Smart
Robots
Design DEPARTMENT OF ENGINEERING
TOHOKU UNIVERSITY
SRD

良い候補は在るのに、
なぜ世界モデルは選べないのか
Why does a world model miss
good candidates that already exist?

公式JEPA-WM / DINO-WMで故障を分解し、PCC再順位付けを検証Failure decomposition and PCC reranking on official JEPA-WM / DINO-WM checkpoints
Tohoku University
奥山 彪太郎 / Kotaro OKUYAMA
東北大学 工学研究科 ロボティクス専攻 修士1年
平田・董研究室 / Smart Robots Design Lab. 2026-08-13
M1, Dept. of Robotics, Graduate School of Engineering, Tohoku University
Hirata & Dong Lab / Smart Robots Design Lab.  2026-08-13
Q何を調べ、何が分かり、次に何を試すのか?What did we test, learn, and decide to try next?

失敗の主因は選択にあり、線形PCCでは改善できなかった。Selection is the main failure source, and linear PCC did not fix it.

256候補をすべて実行して答え合わせすると、難しい壁横断の改善余地は候補生成より選択に集中していた。Ground-truthing all 256 candidates shows that, in hard wall crossing, improvement headroom lies in selection rather than generation.
最大の順位付け失敗PEAK RANKING FAILURE58.8%

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.

線形PCCの対応効果LINEAR-PCC EFFECT−2.6〜+1.2pt

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.

仮説判定DECISIONH2
not supported

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.

次の判断: 選択だけで回収できる余地が最大の層を標的に、潜在特徴も使う非線形リランカを事前登録して検証する。Next decision: preregister and test a nonlinear reranker using latent features on the largest-headroom stratum.
Q初めて読む人は、どの順番で何を理解すればよいのか?What should a first-time reader understand, and in what order?

用語から実装、1エピソード、全体集計へ進めば、結論の根拠を途中で飛ばさず追える。Terms → implementation → one episode → aggregate results makes every inference traceable.

01 / DEFINITIONS

用語と成功条件Terms and success

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.
02 / IMPLEMENTATION

PCCは何を見るかWhat PCC sees

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.
03 / EPISODE

1話で何が壊れるかInspect one episode

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.
04 / AGGREGATE

全体でどれほど起きるかMeasure prevalence

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%.
数式と統計は結論を飾るためではなく、各判断を第三者が再計算できるように後段で示す。Equations and statistics appear later so a third party can recompute each decision—not as decoration.
Part 1 背景と設計
なぜ失敗を「候補なし」と「選び損ね」に分けるのか
Part 1  Background & Design
Why split failure into no-candidate and mis-selection?
01
Qこの発表を読む前に、どの用語をどう理解すればよいか?Which terms must be understood before reading the experiment?

専門語はすべて、「画像から将来を想像し、256候補から1つを選ぶ」処理の部品である。Every technical term names one part of the path from images to one selected plan among 256 candidates.

OBSERVATION観測画像Observation

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.

ENCODER / LATENT符号器/潜在表現Encoder / latent

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.

WORLD MODEL / ROLLOUT世界モデル/rolloutWorld model / rollout

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.

LINEAR PROBE線形probeLinear probe

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.

SELECT / PCC選択/PCCSelection / PCC

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.

候補/候補バンクCandidate / bank
候補は30ステップの行動計画1本。候補バンクは、同じ開始・目標に対して比較する最大256本の集合。A candidate is one 30-step action plan; the bank is the shared set of up to 256 plans.
被覆/順位付け失敗Coverage / ranking failure
被覆は「成功候補が1本以上ある」こと。あるのに選べなければ順位付け失敗、1本もなければ生成失敗。Coverage means at least one candidate succeeds. Missing it is ranking failure; having none is generation failure.
反実仮想実行Counterfactual execution
実際には選ばなかった候補も環境で実行し、「もし選んでいたら成功したか」を答え合わせすること。Execute even unselected candidates to learn what would have happened had each been chosen.
Qなぜ潜在コストだけでは実行可能性を保証できないのか?Why does latent cost alone fail to guarantee feasibility?

世界モデルは、壁を抜ける“近道”を良い候補として選びうる。A world model can prefer a goal-close shortcut that crosses the wall.

  • 潜在世界モデルは、想像ロールアウトの目標への近さで候補行動列を順位付けして実行Latent world models rank candidate action sequences by goal proximity of imagined rollouts
  • しかし「潜在空間で目標に近い」≠「実行可能」— 壁を突き抜けて終端だけ目標に近い候補を「良い」と誤評価しうるBut "close in latent space" ≠ "feasible" — a wall-crossing candidate that merely ends near the goal can be scored as "good"
  • そこで本研究は、全候補を実行して「候補なし」と「選び損ね」を分け、既知幾何による再順位付けが選択だけを改善するかを測る。We therefore execute every candidate to split generation from mis-selection, then test whether known-geometry reranking improves selection alone.
Wallタスク: 潜在コストだけ見ると… gap START GOAL 候補A: 壁を横断 / crosses wall 終端はGOAL近く = コスト小 候補B: ドア経由・実行可能 / feasible 潜在コスト(終端の目標距離)だけでは A と B を区別できない Terminal-proximity cost alone cannot tell A (infeasible) from B (feasible)
Qなぜ成功率や予測誤差だけでなく、全候補を実行する実験が必要なのか?Why execute every candidate instead of reporting only success or prediction error?

失敗を「解ける候補がない」と「解ける候補を選べない」に分けるには、選ばなかった候補の反実仮想実行が必要である。Separating missing candidates from mis-selection requires counterfactual execution of the unchosen candidates.

OBSERVATION

選んだ計画が失敗The selected plan fails

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.

TWO POSSIBLE CAUSES

生成か、順位付けかGeneration or ranking?

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?

IDENTIFY BY EXECUTION

256候補を全て実行Execute all 256

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.

Wallを選ぶ理由 1Why Wall · 1壁と扉の幾何が既知で、経路が通れるかを線分ごとに明示できる。Known wall geometry makes path feasibility explicit segment by segment.
Wallを選ぶ理由 2Why Wall · 2256候補を毎回実行でき、選ばれなかった候補まで真値で監査できる。All 256 candidates are cheap enough to execute, including unchosen ones.
Wallを選ぶ理由 3Why Wall · 3同側/壁越し、扉移動、狭窄、行動雑音を一要因ずつ制御できる。Topology and shifts can be varied one factor at a time.
したがって本実験は、汎用ロボット性能を誇示するベンチマークではなく、世界モデル計画の故障箇所を同定する診断実験である。This is therefore a diagnostic experiment for locating world-model planning failures, not a claim of general robot performance.
counterfactual_audit/PROTOCOL.md · exact identity: 1 − selected success = generation failure + ranking failure.
QWall課題では何を成功・失敗として観測するのか?What counts as success and failure in the Wall task?

評価は、壁を避けて扉を通れる候補を選べるかを問う。The task asks whether the planner selects a route through the gap.

Controlled trajectory illustration (synthetic, NOT OFFICIAL)
概念図:壁と扉の評価設定Concept: wall-and-door evaluationdirect route vs. feasible door route
評価の役割
候補の生成と選択を分けて観測する。
Role in evaluation
Observe generation and selection separately.
候補選択の概念図。 灰色破線 = 壁を突き抜ける direct 候補、紫実線 = ドア経由の候補。評価では各エピソードに256候補を用意し、全候補を実行して「生成失敗」と「順位付け失敗」を分ける。Candidate-selection concept. Gray dashed = direct candidate crossing the wall; purple solid = feasible door route. The evaluation executes all 256 candidates per episode to separate generator failure from ranking failure.
Q公式Wall実験では、モデルに何を見せ、何を動かし、何を成功と数えるのか?What does the official Wall experiment observe, control, and count as success?

65×65の俯瞰画像から2次元点を30ステップ動かし、終端が目標から4.5未満なら成功とする。A point moves for 30 low-level steps from a 65×65 top-down image; success is terminal distance below 4.5.

START GOAL door center y=30open when |y−30| ≤ 4 wall center x=32 · width=665×65 continuous-coordinate world (schematic)
OBSERVATIONRGB俯瞰画像Top-down RGB

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.

STATE / ACTION状態 (x,y)、行動 [-1,1]²State (x,y), action [-1,1]²

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.

HORIZON6×5×2 の行動テンソル6×5×2 action tensor

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.

TASK TOPOLOGY同側/壁越しを半数ずつSame-side / cross-wall

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.

SUCCESS終端距離 < 4.5Terminal distance < 4.5

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.

FROZEN GEOMETRY壁32、扉30、半幅4Wall 32, door 30, half-gap 4

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}
Qチェックポイントは何のファイルで、「公式」「凍結」とは何を意味するのか?What is a checkpoint file, and what do “official” and “frozen” mean?

チェックポイントは、学習で決まった大量の数値「重み」を保存し、学習済みモデルを再現するためのファイルである。A checkpoint stores the learned numerical weights needed to reconstruct a trained model.

SOURCE CODEコードCode

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.

CONFIG設定Configuration

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.

CHECKPOINT学習済み重みLearned weights

訓練で更新された数百万個の数値を保存した .pth.tarA .pth.tar containing millions of numbers learned during training.

3つを 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.

今回読み込んだ2ファイルThe two files loaded here

jepa_wm_wall.pth.tar約202 MiB · SHA-256 8efb0623… · JEPA-WM用≈202 MiB · SHA-256 8efb0623… · for JEPA-WM
dino_wm_wall.pth.tar約257 MiB · SHA-256 ff170be5… · DINO-WM用≈257 MiB · SHA-256 ff170be5… · for DINO-WM
公式 = 論文著者が公開したファイルを使う、という出所の意味。正しさ・最良性能・認証済みを意味しない。
凍結 = 監査中にこの重みを更新しない、という操作の意味。
採用理由 = 再学習やモデル選択の自由度を排し、公開モデルそのものを同条件で検査できる。
SHA-256 = ダウンロードしたファイルの破損・取り違えを検出する固有の指紋。
Official means released by the paper authors; it does not mean certified, correct, or best. Frozen means weights are not updated during the audit. This removes retraining/model-selection freedom and tests the released models themselves. SHA-256 is a fingerprint used to detect a corrupted or wrong download.
語と学習の境界:「公式チェックポイント」は著者配布物、「公式結果」は本監査の6条件を通過した値で、意味が異なる。世界モデルは再学習せず、200座標ラベルの線形probeと、幾何違反の重み λ だけを検証データで決める。Terms and training boundary: “official checkpoint” means an author release; “official result” means a value passing this audit's six gates. The world models are not retrained; only the 200-label probe and geometry-risk weight λ are fitted on validation data.
ENVIRONMENT.json · load_model() · pinned commit 13cf1d9c
QなぜJEPA-WMとDINO-WMを比較し、何を同じに保つのか?Why compare JEPA-WM with DINO-WM, and what is held constant?

これは「2つのJEPA」比較ではなく、同じ公式Wall基盤で遷移予測器の異なるJEPA-WMとDINO-WM基準を監査する比較である。This is not a comparison of two JEPAs; it audits JEPA-WM and its DINO-WM baseline on the same official Wall stack.

JEPA-WM
DINO-WM 基準baseline
位置づけRole
JEPA-WMs研究で提案されたWall用公式チェックポイント。Released Wall checkpoint for the proposed JEPA-WM.
同じ公式リポジトリが配布するDINO-WM比較基準。JEPAそのものではない。Released DINO-WM comparator from the same repository; not itself a JEPA.
共通の視覚表現Shared vision
凍結DINOv2 ViT-S/14、384次元表現、224×224入力。Frozen DINOv2 ViT-S/14, 384-D representation, 224×224 input.
凍結DINOv2 ViT-S/14、384次元表現、224×224入力。Frozen DINOv2 ViT-S/14, 384-D representation, 224×224 input.
主な実装差Main code difference
pred_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.
監査で固定Held fixed
同じWall課題、開始・目標、256候補、seed、PCC式、全候補実行。Same Wall task, starts/goals, 256 candidates, seeds, PCC rule, exhaustive execution.
同じWall課題、開始・目標、256候補、seed、PCC式、全候補実行。Same Wall task, starts/goals, 256 candidates, seeds, PCC rule, exhaustive execution.
採用理由 1:公式性Reason 1 · officialWall用に公開された2チェックポイントを改変せず使える。Both released Wall checkpoints can be used unchanged.
採用理由 2:比較可能性Reason 2 · comparabilityタスクと候補銀行を固定し、故障構成がモデル間で再現するかを見られる。Fixed tasks and banks reveal whether failure composition repeats across models.
採用理由 3:依存性の検査Reason 3 · dependencePCCの結論が単一予測器だけの偶然かを検査できる。Tests whether the PCC conclusion is peculiar to one predictor.
注意:両チェックポイントは複数の設計・学習差を同時に含む。したがって、個々の構成要素の因果効果を断定するアブレーションではない。Caveat: the checkpoints differ in several design and training choices; this is not a causal ablation of any single component.
hubconf.py::_MODEL_CONFIGS · released Wall configs and checkpoints at pinned commit 13cf1d9c.
Q「制御サロゲート」とは何で、なぜ公式チェックポイントより先に試すのか?What is the controlled surrogate, and why use it first?

制御サロゲートは、PCCの成立可能性を切り分ける「陽性対照」であり、公式結果の代用品ではない。The controlled surrogate is a positive control for PCC's viability—not a substitute for official-checkpoint evidence.

TRUE SYSTEM連続2次元状態Continuous 2-D state

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.

PROXY LATENT非線形8次元表現Nonlinear 8-D latent

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.

PROXY WMRidge遷移予測器Ridge transition model

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.

SAME PCC IDEA200ラベルprobe + 幾何200-label probe + geometry

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.

代理モデル+陽性対照Proxy + positive controlサロゲートは、本物を単純化した代理モデル。陽性対照は「この仕組みなら効くはず」と意図的に有利にした確認条件。ここで失敗すればPCC式自体が疑わしく、成功して公式モデルで失敗すれば、画像認識や潜在→座標復号への移行部を疑える。A surrogate is a simplified proxy. A positive control is a deliberately favorable case where the mechanism should work. Failure here indicts the PCC rule; success here but failure on released models points to vision or latent-to-coordinate transfer.
分かったことWhat it established同じ候補生成のまま成功率が 52.14% → 97.57%。PCC式には制御条件下で候補選択を直せる能力がある。With generation fixed, success rose 52.14% → 97.57%; PCC can repair selection under controlled conditions.
分からないことWhat it cannot establish画像入力、公式JEPA-WM/DINO-WM、実ロボットへの転移は証明しない。だから次に公式チェックポイントで同じ式を監査した。It does not establish transfer to visual official models or robots; that is why the same rule is then audited on released checkpoints.
EXPERIMENT_CARD.md · run_experiment.mjs · controlled result only; claim boundary retained.
Q既存の世界モデル計画に対して、何を新しく測るのか?What new quantity do we measure beyond prior world-model planning work?

本研究は予測精度ではなく、“選び損ね”を実行前に診断する。We diagnose mis-selection before execution, rather than prediction accuracy.

世界モデル計画WORLD-MODEL PLANNING

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.

計画可能性の補正PLANNABILITY

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.

実行前診断TEST-TIME DIAGNOSIS

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.

本研究の差分: 公式チェックポイント上で「良い候補が無い」と「在るのに選べない」を分け、幾何再順位付けの効果まで対応比較する。Our gap: on official checkpoints, separate “no good candidate” from “good but unchosen,” then test geometric reranking with paired comparisons.
Qどの部品を凍結し、どこだけを介入するのか?Which components are frozen, and where is the intervention applied?

世界モデルを変えず、復号軌道と幾何リスクを選択器へ加える。The world model stays frozen; decoded geometry only changes candidate selection.

観測・目標Observation & goal

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.

公式エンコーダOfficial encoder

JEPA-WM / DINO-WMの公開重みを固定。Freeze the public JEPA-WM / DINO-WM weights.

潜在ロールアウトLatent rollout

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.

線形復号+PCCLinear probe + PCC

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.

全候補監査Exhaustive audit

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.

FROZEN
encoder・world model・候補bank・checker幾何encoder, world model, candidate bank, checker geometry
LEARNED / VALIDATION ONLY
線形probeとλ(validation 40話、λ∈[0,30])linear probe and λ (40 validation episodes, λ∈[0,30])
AUDIT
実行側だけを条件変更conditions alter execution side only
run_official_audit.py: checkpoint load → rollout → probe/PCC score → selector → exhaustive simulator audit.
Q既存研究と比べて、この評価は何を同時に満たすのか?Which evaluation properties are combined here beyond prior work?

全候補実行・故障分解・凍結チェックポイントを一つの評価に統合する。The study unifies exhaustive execution, failure decomposition, and frozen checkpoints.

研究の主眼Primary focus
凍結WMFrozen WM
既知幾何Known geometry
全候補実行Execute all
生成/選択分解Gen/rank split
対応介入Paired intervention
V-JEPA 2 / DINO-WM
RC-aux / FF-JEPA / Value-Guided
方式依存varies
一部partial
iKCE / consistency objectives
方式依存varies
本研究:公式Wall監査This work: official Wall audit
✓ 256
✓ 500
比較の単位は“モデルの良し悪し”ではなく、失敗をどこまで同定できる評価設計か。The comparison is about diagnostic identifiability, not a blanket model ranking.
注: ✓は本評価で明示的に監査した項目。—は各研究の主眼ではないことを示し、非対応を断定しない。Note: ✓ marks explicitly audited properties; — means not the cited work's primary focus, not proof of absence.
Qこの研究は何を明らかにし、何を成果として残すのか?What does the study resolve, and what does it contribute?

3つの問いを、故障分解・再順位付け・モデル比較で検証する。Three questions map to failure decomposition, reranking, and model comparison.

RQ1 / WHERE失敗はどこで起きる?Where does failure occur?

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?

C1 / AUDIT全候補の反実仮想監査Counterfactual audit of all candidates

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.

RQ2 / INTERVENTION既知幾何で選択を直せる?Can known geometry fix selection?

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.

C2 / NEGATIVE RESULT線形PCCは改善を示さないLinear PCC shows no gain

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.

RQ3 / COMPARISON成功率が近いモデルは同じ?Are similarly successful models alike?

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.

C3 / STRATIFICATION集約が隠す異質性を定量化Quantify hidden heterogeneity

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.

対応する事前登録仮説: RQ1→H1・H3 / RQ2→H2 / RQ3→H4。Preregistered mapping: RQ1→H1/H3, RQ2→H2, RQ3→H4.
Q生成と順位付けのどちらが失敗したと、どう判定するのか?How do we tell generator failure from ranking failure?

全256候補を実行すれば、失敗は2成分に一意に分かれる。Executing all 256 candidates splits every failure into exactly two causes.

  • 復号する: 凍結モデルの想像潜在から、200ラベルの線形回帰で各時刻の (x, y) を推定する。Decode: use a 200-label linear probe to estimate (x, y) at each imagined step.
  • 再順位付けする: 復号軌道が扉の外で壁を横切る最大量をPCCリスクとし、潜在コストへ加える。Rerank: add the maximum outside-gap wall crossing of the decoded path as PCC risk.
  • 監査する: 選ばなかった候補も含めて全件を実行し、失敗の所在を答え合わせする。
    • 1 − 選択成功率 = 生成失敗 + 順位付け失敗
    Audit: execute even the unchosen candidates to ground-truth where failure occurs.
    • 1 − selected success = generator failure + ranking failure
凍結 世界モデル (JEPA-WM / DINO-WM) frozen — no retraining / 再学習しない 想像ロールアウト (潜在) ×256候補 imagined rollouts in latent space, per candidate 線形プローブ: 200ラベルで (x, y) 復号 linear probe decodes physical trajectory PCCスコア = ギャップ外の壁違反の最大値 max wall violation outside the gap 再順位付け: 潜在コスト + リスク項 rerank = latent cost + PCC risk term 答え合わせ: 全候補を実際にシミュレータで実行 → 生成失敗 / 順位付け失敗を分離 ground truth: every candidate executed → separates generator vs ranking failure
Q1エピソードの失敗原因を、どう監査台帳へ落とすのか?How is one episode turned into an auditable failure ledger?

共有候補を4選択器で選び、全256候補の実行結果で答え合わせする。Four selectors share one bank; all 256 executions provide the counterfactual truth.

01 / BANK候補を生成Generate bank

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.

02 / IMAGINE潜在で予測Imagine

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.

03 / SELECT4方式で選ぶSelect four ways

latent / PCC / oracle / random
違うのは候補indexだけ。Only the selected index changes.

04 / EXECUTE全件を実行Execute all

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.

1 − selected success = no successful candidate + successful candidate left unselected
台帳単位: model × condition × seed × episode × topology × budget。候補bankはlatent/PCC間で完全共有。Ledger key: model × condition × seed × episode × topology × budget; bank shared exactly by latent and PCC.
QPCCは候補選択にどう組み込まれているのか?How is PCC inserted into candidate selection?

同じ候補上で潜在コストとPCCを比べ、選択だけの効果を測る。Shared candidates isolate the selection effect of adding PCC to latent cost.

候補バンクCandidate bank
B = {a(1), …, a(256)},  a(j) ∈ [−1, 1]6×5×2マクロ6手 × 低レベル5行動。全手法で共有 (ハッシュ検証)6 macro × 5 low-level actions; shared across methods (hash-verified)
想像ロールアウトImagined rollout
1:T(j) = fθ(z0, a(j))fθ = 凍結した世界モデル (再学習しない)fθ = frozen world model (never retrained)
プローブ復号Probe decode
t(j) = gφ(ẑt(j))gφ = 200ラベルの線形回帰で潜在から座標へgφ = linear regression (200 labels) from latent to (x, y)
PCCリスクPCC risk
R(j) = maxseg max(0, |ycross − ydoor| − whalf)復号軌道が壁中心線を横切る点がドアからどれだけ外れるか。checker幾何は全条件 matched 値 (32, 30, 4) に固定 — シフトはPCCに知らせないhow far centerline crossings fall outside the door; checker geometry fixed at matched (32, 30, 4) — shifts are unseen by PCC
選択則Selection
ĵ = argminj [ ν(Clat(j)) + λ·ν(R(j)) + 0.25λ·1[R(j)>10−5] ]ν = 5–95%分位正規化を[0,2]へclip。λはグリッド{0…30}から検証40件のみで選択ν = 5–95% quantile normalization clipped to [0,2]; λ picked on 40 validation episodes only
成功・被覆Success / coverage
S(j) = 1[ ‖xT,exec(j) − xgoal‖ < 4.5 ],   C = maxj S(j)成功は終端距離のみで判定。衝突は別指標として記録success by terminal distance only; collisions recorded separately
恒等式Identity
1 − E[S(ĵ)] = P(C=0) + P(C=1 ∧ S(ĵ)=0)= 生成失敗 + 順位付け失敗。全候補をシミュレータで実行するので層別に厳密成立= generator + ranking failure; exact per stratum because every candidate is executed
QPCCは単なる衝突分類器ではなく、何を最小化しているのか?What does PCC minimize beyond a binary collision label?

正規化した潜在距離へ、最大壁違反と違反indicatorを加算する。PCC adds normalized maximum wall violation and an indicator to latent cost.

GOAL TERMν(Clatent)

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].

GEOMETRY TERMλ · ν(Rwall)

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.

VIOLATION GATE0.25λ · 1[R>10−5]

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.

λ tuning
validation 40話、grid 0–3040 validation episodes, grid 0–30
checker geometry
(x, y, half-width) = (32, 30, 4)
causal isolation
候補生成は不変、argminだけ変更generation fixed; only argmin changes
run_official_audit.py selection rule: score = ν(C) + λν(R) + 0.25λ·1[R>1e−5].
Q何を満たせば「PCCが改善した」と判定するのか?What must hold before we claim PCC improves planning?

5ポイント以上の改善に加え、偶然・多重比較・候補差では説明できない場合だけ「改善」とする。We claim improvement only when a ≥5-point gain cannot be explained by chance, multiple testing, or changed candidates.

1. 対応比較1. Paired design
(Slat, Spcc)  →  b = #(0→1), c = #(1→0)同一バンク上のペア。b=改善数、c=悪化数 (各モデル×条件 n=500ペア)pairs on the shared bank; b improved, c harmed (n=500 per model×condition)
2. 正確McNemar2. Exact McNemar
p = min(1, 2·P(X ≤ min(b,c))),  X ~ Bin(b+c, 1/2)同じ500話の前後を比べ、「改善した話」と「悪化した話」の偏りが偶然かを検定するtests whether improved versus harmed episodes in the same 500 pairs are plausibly chance
3. Holm補正3. Holm correction
p(1) ≤ α/10,  p(2) ≤ α/9, …  (α = 0.05)10通りを同時に検定すると偶然の当たりが増えるため、合格基準を順に厳しくする。95%CIは5シード単位の再標本化tightens thresholds across 10 tests to control chance hits; 95% CIs resample the five seeds
4. H2実効ゲート4. H2 practical gate
Δ ≥ +0.05 ∧ pHolm ≤ 0.05 ∧ |Δgen| < 10−12順位付け失敗↓ranking failure↓4つ全てを満たして初めて「改善」と主張できる (実行前に凍結した基準)all four must hold before claiming improvement (frozen before the run)
機械claim gateClaim gate
①重みSHA-256照合 ②スモーク分解誤差=0 ③全5シード完走 ④バンク・行動ハッシュ一致 ⑤CI・対検定の生成 ⑥null/harmful含む全数報告(1) weight SHA-256 (2) zero smoke error (3) all 5 seeds (4) bank/action hash identity (5) CIs & tests emitted (6) full reporting incl. null/harmful6条件をすべて通過した値だけが「公式」になるonly values passing all six count as official
Q結果を見た後の調整を、どう防いだのか?How did we prevent post-hoc tuning after seeing outcomes?

モデル・候補・条件・閾値を実行前に固定した。Models, candidates, conditions, and decision thresholds were frozen pre-run.

  • モデル: 公式checkpoint jepa_wm_wall / dino_wm_wall (凍結commit 13cf1d9c、ハッシュ検証済)Models: official checkpoints jepa_wm_wall / dino_wm_wall (frozen commit 13cf1d9c, hash-verified)
  • 設計: 2モデル×5シード=10シャード、5条件、各条件500ペア、候補バンク256・予算 {16,32,64,128,256}
    • 条件: matched / door-center-shift / narrow-door / action-noise / truth-only-obstacle
    Design: 2 models × 5 seeds = 10 shards, 5 conditions, 500 pairs each, bank of 256, budgets {16,32,64,128,256}
    • Conditions: matched / door-center-shift / narrow-door / action-noise / truth-only-obstacle
  • 統計: seed-clusterブートストラップ95%CI (≥2,000 draw・クラスタ=5シードのみ) + 500ペアpooled正確McNemar (Holm補正)。pooled検定のシード依存性は限界として明示。実効ゲート = +0.05利得かつHolm有意Statistics: seed-cluster bootstrap 95% CIs (≥2,000 draws; only 5 seed clusters) + pooled exact McNemar over 500 pairs with Holm. Pooled-test seed dependence is disclosed as a limitation. Practical gate = +0.05 gain AND Holm-significant
  • claim gate 6条件 (ハッシュ検証 / スモーク分解誤差0 / 全シード完了 / バンク一致 / CI・対検定生成 / null・harmfulも報告) 通過後の値のみ公式Only values passing the 6-condition claim gate (hash check / zero smoke error / all seeds done / bank identity / CIs & paired tests / report null & harmful) count as official
Q公式監査は、どの規模で反実仮想を評価したのか?How large was the official counterfactual audit?

2モデル×5条件×500話×256候補=128万候補を実行した。The audit executed 1.28 million candidates: 2×5×500×256.

2 models×5 conditions×500 episodes×256 candidates1,280,000
2 checkpoints

jepa_wm_wall
dino_wm_wall
同一Wall課題、公式公開重み。Same Wall task, public weights.

5 conditions

matched / door shift / narrow door / action noise / truth-only obstacle

4 selectors

latent / PCC / oracle / random
同じ候補bankからindexを選択。Choose indices from the same bank.

5 budgets

{16, 32, 64, 128, 256}
小予算は同一bankのprefix。Smaller budgets are prefixes.

128万は最大予算bankの一意候補数。小予算はprefix再集計なので、別bankを追加生成していない。1.28M counts unique max-budget candidates; smaller budgets reuse prefixes rather than add new banks.
SI_RESULTS_GENERATED.md / frozen commit 13cf1d9c7e476f53c17714d2e0f1dc239a883ce0
Q結果を読むために必要な指標と条件は何か?Which metrics and conditions are needed to read the results?

被覆と順位付け失敗を分ければ、成功率の内訳が読める。Separating coverage from ranking failure reveals what success rate hides.

  • 候補バンク: 最大256本のホライズン付き行動列の固定集合。全手法・全比較で共有 (行動ハッシュ一致を機械検証)。成功 = 実行後の終端と目標の距離 < 4.5 (距離基準のみ。衝突は別指標として記録・報告)Candidate bank: fixed set of up to 256 horizon-length action sequences, shared across methods (hash-verified). Success = executed terminal-to-goal distance < 4.5 (distance only; collisions recorded and reported separately)
  • 被覆 = バンク内に成功候補が≥1本ある割合 / 生成失敗 = 1−被覆 / 順位付け失敗 = 成功候補が在るのに選択候補が失敗する割合 / same-side / cross-wall = 開始と目標が壁の同側か・扉の通過が必要か (各条件250ずつ)Coverage = share of episodes with ≥1 successful bank candidate / generator failure = 1−coverage / ranking failure = successful candidate exists but selection fails / same-side / cross-wall = whether start and goal share the wall side (250 each per condition)
  • seed-clusterブートストラップ: シード(5本)を再抽出単位とし、シード内のエピソード構造を保った95%区間 (≥2,000 draw)Seed-cluster bootstrap: seeds (5) are the resampling units, within-seed episode structure retained; 95% intervals with ≥2,000 draws
  • 5条件: matched=公開設定どおり / door-center-shift=扉中心を6px移動 (シミュレータ真値側) / narrow-door=扉半幅を4→2に狭窄 / action-noise=実行時の各操作に SD 0.08 のノイズ / truth-only-obstacle=実行側のみ第2の壁 (x=48±3, 穴 y=18±4。モデルにもPCCにも非提示の限界試験)5 conditions: matched = published setup / door-center-shift = door center moved 6px (simulator truth) / narrow-door = door half-width 4→2 / action-noise = SD 0.08 execution noise / truth-only-obstacle = second wall on the execution side only (x=48±3, hole y=18±4; shown to neither model nor PCC — a designed limit test)
Q5条件は、何を変え、何をモデルとPCCから隠すのか?What changes in each condition, and what remains hidden from the model and PCC?

基準条件以外では実行環境だけを変え、モデルと固定チェッカーが変化へ追従できるかを測る。Outside the baseline, only execution changes, testing whether the model and fixed checker tolerate mismatch.

条件Condition
実行側の変更Execution-side change
model / PCCに可視?Visible to model/PCC?
診断する問いDiagnostic question
matched
(x,y,half-width)=(32,30,4)
一致matched
基準性能baseline
door-center-shift
真値の扉中心 y + 6 pxtruth door center y + 6 px
不可視hidden
幾何ずれgeometry shift
narrow-door
半幅 4 → 2 pxhalf-width 4 → 2 px
不可視hidden
許容幅低下tolerance loss
action-noise
実行行動ごとにSD 0.08SD 0.08 per executed action
不可視hidden
制御ノイズcontrol noise
truth-only-obstacle
第2壁 x=48±3 / hole y=18±4second wall x=48±3 / hole y=18±4
不可視hidden
未観測障害物unseen obstacle
各条件500話=same-side 250 + cross-wall 250。PCC checkerは常にmatched幾何(32,30,4)のまま。Each cell has 250 same-side + 250 cross-wall episodes; PCC always checks matched geometry (32,30,4).
run_official_audit.py condition constructors and frozen checker constants.
Q各仮説は、何をどの閾値で判定するのか?What does each hypothesis test, and against which threshold?

H1–H4は機構・介入・層別差・モデル差を別々に判定する。H1–H4 separately test mechanism, intervention, strata, and model differences.

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 conditionslatent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 checkpointsindeterminate (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 checkpointsnot_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 stratalatent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 checkpointssupported (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 profilesmatched・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 testsupported (誤差軸のみ・故障率軸は未達)supported (mismatch axis only)
判定契約 (スキーマv3) は公式ラン実行前に凍結。非推定・不足データは自動的に indeterminate = 「証拠なし」扱いで、支持には数えない (fail-closed)。分析器は常に全診断を出力し、都合の悪い条件も抑制しない。The decision contract (schema v3) was frozen before the official run. Non-estimable or under-populated inputs are automatically indeterminate = "no evidence", never counted as support (fail-closed). The analyzer emits all diagnostics and never suppresses unfavorable conditions.
Part 2 結果
選び損ねは大きいが、線形PCCは改善につながらない
Part 2  Results
Mis-selection is large, but linear PCC does not improve success
02
Q計画失敗の最大成分は、候補生成か順位付けか?Is the largest planning failure component generation or ranking?

難しい壁横断では、選択段階で最大58.8%を落としている。In hard wall crossing, mis-selection reaches 58.8%.

  • 難条件のcross-wall層 (latent選択・予算256):
    • door-center-shift: 生成失敗 0.068 に対し順位付け失敗 0.568–0.588
    • narrow-door: 生成失敗 0.204 + 順位付け 0.312–0.316
    Hard-condition cross-wall strata (latent, budget 256):
    • door-center-shift: generator 0.068 vs ranking 0.568–0.588
    • narrow-door: generator 0.204 + ranking 0.312–0.316
  • same-side層は成功 1.000 (truth-only-obstacleを除く) — 失敗は壁横断に幾何的に集中Same-side strata succeed at 1.000 (except truth-only-obstacle) — failures concentrate geometrically at wall crossing
  • = これらの層では良い候補は在る。測定されたheadroomの大部分は選択側= In these strata good candidates exist. The larger measured headroom lies in selection
Official Wall failure decomposition Latent-cost selection, maximum candidate budget N=256; stacked components satisfy 1-success = generator + ranking. Official Wall failure decomposition Latent-cost selection, maximum candidate budget N=256; stacked components satisfy 1-success = generator + ranking. 0% 25% 50% 75% 100% Fraction of episodes dino_wm_wall / action-noise 0.070 dino_wm_wall / door-center-shift 0.318 dino_wm_wall / matched 0.062 dino_wm_wall / narrow-door 0.258 dino_wm_wall / truth-only-obstacle 0.188 jepa_wm_wall / action-noise 0.068 jepa_wm_wall / door-center-shift 0.328 jepa_wm_wall / matched 0.050 jepa_wm_wall / narrow-door 0.260 jepa_wm_wall / truth-only-obstacle 0.176 Generator failure Ranking failure
図1: latent選択・予算256。各モデル×条件 n=500 (各topology層250)。積み上げは恒等式に厳密一致 (監査済SVG、FIGURE_MANIFEST掲載)。Fig.1: latent, budget 256; n=500 per model×condition (250 per stratum). Stacks satisfy the identity exactly (audited SVG, in FIGURE_MANIFEST).
Q「成功候補が在るのに選べない」とは、1エピソードでは何が起きているのか?What does “a good candidate exists but is missed” look like in one episode?

想像上は目標へ近い候補が扉の縁で衝突し、同じバンク内の成功候補を潜在コストが下位に置いた。The imagined shortcut collides at the door edge while a successful candidate in the same bank receives a worse latent cost.

潜在コスト最小の候補 #220Latent argmin candidate #220

latent cost 0.0108
terminal distance 15.18
endpoint error 15.74

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.

同じバンクの成功候補 #210Successful bank candidate #210

latent cost 0.0159
terminal distance 0.15
endpoint error 1.07

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.

復号した想像軌道decoded imagined path実行軌道executed path ● start ○ goal JEPA / seed 8101 / matched / episode 7 / 256候補JEPA / seed 8101 / matched / episode 7 / 256 candidates
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.
Q壁を越える課題では、条件ごとにどの故障成分が増えるのか?Which failure component grows under each cross-wall condition?

扉位置のずれでは選び損ねが、狭い扉と追加障害物では候補不足が大きく増える。Door shift drives ranking failure; narrow doors and hidden obstacles hurt coverage.

Model / conditionDINO
generator
DINO
ranking
DINO
success
JEPA
generator
JEPA
ranking
JEPA
success
action-noise0.8%13.2%86.0%0.8%12.8%86.4%
door-center-shift6.8%56.8%36.4%6.8%58.8%34.4%
matched4.4%8.0%87.6%4.4%5.6%90.0%
narrow-door20.4%31.2%48.4%20.4%31.6%48.0%
truth-only-obstacle15.6%10.0%74.4%15.6%7.6%76.8%
恒等式は各セルで成立:success + generator + ranking = 100%。色は値の大きさであり、有意差を表さない。Identity holds in every cell: success + generator + ranking = 100%. Color encodes magnitude, not significance.
decomposition.csv, method=latent, topology=cross-wall, budget=256, n=250 per cell.
Q候補生成を変えずに、選択だけで回収できる余地はどれだけあるのか?How much success is recoverable by selection alone?

扉位置を6pxずらすと、成功候補が在るにもかかわらず56.8〜58.8%を選び損ねる。Under door shift, 56.8–58.8% of episodes contain a good but unselected candidate.

Model
action-noise
door shift
matched
narrow
truth-only
DINO ranking loss
13.2%
56.8%
8.0%
31.2%
10.0%
JEPA ranking loss
12.8%
58.8%
5.6%
31.6%
7.6%
coverage ceiling

1 − generator failure
bank内に成功候補がある確率。Probability the bank contains a success.

latent success

現行の潜在argminが選ぶ成功率。Success selected by latent argmin.

ranking headroom

coverage ceiling − latent success
選択器だけで理論上回収可能。Theoretically recoverable by selector.

最大headroomがある層こそ次介入の標的。ただしoracleは“どの選択器で回収できるか”までは答えない。Largest-headroom strata are intervention targets; the oracle does not specify which selector can recover them.
decomposition.csv, cross-wall latent rows; bar scale maximum = 60 percentage points.
Q扉位置を変えたとき、なぜ固定PCCは失敗を見抜けないのか?Why can a fixed PCC miss failure after the true door moves?

PCCは既知の扉を通るかだけを検査するため、真の扉が6pxずれると「リスク0」の予測でも実行は衝突する。PCC checks only the known door, so a 6-pixel truth shift can make a risk-zero prediction collide in execution.

PCCが見る扉Checker door

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.

実行側の扉Truth door

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.

観測結果Observed outcome

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.

復号軌道decoded path実行軌道executed path  PCCが仮定する扉checker opening JEPA / seed 8101 / door-center-shift / episode 9JEPA / seed 8101 / door-center-shift / episode 9
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.
QPCC再順位付けは成功率を改善したのか?Did PCC reranking improve success?

全10条件が+5ptゲートに届かず、H2は支持されない。None of 10 cells cleared +5 points; H2 is not supported.

model条件conditionlatentPCCΔ95% CI改善/悪化impr./harmedHolm p
dinoaction-noise0.9300.910−0.020[−0.040, −0.002]8/180.680
dinodoor-center-shift0.6820.662−0.020[−0.044, 0.006]20/301.000
dinomatched0.9380.9380.000[−0.014, 0.020]6/61.000
dinonarrow-door0.7420.716−0.026[−0.034, −0.018]13/260.533
dinotruth-only-obstacle0.8120.816+0.004[−0.012, 0.024]7/51.000
jepaaction-noise0.9320.926−0.006[−0.022, 0.008]9/121.000
jepadoor-center-shift0.6720.666−0.006[−0.022, 0.010]16/191.000
jepamatched0.9500.944−0.006[−0.014, 0.002]3/61.000
jepanarrow-door0.7400.752+0.012[−0.030, 0.052]28/221.000
jepatruth-only-obstacle0.8240.822−0.002[−0.014, 0.016]6/71.000
各セル n=500ペア。実効ゲート (+0.05利得かつHolm有意) に全セル未到達 → H2 not_supported。CIと改善/悪化カウントが一次証拠 (p値のみで判断しない)。生成失敗は全セルで不変 (generator invariant)。n=500 pairs per cell. No cell reaches the practical gate (+0.05 gain AND Holm-significant) → H2 not_supported. CIs and improved/harmed counts are primary evidence (not p-values alone). Generator failure invariant in all cells.
QPCCは同じ話の成功・失敗を、どちら向きへ反転させたのか?In which direction did PCC flip paired episode outcomes?

改善例はあるが、全体では成功→失敗の反転が改善を上回る。PCC creates some wins, but success-to-failure flips dominate overall.

Model / condition
0→1
改善improved
1→0
悪化harmed
Δ
DINO / action-noise
8
18
−2.0pt
DINO / door shift
20
30
−2.0pt
DINO / matched
6
6
0.0pt
DINO / narrow-door
13
26
−2.6pt
DINO / truth-only
7
5
+0.4pt
JEPA / action-noise
9
12
−0.6pt
JEPA / door shift
16
19
−0.6pt
JEPA / matched
3
6
−0.6pt
JEPA / narrow-door
28
22
+1.2pt
JEPA / truth-only
6
7
−0.2pt
paired_contrasts.csv, n=500 pairs per row. Bars scale to 30 discordant episodes.
Q効果量と統計は、どれほどの改善を示したのか?How much improvement do paired effects and statistics show?

効果は−2.6〜+1.2ptに留まり、改善より悪化が多い。Effects stay between −2.6 and +1.2 points, with more harm than gain.

  • 効果量は −0.026〜+0.012 に分布 — 7条件で負・2条件で正・1条件でゼロEffects span −0.026 to +0.012 — negative in 7 cells, positive in 2, zero in 1
  • 未補正CIが0を下回るのは dino の action-noise と narrow-door のみ (= むしろ悪化方向)Only dino action-noise and narrow-door have unadjusted CIs below zero (i.e., harmful direction)
  • Holm補正 p はすべて 0.533〜1.000 で非有意All Holm-adjusted p-values are 0.533–1.000 (non-significant)
  • この凍結された線形PCC実装は、測定されたheadroomを有意な利得に変換できなかった (線形プローブ一般の限界とは主張しない)This frozen linear-PCC reranker did not convert the measured headroom into a significant gain (no claim about linear probes in general)
Paired effect of PCC reranking Seed-cluster bootstrap 95% intervals; exact McNemar p-values adjusted by Holm across emitted contrasts. Paired effect of PCC reranking Seed-cluster bootstrap 95% intervals; exact McNemar p-values adjusted by Holm across emitted contrasts. -0.50 -0.25 +0.00 +0.25 +0.50 dino_wm_wall / action-noise pHolm=0.680 dino_wm_wall / door-center-shift pHolm=1.000 dino_wm_wall / matched pHolm=1.000 dino_wm_wall / narrow-door pHolm=0.533 dino_wm_wall / truth-only-obstacle pHolm=1.000 jepa_wm_wall / action-noise pHolm=1.000 jepa_wm_wall / door-center-shift pHolm=1.000 jepa_wm_wall / matched pHolm=1.000 jepa_wm_wall / narrow-door pHolm=1.000 jepa_wm_wall / truth-only-obstacle pHolm=1.000 Paired success difference (PCC - latent)
Q集約成功率と候補数だけを見てよいのか?Are aggregate success and candidate count sufficient?

集約値は壁横断の弱さを隠し、被覆は16候補で95%に達する。Aggregates hide wall-crossing weakness; coverage is already 95% at 16.

  • H3 supported (両モデル・閾値 gap≥0.10):
    • topology成功率gap 0.305 / 0.310、AUROC gap 0.367 / 0.437
    • 候補被覆の成功率gap 0.871 / 0.868 (被覆AUROC gapは単一クラスのためNA)
    H3 supported (both models, threshold gap≥0.10):
    • Topology success gap 0.305 / 0.310, AUROC gap 0.367 / 0.437
    • Coverage success gap 0.871 / 0.868 (coverage AUROC gap is NA: single-class stratum)
  • 集約成功率は cross-wall 単体の性能より高く見える (層混合の効果。一般的な過大評価の証明ではない)Aggregate success reads higher than the cross-wall estimand (a mixing effect; not proof of generic overestimation)
  • matched の候補被覆: 予算16で95.0% → 256で97.8% (ほぼ天井)。良い候補は低予算でも既に在るMatched coverage: 95.0% at budget 16 → 97.8% at 256 (near ceiling). Good candidates already exist at low budget
Candidate coverage scaling on matched Wall Shared nested candidate banks; coverage means at least one successful candidate exists. Candidate coverage scaling on matched Wall Shared nested candidate banks; coverage means at least one successful candidate exists. 0% 25% 50% 75% 100% 16 32 64 128 256 dino_wm_wall jepa_wm_wall Nested candidate budget Candidate coverage
Q開始と目標が壁の同じ側か反対側かで、難しさはどれほど変わるのか?What does topology stratification reveal that aggregation hides?

壁を越えない課題は97.6%成功するが、扉の通過が必要な課題では66〜67%まで落ちる。Success is 97.6% on same-side episodes but only 66–67% on cross-wall.

Model / topology
Success
AUROC
Generator
Ranking
JEPA / same-side
97.60%
0.492
1.36%
1.04%
JEPA / cross-wall
67.12%
0.859
9.60%
23.28%
DINO / same-side
97.60%
0.407
1.36%
1.04%
DINO / cross-wall
66.56%
0.844
9.60%
23.84%
JEPA gap

success 30.48pt / AUROC 0.367
凍結閾値0.10を超える。Exceeds frozen 0.10 thresholds.

DINO gap

success 31.04pt / AUROC 0.437
両モデルでH3 supported。H3 supported for both models.

AUROC = 同じ候補バンクで、選択スコアが成功候補を失敗候補より上に並べる確率(0.5=無作為、1=完全)。高い値は全体性能を意味せず、cross-wallの成否が分離しやすい場合もある。AUROC = probability the selector ranks a successful candidate above a failed one in the same bank (0.5=random; 1=perfect). Higher is not overall quality; cross-wall outcomes may simply be easier to separate.
hypotheses.json, H3 pooled across five conditions; n=1,250 per topology and model.
QH1とH4から、何を主張できるのか?What can H1 and H4 support?

予測誤差差は確認できたが、機構シフトH1は判定不能だった。Prediction-error mismatch is supported; mechanism shift remains indeterminate.

H4 supportedH4 SUPPORTED

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.

H1 indeterminateH1 INDETERMINATE

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.

この発表の主張範囲: 公式2モデルのWall評価で、故障分解・線形PCCの負結果・層別異質性・予測誤差差を報告する。Claim scope: on two official Wall checkpoints, report failure decomposition, the linear-PCC null result, stratified heterogeneity, and prediction-error mismatch.
Q基準条件で成功率が近い両モデルは、軌道予測の誤差まで同じなのか?Do similar matched success rates imply similar prediction errors?

成功率の差は1.2ptでも、JEPAの終端誤差は0.366、経路誤差は0.187小さい。Despite a 1.2-point success gap, JEPA lowers endpoint error by 0.366 and path error by 0.187.

matched / latent / budget 256matched / latent / budget 256
JEPA
DINO
Success
95.0%
93.8%
Generator failure
2.2%
2.2%
Ranking failure
2.8%
4.0%
Endpoint prediction error
1.123
1.489
Path prediction error
0.721
0.908
H4: success gap 0.012 ≤ 0.05 かつ mismatch gap ≥ 0.05 → supported。これはモデル同等性の証明ではない。H4: success gap 0.012 ≤ 0.05 and mismatch gap ≥ 0.05 → supported; this is not an equivalence claim.
hypotheses.json, H4 matched / latent / maximum budget, n=500 per model.
Qモデルに見えない障害物を足すと、失敗は候補生成側へ移るのか?Did the hidden obstacle shift failure composition toward generation?

候補不足の比率は約20〜24pt増えたが、基準条件の失敗数が少なく仮説H1は確定できない。Generator share rose 20–24 points, but too few matched failures make H1 indeterminate.

Model / condition
失敗内構成:生成 / 順位付けFailure composition: generation / ranking
fail n
JEPA / matched
44.0%
56.0%
25
JEPA / truth-only
63.6%
36.4%
88
DINO / matched
35.5%
64.5%
31
DINO / truth-only
59.6%
40.4%
94
JEPA shift +19.6pt

閾値20ptに僅かに未達、matched n=25。Just below 20pt threshold; matched n=25.

DINO shift +24.1pt

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.
Part 3 考察と限界
診断の成功と介入の不発を分け、主張範囲を定める
Part 3  Discussion & Limitations
Separate diagnosis from intervention and bound the claim
03
Q代理モデルで効いたPCCは、公開モデルでも効くのか?Does PCC transfer from the surrogate to public checkpoints?

サロゲートの45.4pt改善は、公式モデルでは再現しなかった。The surrogate's 45.4-point gain did not reproduce on official models.

制御サロゲートControlled surrogate
公式チェックポイントOfficial checkpoints
意味Meaning
真の状態を直接使う小さな代理系。世界モデルの表現誤差と座標プローブの誤差を取り除く。A small proxy that reads true state, removing world-model and probe errors.
公開されたJEPA-WM / DINO-WMの学習済み重み。再学習せず凍結して使う実評価対象。Released JEPA-WM / DINO-WM weights, frozen and used without retraining.
採用理由Why use it
PCCの数式自体に、正しい状態が与えられれば選択を改善する能力があるかを単独で確かめるため。To test whether the PCC rule can improve selection when given correct state.
提案が都合のよい代理表現だけでなく、第三者が入手できる実際の表現へ移るかを確かめるため。To test transfer to real, publicly obtainable representations.
言える範囲Claim scope
機構として可能か。外的妥当性や公開モデルでの効果は主張しない。Mechanistic possibility only; no transfer claim.
公開重み・線形プローブ・本Wall設定で実際に効くか。Actual effect for public weights, linear probe, and this Wall setup.
制御サロゲートSurrogate
52.14% → 97.57%
公式10条件10 official cells
−2.6〜+1.2pt
機構検証用の制御サロゲート 既知状態でPCC単体を分離評価 52.14% latent 97.57% +PCC 大改善に見えた 再現せず did not transfer 公式チェックポイント official / 全10セルの対応効果 −0.25 0 +0.25 効果 −0.026〜+0.012 ・ Holm後すべて非有意 この「ずれ」自体が知見 / the divergence itself is the finding
Q線形PCCがheadroomを回収できない原因を、どう切り分けるのか?How can we localize why linear PCC fails to recover headroom?

候補生成ではなく、復号・時間集約・較正の3段階を個別に検証する。Test decoding, temporal aggregation, and calibration separately from generation.

① 潜在→座標① Latent → coordinates

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.

② 時間→max② Time → max

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.

③ 分位正規化③ Quantile scaling

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.

④ argmin反転④ Argmin flip

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.

Probe ablation真値座標 vs 線形復号で、復号損失を分離。Truth coordinates vs linear decode.
Risk ablationmax vs 積分 / sequence-aware risk。Max vs integrated or sequence-aware risk.
Calibration ablation単一λ vs 条件別 / 非線形reranker。Single λ vs condition-aware or nonlinear reranker.
これは実装から導いた検証可能なボトルネック仮説であり、今回の結果だけで因果確定した説明ではない。These are implementation-grounded, testable bottleneck hypotheses—not causal conclusions from the current run.
Qこの結果はどこまで言えて、どこから言えないのか?How far do these results generalize?

結論はWallシミュレーションと線形プローブに限定される。The conclusion is limited to Wall simulation and this linear probe.

  • 構成概念PCCは「既知幾何との実行前照合」— 未知の制約は原理的に発見できない (設計上の限界)。未知制約下の挙動は本評価の範囲外ConstructPCC checks consistency with known geometry — unknown constraints are undiscoverable by design; behavior under them is out of scope
  • 統計効果量・区間・改善/悪化カウントを一次証拠とし、p値のみで実務的重要性を判断しない。null・harmful条件も除外せず全報告StatisticalEffect sizes, CIs and improved/harmed counts are primary; practical importance is never inferred from p-values alone. Null and harmful conditions are all reported
  • 外的単一環境 (Wall)・シミュレーション限定・線形プローブのみ。実機・画像入力・複数環境・強baselineは対象外 — arXiv級一般化には各ゲートの通過が必要ExternalSingle environment (Wall), simulation-only, linear probe only. Real robots, image inputs, multi-environment and strong baselines are out of scope — arXiv-level generalization requires those gates
  • 解釈順位付け失敗が大きいことは改善可能な場所の診断であり、任意の再順位付け器が成功する保証ではない。本実験が示したのは線形PCCの効果までInterpretationLarge ranking failure diagnoses where headroom exists; it does not guarantee that any reranker will work. This experiment identifies the effect of this linear PCC only
Qこの研究で、何が確定したのか?What has this study established?

診断は選び損ねを定量化したが、線形PCCは埋められなかった。The audit quantifies mis-selection, but linear PCC does not close it.

診断DIAGNOSIS

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.

介入INTERVENTION

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.

次の判断NEXT DECISION

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.

Q4仮説の結論は、閾値と観測値から再計算できるか?Can each hypothesis decision be recomputed from threshold and observation?

閾値・観測値・判定を一対一で並べると、H3/H4だけが支持される。A one-to-one evidence ledger supports H3/H4, but not H1/H2.

Hyp.
問いQuestion
凍結閾値Frozen threshold
観測値Observed
判定Decision
H1
障害物で生成比率が増える?Does obstacle shift generation share?
shift ≥ .20
failures ≥ 50
+.196 / +.241
matched n=25 / 31
INDETERMINATE
H2
PCCで成功率+5pt?Does PCC add 5pt?
Δ≥+.05 + Holm
gen invariant + rank↓
range −.026…+.012
Holm: no cell
NOT SUPPORTED
H3
topology差≥10pt?Topology gap ≥10pt?
success & AUROC gap ≥.10
n≥50
success .305/.310
AUROC .367/.437
SUPPORTED
H4
成功近似・誤差差あり?Similar success, different error?
success gap≤.05
mismatch gap≥.05
success .012
endpoint .366 / path .187
SUPPORTED
否定結果も判定不能も残すことで、改善主張ではなく“どこまで分かったか”を追跡可能にする。Keeping negative and indeterminate outcomes makes the knowledge boundary traceable.
hypotheses.json + paired_contrasts.csv + frozen decision contract schema v3.
Part 4 今後の計画
故障診断を発表の主題にし、次の介入を検証する
Part 4  Roadmap
Lead with diagnosis, then test a stronger intervention
04
Qこの結果を学会発表では、どう構成するのか?How should SI2026 frame these results?

SI2026ではPCC改善ではなく、故障分解による診断を主題にする。SI2026 should lead with diagnosis by failure decomposition, not PCC gain.

01 / PROBLEM

成功率だけでは故障箇所が分からないSuccess rate hides the failure location

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.

02 / METHOD

全候補の反実仮想監査で分解するCounterfactual audit of every candidate

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.

03 / RESULT

難条件では順位付けが最大成分Ranking dominates hard 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.

04 / SCOPE

診断を主張し、介入の負結果も報告するLead with diagnosis; retain the null result

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.

発表タイトル案:Proposed title: 候補生成と順位付けの故障分解による潜在世界モデルMPCの実行前診断評価Pre-execution diagnosis of latent world-model MPC by decomposing generation and ranking failures
Q発表可能な研究成果にする次の工程は何か?What work turns this result into a presentation-ready contribution?

主張固定→原稿→再現資料→発表の順で12月までに仕上げる。Freeze claims, write, package reproducibility, then prepare the December talk.

2026年8月9月10–11月12月 講演申込 締切 8/31・タイトル+概要は確定済 主張・タイトル・概要を固定 著者・貢献範囲を確定 申込・原稿の表記へ反映 原稿提出 9/11–30 数値確定版ドラフト完成済 (整形のみ) 強リランカ実験 事前登録 → 実行 → 凍結ゲート判定 (次頁) SICE SI2026 発表 12/8–10 グランキューブ大阪 再現性パッケージと外的検証を並行して整備 凍結結果・runner・ハッシュ照合を公開し、第2環境PointMazeで診断の再現性を確認する。
Q線形PCCの次に、何をどの判定基準で試すのか?What intervention comes next, and how will it be gated?

選び損ねが最大の条件へ、潜在特徴も使う非線形の衝突リスク推定を適用する。The next test targets the largest headroom with nonlinear collision risk.

01 / TARGET

最大headroom層Largest headroom

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).

02 / MODEL

非線形の衝突リスクNonlinear collision risk

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.

03 / TEST

同じ候補・同じゲートSame bank, same gate

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.

成功条件は、選択だけで回収できる余地(headroom)が最大の層を狙う介入が、+5pt・Holm有意・生成不変・順位付け失敗減少の凍結ゲートを同時に通ること。Success means an intervention targeting the stratum with the largest selection-recoverable headroom clears all frozen gates: +5 points, Holm significance, unchanged generation, and fewer ranking failures.
Q第三者は、どこから結果と実装を追えるのか?Where can a third party trace the results and implementation?

全数値は凍結結果・公開runner・ハッシュ照合から再現できる。Frozen results, the public runner, and hashes make every number traceable.

01 / FREEZEcommitと重みを固定Freeze code and weights

commit 13cf1d9cと公式checkpointのSHA-256を照合。Pin commit 13cf1d9c and verify official checkpoint hashes.

02 / BANK候補を全手法で共有Share candidate banks

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.

03 / AUDIT128万候補を実行Execute 1.28M candidates

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.

04 / CLAIMゲート通過後に数値化Publish after claim gates

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.

追跡先:Traceability: SI_RESULTS_GENERATED.md / 公式数値レポート / 成果・レポート一覧 / 平易版7枚デッキ
論文ドラフトの出所識別子: commit 13cf1d9c / contract schema v3 / claim gate 6条件。Manuscript provenance: commit 13cf1d9c / contract schema v3 / six claim gates.
→/Space 次 ・ ← 戻る ・ F 全画面 ・ O 一覧 ・ L 日英 ・ P 印刷(PDF)