus the goal.
JEPA-WM / DINO-WMの公開重みを固定。Freeze the public JEPA-WM / DINO-WM weights.
同一の最大256行動候補を想像する。Imagine the same bank of up to 256 action sequences.
200ラベルで座標復号し、壁違反を選択コスト化。Decode coordinates from 200 labels and score wall violation.
真値simulatorで全候補を実行し、反実仮想を確定。Execute every candidate in the truth simulator.
run_official_audit.py: checkpoint load → rollout → probe/PCC score → selector → exhaustive simulator audit.良い候補が無いのか、在るのに選べないのか。No good candidate, or a good candidate left unchosen?
1−成功率を生成失敗と順位付け失敗へ厳密分解する。Exactly decompose 1−success into generation and ranking failure.
世界モデルを再学習せず、実行前に候補を再順位付けする。Rerank before execution without retraining the world model.
全10条件で事前登録した改善ゲートに未到達。No cell passes the preregistered improvement gate.
故障構成と予測誤差プロファイルまで比較する。Compare failure composition and prediction-error profiles.
topologyと候補被覆で層別し、両モデルの差を測る。Stratify by topology and coverage on both checkpoints.
seed・episode・行動tensorを保存し、SHAハッシュで同一性を検証。Save seed, episode, and action tensor; verify identity by hash.
凍結WMで各候補をrolloutし、潜在距離と復号軌道を得る。Roll out each candidate with the frozen WM.
latent / PCC / oracle / random
違うのは候補indexだけ。Only the selected index changes.
選択外も含む256候補を真値simulatorで実行し、成功vectorを保存。Execute all 256, including unselected candidates.
想像終端と目標の潜在距離。候補bank内5–95%分位で正規化し[0,2]へclip。Latent goal distance, normalized by within-bank 5–95% quantiles and clipped to [0,2].
復号軌道が扉外で壁中心線を横切る最大逸脱。時間方向はmaxで集約。Maximum decoded outside-gap wall crossing; time is collapsed by max.
僅かな非零違反にも固定penalty。score最小の候補を実行する。A fixed penalty for any nonzero violation; execute the minimum-score candidate.
run_official_audit.py selection rule: score = ν(C) + λν(R) + 0.25λ·1[R>1e−5].jepa_wm_walldino_wm_wall
同一Wall課題、公式公開重み。Same Wall task, public weights.
matched / door shift / narrow door / action noise / truth-only obstacle
latent / PCC / oracle / random
同じ候補bankからindexを選択。Choose indices from the same bank.
{16, 32, 64, 128, 256}
小予算は同一bankのprefix。Smaller budgets are prefixes.
SI_RESULTS_GENERATED.md / frozen commit 13cf1d9c7e476f53c17714d2e0f1dc239a883ce0run_official_audit.py condition constructors and frozen checker constants.| 仮説Hyp. | 内容Statement | 凍結閾値 (2026-07-19・実行前)Frozen thresholds (2026-07-19, pre-run) | 判定 (公式)Decision (official) |
|---|---|---|---|
| H1 | 機構シフト: 失敗の主因が条件で移るMechanism shift across conditions | latent選択・予算256で matched の順位付けshare ≥0.50 かつ (truth-only生成share − matched生成share) ≥+0.20。両層とも失敗 n≥50 (チェックポイント毎)。不足= indeterminate / 推定可能な閾値未達= not_supported。集約支持は両チェックポイント必要At latent selection, budget 256: matched ranking-failure share ≥0.50 and (truth-only generator share − matched generator share) ≥+0.20; ≥50 failures per stratum, per checkpoint. Missing = indeterminate; estimable failure = not_supported; aggregate needs both checkpoints | indeterminate (matched失敗 n<50) = 証拠なしindeterminate (n<50) = no evidence |
| H2 | 標的介入: PCC再順位付けが成功率を改善Targeted reranking improves success | 生成失敗差が 1e−12 未満で不変 かつ 順位付け失敗が厳密減少 かつ 利得 ≥+0.05 かつ Holm補正 正確McNemar p≤0.05 (10確証対比)。集約判定は両モデルの matched 条件を用いるGenerator-failure delta strictly below 1e−12 and ranking failure strictly decreases and gain ≥+0.05 and Holm-adjusted exact McNemar p≤0.05 (10 confirmatory contrasts). Aggregate uses matched for both checkpoints | not_supported (全10セル)not_supported (all 10 cells) |
| H3 | 集約交絡: 層別で材料差が出るAggregation confounding under strata | latent選択・予算256。topology 2群と被覆 2群が存在し全群 n≥50。推定可能な成功率/AUROCの max−min gap ≥0.10 で支持。単一クラス層のAUROCはNA可。集約は両チェックポイント必要Latent, budget 256; both topology and both binary-coverage groups exist with n≥50 each; any estimable success/AUROC max−min gap ≥0.10 supports; one-class AUROC may be NA; aggregate needs both checkpoints | supported (両モデル)supported (both models) |
| H4 | モデル別プロファイル: 成功率類似でも故障/誤差が分離Checkpoint-specific profiles | matched・latent・予算256・各モデル n=500。成功|Δ|≤0.05 かつ (生成/順位付け故障率gap ≥0.10 または endpoint/path予測誤差gap ≥0.05)。点閾値診断であり同等性検定ではないMatched, latent, budget 256, n=500/checkpoint. |success gap|≤0.05 AND (generator/ranking failure gap ≥0.10 or endpoint/path mismatch gap ≥0.05). A point-threshold diagnostic, not an equivalence test | supported (誤差軸のみ・故障率軸は未達)supported (mismatch axis only) |
復号軌道は扉を通過して目標へ向かうが、実行軌道は扉の下端で衝突し、その場に留まる。The decoded path passes the door, but execution collides at the lower edge and stalls.
潜在コストはわずかに高いが、実行では扉を抜け、終端距離4.5未満を満たす。Its latent cost is slightly worse, yet execution clears the door and meets the 4.5 threshold.
seed_8101_matched_counterfactuals.npz + seed_8101_per_candidate.csv. 説明用の実例であり、代表性は集計スライドで担保する。Illustrative episode; prevalence is established by the aggregate slides.| Model / condition | DINO generator | DINO ranking | DINO success | JEPA generator | JEPA ranking | JEPA success |
|---|---|---|---|---|---|---|
| action-noise | 0.8% | 13.2% | 86.0% | 0.8% | 12.8% | 86.4% |
| door-center-shift | 6.8% | 56.8% | 36.4% | 6.8% | 58.8% | 34.4% |
| matched | 4.4% | 8.0% | 87.6% | 4.4% | 5.6% | 90.0% |
| narrow-door | 20.4% | 31.2% | 48.4% | 20.4% | 31.6% | 48.0% |
| truth-only-obstacle | 15.6% | 10.0% | 74.4% | 15.6% | 7.6% | 76.8% |
decomposition.csv, method=latent, topology=cross-wall, budget=256, n=250 per cell.1 − generator failure
bank内に成功候補がある確率。Probability the bank contains a success.
現行の潜在argminが選ぶ成功率。Success selected by latent argmin.
coverage ceiling − latent success
選択器だけで理論上回収可能。Theoretically recoverable by selector.
decomposition.csv, cross-wall latent rows; bar scale maximum = 60 percentage points.中心 y=30、半幅4。復号軌道は x=32 を約 y=28.5 で横切るため、実装の pcc_risk() は0を返す。Center 30, half-width 4. The decoded crossing is near y=28.5, so pcc_risk() returns zero.
中心だけを y=36へ移動し、モデルとPCCには知らせない。図の灰色壁と緑の開口が真値。Only truth moves to center 36; model and PCC are unchanged. Gray walls show the true opening.
candidate #14は risk 0 だが衝突し、終端距離は 6.58。成功閾値4.5を満たさない。Candidate #14 has risk 0, collides, and ends 6.58 away—outside the 4.5 success threshold.
pcc_risk() in run_official_audit.py + official NPZ/CSV. これは「PCCは未知の幾何を発見する器ではない」という採用範囲を具体化する。This concretizes the scope: PCC is not a detector for unknown geometry.| model | 条件condition | latent | PCC | Δ | 95% CI | 改善/悪化impr./harmed | Holm p |
|---|---|---|---|---|---|---|---|
| dino | action-noise | 0.930 | 0.910 | −0.020 | [−0.040, −0.002] | 8/18 | 0.680 |
| dino | door-center-shift | 0.682 | 0.662 | −0.020 | [−0.044, 0.006] | 20/30 | 1.000 |
| dino | matched | 0.938 | 0.938 | 0.000 | [−0.014, 0.020] | 6/6 | 1.000 |
| dino | narrow-door | 0.742 | 0.716 | −0.026 | [−0.034, −0.018] | 13/26 | 0.533 |
| dino | truth-only-obstacle | 0.812 | 0.816 | +0.004 | [−0.012, 0.024] | 7/5 | 1.000 |
| jepa | action-noise | 0.932 | 0.926 | −0.006 | [−0.022, 0.008] | 9/12 | 1.000 |
| jepa | door-center-shift | 0.672 | 0.666 | −0.006 | [−0.022, 0.010] | 16/19 | 1.000 |
| jepa | matched | 0.950 | 0.944 | −0.006 | [−0.014, 0.002] | 3/6 | 1.000 |
| jepa | narrow-door | 0.740 | 0.752 | +0.012 | [−0.030, 0.052] | 28/22 | 1.000 |
| jepa | truth-only-obstacle | 0.824 | 0.822 | −0.002 | [−0.014, 0.016] | 6/7 | 1.000 |
paired_contrasts.csv, n=500 pairs per row. Bars scale to 30 discordant episodes.success 30.48pt / AUROC 0.367
凍結閾値0.10を超える。Exceeds frozen 0.10 thresholds.
success 31.04pt / AUROC 0.437
両モデルでH3 supported。H3 supported for both models.
hypotheses.json, H3 pooled across five conditions; n=1,250 per topology and model.成功率は近い (jepa 0.950 / dino 0.938) 一方、終端誤差差0.366・経路誤差差0.187で、予測誤差プロファイルは一致しない。Success is similar (0.950 vs 0.938), but endpoint and path error differ by 0.366 and 0.187.
matchedの失敗数が情報ゲート50件に届かない (jepa 25 / dino 31)。機構シフトの有無は追加データで判定する。Matched failures miss the 50-case information gate (25 / 31); more data are needed to judge mechanism shift.
hypotheses.json, H4 matched / latent / maximum budget, n=500 per model.閾値20ptに僅かに未達、matched n=25。Just below 20pt threshold; matched n=25.
効果閾値は通るがmatched n=31<50。Effect clears threshold, but matched n=31<50.
hypotheses.json, H1 failure-composition diagnostic.200ラベルの線形probeが、扉付近の軌道を十分に復号できない可能性。A 200-label linear probe may miss near-door trajectory details.
最大逸脱だけでは、一瞬の誤差と持続的な壁抜けを同じ尺度へ潰す。Max collapse treats a transient error like persistent wall crossing.
bank内5–95%正規化と単一λが、条件間で同じ意味を持たない可能性。Within-bank scaling and one λ may not calibrate across conditions.
結果として実行可能な迂回路を過罰し、別の失敗候補を選ぶ。The selector may over-penalize feasible detours and choose another failure.
難しいcross-wall層では、良い候補が無いことより、在る候補を選べないことが最大の失敗成分だった。In hard cross-wall strata, failing to select an available good candidate is the largest component.
線形PCCはそのheadroomを成功率へ変換できず、H2はnot supportedとなった。Linear PCC does not convert that headroom into success; H2 is not supported.
選択だけで回収できる余地が最大の層へ非線形リランカを当て、複数環境と強い比較手法で診断の再現性を確かめる。Target the largest-headroom stratum with a nonlinear reranker, then test diagnosis across environments and stronger baselines.
hypotheses.json + paired_contrasts.csv + frozen decision contract schema v3.候補が無いのか、在るのに選べないのかを分ける必要がある。We must separate no-good-candidate from good-but-unchosen.
公式2モデル・5条件・各500ペアで、生成と順位付けを同じ候補上で比較する。Compare generation and ranking on shared banks across two models and five conditions.
一方、線形PCCの効果は−2.6〜+1.2ptで改善ゲートに届かない。Linear PCC stays at −2.6 to +1.2 points and misses the improvement gate.
Wallシミュレーションと線形PCCの範囲を明示し、改善を過大に主張しない。State the Wall/linear-PCC scope and avoid overstating improvement.
door-center-shiftのcross-wall。順位付け失敗はdino 56.8%、jepa 58.8% (各n=250)。door-center-shift cross-wall: ranking failure is 56.8% / 58.8% (n=250 each).
復号軌道と潜在特徴を入力し、線形PCCが拾えなかった衝突信号を学習する。Use decoded trajectories plus latent features to learn signals missed by linear PCC.
検証データだけで学習し、+5pt・Holm有意・生成不変・順位付け減少を実行前に固定する。Train on validation only; pre-freeze +5pt, Holm, generation invariance, and ranking reduction.
commit 13cf1d9cと公式checkpointのSHA-256を照合。Pin commit 13cf1d9c and verify official checkpoint hashes.
行動ハッシュを一致させ、潜在選択とPCCを対応比較。Match action hashes for paired latent-vs-PCC comparisons.
2モデル×5条件×500ペア×256候補を全件答え合わせ。Ground-truth 2×5×500×256 candidates.
分解恒等式・全シード・CI・対応検定を機械検証。Verify decomposition, seeds, CIs, and paired tests mechanically.