{"id":"eb545b5f-ff37-4e90-b983-c323168ab497","arxiv_id":"2609.31394","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"InternW0-Δ couples a pretrained video expert with an action expert in a directed mixture-of-transformers, learns future-relevant change representations without future-video rollout, and reports top scores on LIBERO-Plus, RoboTwin 2.0, EBench, and RoboDojo after 20K+ hour pretraining.","lead":"InternW0-Δ is a robot control network that couples video prediction with action generation in one mixture-of-transformers model and trains on over 20,000 hours of robot and human data. It reports top success rates on four simulation benchmarks and four real robot platforms, and promises to release code, weights, and data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The 'outperforms prior methods' claim rests on 1.3-2.5-point leads over baselines harvested from other papers (Sec. 7.1) with no error bars; on EBench the lead sits entirely on a composite score despite a lower SR (49.2 vs 49.4), and the authors' own ablations move results by up to 8.7 points.","rationale":"This is a large, well-executed systems paper whose central claim is comparative empirical performance on external benchmarks, so correctness risk concentrates exactly where the reader put it: on the comparison against prior numbers, not on the architecture's internal soundness. The directed-mask construction (Sec. 3.6, Fig. 3c) genuinely prevents a forward path from future tokens to the action path, and the inference procedure (Sec. 3.8) requires no future sampling; Causal Imprint and 4D distillation are training-only as claimed. I found no internal-consistency flaw in the design: the causal mask, the flow objectives, and the caching scheme are consistent with each other, and the latency ablation (Table 8) is well-controlled with success rate nearly flat (92.78 to 92.23) across configs.\n\nThe load-bearing weakness is the size of the lead over the strongest baselines, all harvested rather than re-run (Sec. 7.1). The margins are 1.4 points (LIBERO-Plus), 2.5 points (RoboTwin Clean2Random), 1.3 points (EBench score), and 1.43 points (RoboDojo over GPT-6 Astra), with no seed variance anywhere in Tables 9-12. This is not merely a missing-statistics style point: the paper's own ablations demonstrate that near-identical configurations move results by 3.9 to 8.7 points (VLM swap alone is 8.71 in Table 14; teacher choice spans ~4.7 in Table 15). A headline lead smaller than the method's own configuration sensitivity is exactly the regime where run-to-run or protocol-mismatch noise can flip the ranking. EBench is the clearest case: the paper's SR (49.2) is below OpenWAM-a's (49.4), and the 'best' claim rests entirely on a composite score the paper does not decompose per task.\n\nThe real-robot pretraining comparison (Table 13) that the reader also flagged is, in my read, less load-bearing: 0/20 vs 19/20 is a statistically robust gap even with wide confidence intervals (Fisher exact p ~ 1e-6), so n=20 does not threaten that specific claim. I also note a smaller unexplained gap: the full ablation stack reaches 78.37 (Table 14) and robot-only pretraining 83.73 (Table 16), yet the headline is 92.8, a ~9-point jump not attributable to any ablated component; this reinforces, rather than replaces, the central concern that the headline margin is not independently verifiable from the manuscript. A secondary credibility tension is the 'largest open-source corpus of its kind' claim versus the 100K+ hour Xiaomi-Robotics-1 dataset the paper itself cites in Table 12.\n\nMy read agrees with the reader's weakest assumption and confirms the CONDITIONAL verdict; nothing moves it, so the verdict is UNCHANGED. The coupled multi-seed re-run described in concrete_test is the single check that would settle whether the headline margins are real.","tokens_in":47367,"tokens_out":7920,"duration_ms":52800,"concrete_test":"Re-run InternW0-Delta and the two strongest baselines (Qwen-RobotManip-Context; OpenWAM-a) on LIBERO-Plus, RoboTwin 2.0, EBench, and RoboDojo under identical protocol with at least 3 seeds over full eval sets, and separately re-run the Table 14 VLM-swap row (Qwen3.5 vs RynnBrain) and Table 15 teacher rows with 3 seeds to measure the method's own variance. If the 92.8 vs 91.4 (LIBERO-Plus), 71.9 vs 69.4 (RoboTwin), and 66.0 vs 64.7 (EBench) gaps fall inside the joint seed-variance band, the 'outperforms' claim fails at confidence; if the gaps exceed twice that band, the claim survives. This single check settles whether the headline margin is real or noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The decisive claims (Abstract, b1: 'outperforms prior methods across diverse simulation benchmarks') are comparative, so correctness risk concentrates on the comparison against prior numbers. Section 7.1 states 'Baseline results are taken from the corresponding papers or public benchmark results under matching evaluation protocols,' i.e., no baseline was re-run. The largest leads over the strongest baselines are thin on every benchmark: LIBERO-Plus 92.8 vs 91.4 over Qwen-RobotManip-Context; RoboTwin Clean2Random 71.9 vs 69.4; EBench score 66.0 vs 64.7; RoboDojo avg 23.91 vs 22.48 over GPT-6 Astra and 19.34 over DM0.5 (the 'nearly double' phrasing holds only versus WAM baselines, 11.92 for OpenWAM-a). None of Tables 9-12 reports seed variance, so a 1.3-1.4-point lead is indistinguishable from run-to-run noise. The sharpest case is EBench: InternW0-Delta's success rate (49.2) is below OpenWAM-a's (49.4), so the 'highest overall score' claim rides entirely on the benchmark's composite score metric, which is never decomposed per task. The concern is corroborated internally: Table 14 shows a frozen-VLM swap (Qwen3.5 to RynnBrain) moves LIBERO-Plus by 8.71 points and Table 15 teacher choice spans 73.63-78.37; a 1.4-point headline lead is smaller than the method's own demonstrated configuration sensitivity. The real-robot ablation (Table 13, 0/20 vs 19/20) is statistically robust despite n=20 (Fisher exact p ~ 1e-6) and is not the load-bearing weakness. Secondary unexplained gap: the full ablation stack tops out at 78.37 (Table 14) and robot-only pretraining at 83.73 (Table 16), yet the headline is 92.8, a jump not attributable to any reported component.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper introduces InternW0-Δ, a World Action Model that couples a pretrained video expert (Wan2.2-TI2V-5B) and a randomly initialized action expert (ActionDiT) inside a 30-block Mixture-of-Transformers, with semantic conditioning from a frozen VLM, a sparse anchor/recent/current visual memory, a training-only future-supervision objective called Causal Imprint, and a training-only 4D-aware distillation from a Track4World teacher. The model is pretrained on a 23,072-hour heterogeneous corpus (robot demos, UMI, egocentric, and Ego2Robot data) and post-trained on four simulation benchmarks (LIBERO-Plus, RoboTwin 2.0, EBench, RoboDojo) and four real-robot platforms (two gripper, two dexterous-hand). The headline claims are state-of-the-art success rates on LIBERO-Plus (92.8%) and RoboTwin 2.0 Clean2Random (71.9%), the highest EBench overall score (66.0), the best RoboDojo average (23.9%), and real-robot deployment at 152.8 ms round-trip latency. The authors report component-wise, data-source, and teacher-choice ablations, and openly state that raw egocentric video provides little measurable benefit on their current gripper-centric benchmarks (b606–b610).","tokens_in":47858,"tokens_out":5977,"duration_ms":33355,"significance":"If the benchmark claims hold, this is a substantial systems contribution: the directed information-flow design (Causal Imprint and 4D-aware distillation with the teacher discarded at inference) is clean and falsifiable, the inference engineering (single video prefill, cached K/V, action-only denoising) is practical, and the planned 23K-hour open corpus with release of code, weights, and filtered-data indices would be a genuine community resource. The empirical support is unusually thorough for this area: cumulative component ablations (Table 14), per-source data ablations (Tables 16–17), teacher-choice ablations (Table 15), and an honest acknowledgment that egocentric data barely helps on the current benchmarks all reflect careful, transparent experimentation. The significance is contingent on the comparative 'outperforms prior methods' claim surviving statistical scrutiny, since the headline margins over the strongest baselines are 1.3–2.5 points and no error bars or re-run baselines are reported; this is the focus of the major comments.","major_comments":[{"comment":"The load-bearing comparative claim 'outperforms prior methods' (Abstract, b1) rests on baseline numbers harvested from prior papers, per Section 7.1 ('Baseline results are taken from the corresponding papers or public benchmark results under matching evaluation protocols'), with no re-run of baselines and no seed variance or error bars reported in Tables 9–12. The margins over the strongest baselines are thin on every benchmark (LIBERO-Plus 92.8 vs 91.4; RoboTwin Clean2Random 71.9 vs 69.4; EBench score 66.0 vs 64.7; RoboDojo average 23.91 vs 22.48), and the paper's own ablations demonstrate configuration sensitivity far larger than these leads: swapping the frozen VLM moves LIBERO-Plus by 8.71 points (Table 14) and teacher choice spans 73.63–78.37 (Table 15). Without multi-seed variance on the authors' own runs, or a re-run of at least the closest baselines under identical protocols, a 1.3–1.4-point headline lead is indistinguishable from run-to-run noise and does not substantiate the comparative claim as stated. I request either seed variance plus a statistical test on all four benchmarks, or a re-run of the nearest VLA baselines reported in the same tables.","section":"§7.1, Tables 9–12","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"On EBench, InternW0-Δ's overall success rate (49.2%) is lower than OpenWAM-α's (49.4%), so the 'highest overall score of 66.0' claim rests entirely on the benchmark's composite score, which is never decomposed per task in the paper. Since the composite is the sole basis for the EBench headlne, the paper should report the per-task score decomposition and state how the composite is weighted relative to raw success rate; without this, the EBench 'outperforms' claim is not auditable.","section":"§7.2.3, Table 11 (EBench)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The paper states that eleven tasks across four real-robot platforms were evaluated, but quantitative success rates are reported only for two AC-One tasks (toast bread and Luminol reaction, Table 13, n=20 per condition); the other nine tasks (get a drink, MOF experiment, magnetic stirrer, place tube, pour water, place fruit into box, stack cups, use dropper, make a sandwich) are supported only by qualitative frame selections in Figure 10. The abstract's claim of strong real-robot performance across platforms therefore currently rests on two tasks on a single platform. I request success rates with trial counts for the remaining tasks, or an explicit statement that those platforms were demonstrated rather than quantitatively benchmarked.","section":"§7.3, Table 13","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The sentence 'The weights for the video, Causal Imprint, and future-feature alignment loss are set to 0.5, 1.0, 0.5, and 0.1, respectively' lists four values for three losses and contradicts Table 4 (λ_video=0.5, λ_action per data source, λ_Δ=0.5, λ_align=0.1); please correct the enumeration.","section":"§5.1, Table 4","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The phrase 'nearly double that of the strongest prior WAM' (b34) is accurate against OpenWAM-α (11.92 to 23.91 on RoboDojo) but the margin over the best non-WAM baselines (GPT-6 Astra at 22.48) is only 1.4 points; please qualify the sentence so the reference class (WAM baselines only) is explicit.","section":"Abstract and §7.2.4","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The expansion from 1,101.55 source hours to 5,633.77 robot-hours (a factor of about 5) is attributed to multi-embodiment coverage and episode segmentation; a worked example in the caption or a footnote would make this accounting reproducible.","section":"Table 3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"In the RTC continuity comparison, the VJP trace is described as a single individual run while hard prefix and training-time RTC are shown as mean curves; please report the number of runs and dispersion for all three strategies.","section":"§7.3, Figure 12","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Read this one. It is a big, well-built systems paper: 23K hours of curated heterogeneous data, a genuinely new prediction objective, and the most complete ablation set I have seen in a WAM paper in a while. If you work on world action models, it is the current reference point.\n\nWhat is actually new: Causal Imprint. Instead of feeding future frames to the policy, they regress adjacent clean-latent differences plus align Causal Imprint tokens with future video features, at zero future-access cost at inference. The directed attention mask genuinely prevents future leakage into the action stream. That is a real design decision with real ablations behind it (Table 14: CI alone +1.7, alignment +5.7). Second, the Ego2Robot pipeline: 5,633 robot-hours from egocentric video via coupled base search, trajectory IK, and inpainting. Substantial engineering, and they are honest that raw ego video barely helps (Tables 16–17) — they say so plainly in the limitations. The ablations are unusually good: component-wise cumulative, teacher choice, data source. They earn credit.\n\nThe soft spots are where the headline claims live. All baselines are harvested from papers, not re-run (Sec. 7.1), and there is no seed variance reported anywhere. The largest leads over the strongest per-benchmark method are 1.4–2.5 points — indistinguishable from run-to-run noise. The sharpest case is EBench: success rate 49.2 vs OpenWAM's 49.4, lower, yet they claim the best result via the composite score. And 'nearly double' on RoboDojo only holds versus WAM baselines (11.92 for OpenWAM-a), not versus the LLM-as-policy (GPT-6 Astra at 22.48). The paper's own ablations move results by 8.7 points on a VLM swap, so a 1.4-point headline lead is smaller than their own configuration sensitivity. The stress-test note is right on both counts.\n\nOne thing I noticed that neither the reader nor the stress test fully flagged: the full ablation stack tops out at 78.37 and robot-only pretraining at 83.73, yet the headline is 92.8. Nothing in the reported components explains that jump. Either there is a scaling effect not decomposed, or the ablated runs differ in some unreported way. Minor, but worth asking the authors.\n\nAlso: 'largest open-source corpus of its kind' sits awkwardly next to Xiaomi-Robotics-1 (100K+ hours), which they themselves cite. And code, weights, and data are promised, not shipped.\n\nNet: the architecture is sound and the objective is new. The 'outperforms' claim needs either re-run baselines with error bars or a softened phrasing. A reviewer should push on that, not on the method. Send it out — it is the kind of paper the field needs refereed properly.","headline":"A genuinely new training objective (Causal Imprint) and a serious 23K-hour data pipeline wrapped around a solid WAM, but the 'outperforms prior methods' headline rides on thin margins over baselines harvested from other papers.","tokens_in":48796,"tokens_out":2897,"duration_ms":18325},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"InternW0-Δ claims a world-action model that turns training-only future supervision into action-relevant features, reports the best results on four manipulation benchmarks and four real robot platforms, and does all of it without…","keywords":["world action model","robot manipulation","Causal Imprint","video prediction","Mixture-of-Transformers","4D representation distillation","heterogeneous data corpus","ego-to-robot learning"],"falsifier":"Re-running the strongest compared VLA and WAM baselines (e.g., Qwen-RobotManip-Context and OpenWAM-α) under identical LIBERO-Plus, EBench, and RoboDojo protocols with repeated seeds would show whether the 1.4-point and 1.3-point margins persist or fall within run-to-run noise; separately, repeating the real-robot pretraining ablation with more than 20 trials per condition would test whether the 0% to 95% jump on Luminol reaction is real or an artifact of a small sample.","tokens_in":48818,"feed_emoji":"🤖","tokens_out":4739,"duration_ms":31138,"temperature":0.7,"pith_summary":"InternW0-Δ is a world action model for generalist robot manipulation that claims to make video-prediction knowledge directly useful for control without paying the cost of future-video generation at run time. The paper argues that the right architecture—a directed mixture-of-transformers that couples a pretrained video-diffusion expert with a separate action expert—lets training-only future supervision improve action prediction while the policy never sees or decodes a future frame. To back this up, it builds a 23,072-hour heterogeneous corpus of robot demonstrations, UMI data, egocentric human video, and ego-to-robot conversions, unifies them under a canonical 80-dimensional state-action representation, and reports the best published success rates on LIBERO-Plus (92.8%), RoboTwin 2.0 (71.9% Clean2Random, 90.0% Clean2Clean), the highest EBench overall score (66.0), and the best RoboDojo average (23.9%, roughly double the strongest prior WAM). If the claims hold, the work shows that predictive and geometric priors can be folded into robot policies during training and then discarded at inference, with a 152.8 ms round-trip latency on a single RTX 5090.","feed_headline":"Future-video training, no future-video inference: 92.8% on LIBERO-Plus","feed_subtitle":"A world action model distills predictive and 4D priors into actions on 23K open hours, then skips video decoding at run time.","key_machinery":"The load-bearing object is the directed World-Action Mixture-of-Transformers with its attention mask, which separates two token streams: video tokens (anchor, recent, current, and training-only future) and action tokens, coupled through masked joint self-attention but forced to exchange information only along allowed directions. Causal Imprint tokens sit between the streams: they aggregate recent and current visual features, are trained against adjacent differences of clean future video latents ($\\Delta z_{t,i} = z_{t+i\\rho_v} - z_{t+(i-1)\\rho_v}$) plus a cosine alignment to terminal future-video features, and their hidden states are exposed to the action expert at inference. The 4D-aware distillation adds a query-based student branch aligned via MSE to cached 1,430-dimensional Track4World descriptors. All of it rests on a canonical 80-dimensional state-action representation and a two-stage recipe (pretrain on the 23,072-hour corpus, then post-train per embodiment), with K/V caching and CUDA-graph replay making action-only denoising fast enough for real-time control.","core_discovery":"The central claim is that predictive visual dynamics can be separated from action inference without losing their benefit: future supervision is used only to shape representations, never as policy input. InternW0-Δ couples a pretrained Wan2.2-TI2V-5B video expert with a randomly initialized ActionDiT action expert inside a directed Mixture-of-Transformers, where a frozen VLM supplies scene semantics to the action stream. A set of learnable Causal Imprint tokens, supervised by adjacent clean-latent differences and aligned to future video-expert features, encode future-relevant scene changes from anchor, recent, and current observations, and a training-only 4D-aware distillation from a Track4World teacher injects geometric and motion priors. A directed attention mask guarantees that no forward path exposes realized future frames to the action expert. On this design, the paper reports the best results among compared WAMs and VLAs across LIBERO-Plus, RoboTwin 2.0, EBench, and RoboDojo, and adaptation of one pretrained checkpoint to two gripper and two dexterous-hand real platforms.","pith_inferences":["Causal Imprint is a general recipe beyond manipulation: supervising a small token set with future-latent differences could make any video-pretrained backbone action-relevant, so the pattern may extend to navigation, humanoid locomotion, or other predictive-sensor control settings.","The strongest comparative margins are against other WAMs rather than leading VLAs—the 1.4-point LIBERO-Plus lead and 1.3-point EBench lead over the best VLA are near the range of run-to-run noise, so the method's standing relative to top VLAs is less settled than its standing among WAMs.","If the Ego2Robot conversion pipeline holds up at scale, the implication is that abundant egocentric human video can be recycled into robot training data, potentially easing the demonstration bottleneck for generalist policies.","A testable extension would be measuring whether Causal Imprint's future-supervision benefit grows on dexterous-hand tasks, where fine-grained contact and motion cues matter more than on gripper tasks, which the paper itself flags as under-covered by robot datasets."],"forward_implications":["Future-video generation can be dropped from the WAM inference path entirely, since Causal Imprint and the directed mask let the policy act without sampling or decoding future frames, cutting per-step compute to one video prefill plus action-only denoising.","A single pretrained checkpoint transfers across embodiments: the same 23K-hour pretraining adapts through platform-specific adapters to two gripper and two dexterous-hand setups with different action spaces.","Distribution-shift robustness is achievable without sacrificing clean-distribution performance, since post-training on unperturbed demonstrations alone yields both best Clean2Random (71.9%) and best Clean2Clean (90.0%) on RoboTwin 2.0.","Training iteration cost drops substantially from infrastructure alone: VAE/VLM feature caching plus layerwise MoT compilation gives 3.02× and 2.11× end-to-end speedups on LIBERO and RoboTwin.","Real-time dexterous control is feasible on a single consumer GPU, with a 5.11× inference speedup to a 152.8 ms controller-observed round-trip latency via context caching, grouped action execution, and CUDA-graph replay."],"supporting_citations":[{"why":"Establishes the design lineage of separating future-video supervision from the action inference path, which InternW0-Δ follows and extends.","marker":"Yuan et al., 2026b (Fast-WAM)"},{"why":"Supplies the pretrained video expert and T5 language-conditioning pathway that internally initializes the video stream and preserves its video-generation supervision.","marker":"Wan Team, 2025 (Wan2.2-TI2V-5B)"},{"why":"Is the frozen 4D teacher whose cached clip-level geometry and motion descriptors drive the training-only distillation loss.","marker":"Lu et al., 2026 (Track4World)"},{"why":"Provides the data-alignment and filtering practices the corpus pipeline builds on, and serves as the strongest VLA baseline on LIBERO-Plus and RoboTwin 2.0.","marker":"Yuan et al., 2026a (Qwen-RobotManip)"},{"why":"Is the frozen VLM whose scene-semantic features condition the action expert; the ablation shows VLM choice is the single largest source of gain.","marker":"Dang et al., 2026 (RynnBrain1.1-2B)"},{"why":"Is the strongest prior WAM baseline that InternW0-Δ claims to nearly double on RoboDojo (11.92% to 23.91% average success).","marker":"Wang et al., 2026d (OpenWAM-α)"},{"why":"Inspires the ego-to-robot conversion pipeline for kinematic alignment, which InternW0-Δ extends with coarse-to-fine trajectory validation.","marker":"Wang et al., 2026c (Ego2Robot)"},{"why":"Supplies the training-time RTC prefix-conditioning method that the real-robot asynchronous deployment relies on for continuous execution.","marker":"Black et al., 2025d (Training-time action conditioning)"}],"fun_headline_variants":["Future-video training, no inference: 92.8% LIBERO-Plus with open data","Train on future video, act without it: InternW0-Δ hits 92.8%","20K hours open data, zero future video at run time","Causal Imprint skips video rollout: 92.8% LIBERO-Plus"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The headline that InternW0-Δ outperforms prior methods assumes that the baseline numbers cited from other papers were measured under statistically equivalent evaluation conditions, since no baselines are re-run in-house and no cross-method error bars are reported.","fun_headline_variants_meta":{"raw":{"variants":["Future-video training, no inference: 92.8% LIBERO-Plus with open data","Train on future video, act without it: InternW0-Δ hits 92.8%","20K hours open data, zero future video at run time","Causal Imprint skips video rollout: 92.8% LIBERO-Plus"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000226,"raw_usage":{"cost":0.000226128,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000226128,"upstream_inference_prompt_cost":0.000128688,"upstream_inference_completions_cost":9.744e-05},"total_tokens":1472,"prompt_tokens":1124,"completion_tokens":348,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":263}},"tokens_in":1124,"tokens_out":348,"cost_source":"openrouter_usage","duration_ms":2848,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":263,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-09-28T04:41:57.526797+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Re-running the strongest compared VLA and WAM baselines (e.g., Qwen-RobotManip-Context and OpenWAM-α) under identical LIBERO-Plus, EBench, and RoboDojo protocols with repeated seeds would show whether the 1.4-point and 1.3-point margins persist or fall within run-to-run noise; separately, repeating the real-robot pretraining ablation with more than 20 trials per condition would test whether the 0% to 95% jump on Luminol reaction is real or an artifact of a small sample.","supporting_citations":[],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}