{"id":"f27cfd46-12e4-4e98-a521-887ca7265dbb","arxiv_id":"2610.10270","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"VPP2 shows that event-level video pretraining with detailed captions, followed by single-step distillation and a mixture-of-transformers action expert, yields strong zero-shot robot manipulation and video prediction.","lead":"The authors build VPP2, a robot \"world action model\" that learns to predict manipulation videos from detailed captions, then distills that video model into a fast single-step visual planner that also outputs actions. On real and simulated benchmarks it reports large gains over prior world models and vision-language-action policies in both zero-shot and post-trained settings.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"Event-level video prediction is credited for the action gains, but the only ablation (VPP2-Fixed-Step, Table 2) is scored by GPT video-instruction-following, never by robot success; the ALOHA baseline comparison also gives VPP2 much broader action-pretraining data.","rationale":"To support 'strong zero-shot generalization in both video prediction and action generation', the paper needs to show that the video-prediction mechanism is what carries the action results. It does not: VPP2-Fixed-Step, the only variant that differs on the key mechanism, is scored only by GPT-based video instruction following (Table 2), and no policy-level result uses this variant. The reader's CONDITIONAL verdict is therefore appropriate. I mark partial agreement because the reader identifies the causal-attribution gap, and I add that the ALOHA zero-shot comparison additionally conflates the broad action-pretraining data available to VPP2 with the narrower ALOHA-only finetuning given to pi0.5 and Fast-WAM (Figure 8 caption), so the 58.5% versus 40.0% margin has an alternative explanation independent of event-level video prediction. This does not make the results look fabricated; the system-level numbers can all be true while the causal story remains unsupported. With the fixed-step action ablation, the paper would either directly validate its central attribution or reveal that the policy gains come from other components such as chunk post-training, distillation, the MoT action expert, or data scale. Since the requested check is a missing experiment rather than a detected error, the reader's CONDITIONAL verdict is unchanged.","tokens_in":22150,"tokens_out":9452,"duration_ms":59879,"concrete_test":"Train the Stage-3 MoT action expert on top of the VPP2-Fixed-Step video backbone (same action data, same LoRA recipe, same VLM planner) and evaluate the same zero-shot ALOHA and LIBERO-Pro/OOD protocols. If the fixed-step policy's action success stays within a few points of VPP2 despite the large video-instruction-following gap, the event-level-video attribution is not supported; if it drops by a large margin, the attribution receives its first direct evidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"VPP2's central claim is causal: next-event video prediction with detailed captions is asserted to be what makes action generation generalize. Section 4.2 states this attribution directly ('We attribute these gains to next-event video prediction with detailed captions'), and the abstract repeats it. The only ablation that differs on this mechanism, VPP2-Fixed-Step, is evaluated solely with a GPT-based instruction-following metric on 50 image-instruction pairs per domain (Section 4.1, Table 2). It is never run through the Stage-3 action expert, so the paper never tests whether the large video gap (0.80 vs 0.48 human, 0.90 vs 0.42 robot) propagates to action success. The zero-shot ALOHA comparison cannot fill this gap because it is also confounded: per the Figure 8 caption, pi0.5 and Fast-WAM are 'finetuned on all ALOHA datasets exposed in VPP2's training data', whereas VPP2's action expert was pretrained on the full bimanual corpus (AgiBot World-Beta, RoboCOIN, RDT, GM-100, ABC-130K, self-collected ALOHA). The 18.5-point ALOHA margin could therefore come from action-pretraining data volume and embodiment familiarity rather than from event-level video prediction. This is a structural gap: the claimed mechanism and the measured outcome are never connected in any single experiment.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper introduces VPP2, a world action model built from four components: large-scale continued pretraining of a video foundation model on manipulation videos annotated with detailed captions and trained at the event level; chunk-level post-training with a fixed prediction horizon; consistency distillation to a single-step visual planner; and a mixture-of-transformers action expert trained on unified bimanual action data. The reported results are threefold: VPP2-14B achieves 0.80/0.90 instruction-following success on open-ended human/robot video prediction tasks versus 0.70/0.78 for Cosmos3-64B; it attains 58.5% average success on zero-shot real-world ALOHA tasks versus 40.0% for pi0.5 and 20.5% for Fast-WAM; and after benchmark post-training it reports 45.0%, 63.9%, and 32.26% on LIBERO-Pro, LIBERO-OOD, and RoboDojo respectively. The authors attribute the policy gains to next-event video prediction with detailed captions, which they argue establishes consistent instruction-to-trajectory mappings that transfer to action generation.","tokens_in":22597,"tokens_out":5200,"duration_ms":33800,"significance":"If the causal claim is supported, this is a useful and timely result: the paper contributes a large-scale captioning and data-processing pipeline, a concrete event-level pretraining recipe, a fixed-horizon distillation procedure with a reported 0.22-second per-chunk latency, and strong results on three external benchmark suites. The inclusion of a fixed-step video ablation, human pairwise win rates, and comparisons to many strong baselines are genuine strengths. The main weakness is evidentiary: the central mechanistic attribution is only tested with a video-side metric, and the real-robot comparison carries a data-scale confound. The benchmark results themselves are valuable even if the attribution is not fully established, but the paper's headline framing depends on showing that video-prediction quality is what drives action success.","major_comments":[{"comment":"The paper's central causal claim—that next-event video prediction with detailed captions is what produces the action gains—is not directly tested. The only ablation isolating this mechanism, VPP2-Fixed-Step, is evaluated solely by video instruction-following on 50 image-instruction pairs per domain (Table 2), and it is never run through the Stage-3 action expert or evaluated on ALOHA, LIBERO, or RoboDojo. The sentences in Section 4.2, 'We attribute these gains to next-event video prediction with detailed captions,' therefore connect a mechanism to outcomes that the reported experiments never jointly measure. I would ask for an action-level comparison of the event-level and fixed-step variants, or an explicit control that removes Stage 1 while keeping Stage 2 and Stage 3.","section":"Section 4.1, Table 2, and Section 4.2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The zero-shot ALOHA comparison is confounded by action-pretraining data breadth. The text says pi0.5 and Fast-WAM are trained on 'the same data used in VPP2 training,' while the Figure 8 caption says they are 'finetuned on all ALOHA datasets exposed in VPP2's training data'; VPP2's action expert was pretrained on the full bimanual corpus including AgiBot World-Beta, RoboCOIN, RDT, GM-100, ABC-130K, and self-collected ALOHA data, plus the large-scale video-pretraining corpus. These two statements describe different comparisons, and if the baselines saw only the ALOHA subset, the 18.5-point margin could reflect pretraining scale and embodiment familiarity rather than event-level video prediction. The section should state exactly which data each baseline received and report trial counts and standard errors per task category.","section":"Section 4.2 and Figure 8","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The LIBERO and RoboDojo results are reported as benchmark-specific post-training outcomes, but VPP2 differs from the baselines not only in the video-prediction stage but also in large-scale video and action pretraining. For example, all LIBERO models are stated to be trained exclusively on the four standard LIBERO suites, yet VPP2 additionally enters that post-training stage with the pretrained video and action modules described in Section 3. As a result, these tables cannot by themselves separate the contribution of event-level video prediction from the contribution of pretraining scale. This is not a flaw in the reported success rates, but it limits the stated attribution in Section 4.2.","section":"Table 3 and Table 4","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The video instruction-following evaluation described in Section 4.1 does not make fully explicit whether all models received the same input prompt format. VPP2's training uses detailed caption-template annotations, and Section 3.3 applies a VLM to expand user instructions into detailed subtask descriptions at execution time; if the baseline video models were evaluated with shorter, unexpanded instructions while VPP2 received caption-template prompts, the GPT-based instruction-following score could partly measure prompt format rather than prediction quality. The evaluation protocol should state the exact prompt given to each model and, ideally, report human-GPT agreement on the 50-pair evaluation sets.","section":"Section 4.1","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The consistency distillation objective is introduced with the notation F_theta for the student and an EMA parameter theta-bar, but the boundary condition F_theta(x,1,c)=x and the teacher integration step from (x_s,s) to (x_{s'},s') would benefit from one sentence clarifying which network supplies the target when s'=1 is sampled.","section":"Section 3.1, Eq. (5)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The transform notation in the unified action space section is ambiguous: A_d and B_d are defined as products of world and end-effector frames, but the text does not state whether these are maps from the dataset frame to the canonical frame or the reverse. A brief frame-diagram or explicit left/right convention would remove the ambiguity.","section":"Section 2, Eq. (6)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Table 2 reports point estimates without confidence intervals or trial counts beyond the stated 50 pairs per domain; given that several differences are around 10 percentage points, error bars or a short protocol paragraph would materially increase confidence in the comparisons.","section":"Section 4.1, Table 2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The win-rate bars in Figure 6 are visually dense and the baseline/win percentages are not aligned with the bars in the printed layout; a small table or a consistent ordering of 'Baseline wins / Ties / VPP2 wins' would make the human-evaluation results much easier to verify.","section":"Figure 6","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The reproducibility statement says detailed training configurations, code, and checkpoints are on the project website, but the manuscript itself does not list the main hyperparameters for Stage 2 and Stage 3 or the captioning VLM used for data annotation; including these in the appendix would strengthen the reproducibility claim independent of the website.","section":"Reproducibility Statement","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Quick take: this is a serious, well-engineered systems paper with headline results worth taking seriously, but its central causal claim — event-level video prediction with detailed captions drives the action gains — is not tested by any single experiment. Read the attribution one step more cautiously than the abstract does.\n\nWhat's actually new is the integrated recipe. Event-level continued pretraining of Wan2.1-14B on a large, caption-annotated manipulation corpus; fixed-horizon post-training; consistency distillation to one-step video generation; then a 0.9B mixture-of-transformers action expert over the video latents, with coordinate alignment and T-shaped multi-view inputs. Each ingredient has a prior home, but the combination plus the evaluation is a legitimate contribution. Benchmark results are strong and externally grounded: 45.0% on LIBERO-Pro against 11.0% for the best baseline, 63.9% on LIBERO-OOD, 32.26% on RoboDojo, and LIBERO-ID near 99%, so the gains don't come from a memorization tradeoff. The Fixed-Step ablation (0.48/0.42 vs 0.80/0.90 on the video metric) does show event-level prediction with detailed captions matters for the video model itself.\n\nSoft spots, in proportion. The stress-test's structural gap holds up: Section 4.2 attributes the gains to next-event video prediction with detailed captions, but VPP2-Fixed-Step never goes through the Stage-3 action expert. The mechanism and the action-level outcome are never connected, and the same gap applies to the LIBERO and RoboDojo attribution. The 11-point video headline rests on a GPT evaluator over 50 self-collected image-instruction pairs per domain, and the human pairwise data don't clearly corroborate the Cosmos3-64B advantage (28%/24% VPP2 wins, mostly ties). There's mild VLM-in/VLM-out circularity there — VLM captions, GPT judge — though the human judgments and the robot benchmarks keep it from being decisive. The ALOHA comparison is confounded: the baselines were fine-tuned only on ALOHA data, while VPP2's action expert saw the full bimanual corpus; 'zero-shot' means no task fine-tuning, not unseen embodiment. No trial counts or error bars. Two minor items: RoboDojo's Memory dimension is a real weak spot (4.1% vs GPT-6-Astra's 38.7%) hidden by the SOTA average, though the appendix tables make it visible; and code and checkpoints are promised on the project page but not attached. The citation pattern looks fine.\n\nBottom line: a serious empirical paper with a useful recipe and benchmark results worth engaging. The fixes — an event-vs-fixed-step action ablation, released artifacts, error bars — are all within reach. Send it to peer review; expect revision on the attribution. I'd cite it, and it would make for a good reading-group conversation on how WAM ablations get evaluated.","headline":"A credible, well-engineered WAM with real benchmark gains whose core attribution — event-level, caption-rich video prediction drives the action results — is never tested at the action level; send to review, expect revision.","tokens_in":23150,"tokens_out":11775,"duration_ms":66586},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"VPP2 claims that action quality is set by video prediction quality: predict whole captioned manipulation events, distill to one step, then add an action head, reaching 58.5% zero-shot ALOHA success and 45.0% on LIBERO-Pro.","keywords":["world action model","video prediction policy","event-level video pretraining","detailed video captioning","consistency distillation","mixture-of-transformers","zero-shot robot manipulation","instruction following"],"falsifier":"Train VPP2's exact pipeline but replace event-level sampling with fixed-horizon chunk sampling, the VPP2-Fixed-Step configuration, on identical data and captions, then run the zero-shot ALOHA and LIBERO-Pro protocols; success rates matching 58.5% and 45.0% would falsify the paper's attribution of its gains to event-level prediction. A second check strips the detailed template captions down to one-line instructions; if robot success rates stay at VPP2's levels, the detailed-caption claim falls.","tokens_in":23651,"feed_emoji":"🤖","tokens_out":23239,"duration_ms":125982,"temperature":0.7,"pith_summary":"World action models — robot policies that lean on video prediction to generate actions — tend to fail in open-ended environments, and the paper identifies why: video foundation models are tuned for creative generation rather than physical manipulation, and grafting action modules onto them degrades the very generalization they were chosen for. VPP2's answer is to fix the video model first: continue-pretrain a 14B video foundation model on a large, captioned collection of manipulation clips at the level of whole events (a complete sub-task from start to finish), then post-train it for fixed-horizon chunks and distill it to a single forward pass, and only then attach a small mixture-of-transformers action expert that turns the predicted video latents into action chunks. The paper reports that this ordering works: VPP2-14B outperforms a 64B video model at instruction-following (0.80 vs 0.70 on human-hand prompts, 0.90 vs 0.78 on robot prompts), reaches 58.5% success on zero-shot real-world ALOHA tasks versus 40.0% for the strongest baseline, and leads on LIBERO-Pro (45.0%), LIBERO-OOD (63.9%), and RoboDojo (32.26%) after benchmark-specific post-training. The paper's bet is that generalization is won or lost at the video-prediction stage, before actions are ever considered, and that predicting better is what makes the policy act better.","feed_headline":"Full-event video prediction lifts zero-shot robot success to 58.5%","feed_subtitle":"A 14B robot policy beats a 64B video model at instruction following and lifts zero-shot ALOHA success from 40% to 58.5%.","key_machinery":"The load-bearing mechanism is event-level video pretraining with detailed captions. Training examples are whole sub-task events, with $N$ future frames sampled uniformly across the event, $y_{\\text{event}} = (o_0, o_{\\lfloor T/N \\rfloor}, o_{\\lfloor 2T/N \\rfloor}, \\dots, o_T)$, optimized by the flow-matching objective $\\mathcal{L}_{\\text{FM}}(\\theta) = \\mathbb{E}\\left[\\|v_\\theta(x_s, s, c) - (x_1 - \\epsilon)\\|_2^2\\right]$ on the interpolated latent $x_s = (1-s)\\epsilon + s x_1$. This is precisely the factor that separates VPP2 from its own ablation VPP2-Fixed-Step, which predicts a fixed-duration chunk instead and scores 0.48/0.42 versus VPP2's 0.80/0.90 on instruction-following. Four supporting mechanisms carry the rest of the argument: a captioning template that resolves four sources of future uncertainty (task sequence and manner, active end-effector identity, target-object ambiguity, and camera motion); a two-level coordinate alignment using workspace alignment $A_d = \\bar{W}^T W_d$ and end-effector alignment $B_d = E_d^T \\bar{E}$ so that bimanual data from many embodiments share one action space; consistency distillation, which enforces $\\mathcal{L}_{\\text{CD}}$ agreement between adjacent flow times with the terminal condition $F_\\theta(x, 1, c) = x$, enabling single-step video generation; and a mixture-of-transformers action DiT that consumes the video model's latents so that action learning does not backpropagate through the full video diffusion process.","core_discovery":"The central claim is that a world action model's generalization is determined before any action module is trained: everything hinges on whether the video backbone has learned a stable, near-unique mapping from a manipulation instruction to a future trajectory. VPP2 pursues that mapping through event-level continued pretraining, which samples frames uniformly across an entire sub-task event rather than a short fixed window, on clips annotated with detailed captions that name the active end effector, disambiguate the target object, and describe camera motion. The paper's stated rationale is that a detailed caption together with a complete trajectory leaves one concentrated distribution of plausible futures, whereas a short caption paired with a short chunk leaves many competing modes, so the model falls back on spurious short-horizon correlations. The trained video model is then post-trained to a fixed prediction horizon (2 seconds for human hands, 8 seconds for robots), distilled with a consistency objective so a full future chunk is generated in a single roughly 0.12-second forward pass, and finally joined to a 0.9B mixture-of-transformers action expert that emits 2-second action chunks at a total latency near 0.22 seconds. The empirical payload is that this sequence produces strong zero-shot generalization in both prediction and action, and that the improvements concentrate in tasks requiring spatial understanding and novel task specifications.","pith_inferences":["The paper attributes its benchmark gains to event-level prediction with detailed captions, but the only ablation isolating that factor, VPP2-Fixed-Step, is scored by a GPT-based evaluator on 50 image-instruction pairs per domain and is never run on robot success rates; whether the video-side gains transfer to executed actions is an open question the action benchmarks do not settle.","The data pipeline bundles several changes at once — larger and more diverse training data, unified workspace and end-effector coordinate frames, T-shaped multi-view inputs, and detailed captions — so the reported margins probably mix curation effects with the event-level modeling choice; matching ablations measured on ALOHA or LIBERO success would decompose them.","If the recipe transfers, it suggests that other world action models could adopt the same ordering — generalizable video prediction first, distillation second, action expert last — without needing a 64B backbone, since the headline gains come from a 14B model.","The subtask-planning results on RoboDojo (for example, tic-tac-toe rising from 0% to 38%) hint that the remaining bottleneck on long-horizon tasks is high-level sequencing rather than low-level motion prediction, so future gains may come as cheaply from better planners as from better video models."],"forward_implications":["A 14B policy with a manipulation-tuned video backbone can beat a 64B general video foundation model at instruction-following, so model scale is not the binding constraint for world action models.","Video prediction can be made fast enough for closed-loop control: the distilled single-step video planner and the action expert together yield a 2-second action chunk in roughly 0.22 seconds.","Zero-shot deployment on a seen embodiment improves to 58.5% average success over 10 real-world ALOHA task categories, against 40.0% for π0.5 and 20.5% for Fast-WAM, with VPP2 best in 9 of 10 categories.","After training only on the standard LIBERO suites, VPP2 reaches 45.0% on LIBERO-Pro where the strongest baseline manages 11.0%, and 63.9% on LIBERO-OOD; the authors take the large margin on task-specification perturbation as evidence that the policy follows instructions rather than replaying memorized trajectories.","On the 42-task RoboDojo bimanual benchmark, VPP2 reports state-of-the-art results (32.26% success, 39.26 average score), and adding a VLM subtask planner lifts five selected task groups from 27.6% to 57.6% success."],"supporting_citations":[{"why":"Supplies the Wan2.1-I2V-14B base video model that VPP2 continue-pretrains, post-trains, and distills, plus the detailed-caption prompt-enhancement practice.","marker":"(Wan et al., 2025)"},{"why":"Supplies the flow-matching objective used for both event-level pretraining and fixed-horizon post-training.","marker":"(Lipman et al., 2023)"},{"why":"Supplies consistency distillation, the method that compresses the multi-step video model into a single-step planner.","marker":"(Song et al., 2023)"},{"why":"Supplies the mixture-of-transformers architecture used for the action expert.","marker":"(Liang et al., 2025)"},{"why":"The original Video Prediction Policy; establishes the premise that video prediction priors transfer to action learning, which VPP2 extends.","marker":"(Hu et al., 2024)"},{"why":"Cosmos3 is the 64B video foundation model that VPP2-14B outperforms by 11 points on instruction-following success.","marker":"(Agarwal et al., 2026)"},{"why":"π0.5 is the strongest zero-shot ALOHA baseline (40.0%), retrained on VPP2's data for a fair comparison.","marker":"(Intelligence et al., 2025)"},{"why":"LIBERO-Pro supplies the position and task perturbation protocol that measures generalization beyond memorization.","marker":"(Zhou et al., 2025)"},{"why":"Supplies the LIBERO-OOD compositional evaluation protocol and the Temporal Ratio method, the strongest prior approach to that generalization gap.","marker":"(Mishra et al., 2026)"},{"why":"RoboDojo supplies the 42-task bimanual simulation benchmark and the official leaderboard baselines VPP2 is compared against.","marker":"(Chen et al., 2026)"}],"fun_headline_variants":["Event-level video pretraining lifts zero-shot ALOHA success to 58.5%","VPP2 outperforms a 64B video model at instruction following by 11 points","Full-event video predictions improve zero-shot robot actions, VPP2 shows","VPP2 distills future-video prediction into a single-step robot planner","Video prediction policy 2 raises zero-shot ALOHA success by 18.5 points"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The causal story rests on the assumption that automatic video-instruction-following scores predict real robot success: the only ablation isolating event-level prediction with detailed captions (VPP2-Fixed-Step) is evaluated by a GPT-based metric on 50 image-instruction pairs per domain, and is never evaluated on the ALOHA, LIBERO, or RoboDojo action benchmarks that produce the headline numbers.","fun_headline_variants_meta":{"raw":{"variants":["Event-level video pretraining lifts zero-shot ALOHA success to 58.5%","VPP2 outperforms a 64B video model at instruction following by 11 points","Full-event video predictions improve zero-shot robot actions, VPP2 shows","VPP2 distills future-video prediction into a single-step robot planner","Video prediction policy 2 raises zero-shot ALOHA success by 18.5 points"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000617,"raw_usage":{"cost":0.000616756,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000616756,"upstream_inference_prompt_cost":0.000112476,"upstream_inference_completions_cost":0.00050428},"total_tokens":3014,"prompt_tokens":1213,"completion_tokens":1801,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":512,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":1704}},"tokens_in":1213,"tokens_out":1801,"cost_source":"openrouter_usage","duration_ms":10948,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":1704,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-09T01:39:51.445099+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":"Train VPP2's exact pipeline but replace event-level sampling with fixed-horizon chunk sampling, the VPP2-Fixed-Step configuration, on identical data and captions, then run the zero-shot ALOHA and LIBERO-Pro protocols; success rates matching 58.5% and 45.0% would falsify the paper's attribution of its gains to event-level prediction. A second check strips the detailed template captions down to one-line instructions; if robot success rates stay at VPP2's levels, the detailed-caption claim falls.","supporting_citations":[],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":2}