DataClaw0: Agentic Tailoring Multimodal Data from Raw Streams

• Data Tailoring turns noisy, redundant, weakly structured streams into dense, verifiable, intent-aligned supervision, aiming to replace repeated proprietary-VLM annotation with a reusable learned capability.The introduction motivates this framing by noting that direct annotation can produce hallucinated, fragmented, or low-density outputs, while API annotation incurs recurring cost per sample.Source: §1 · pp. 1–2
• A two-stage pipeline extracts deterministic Factual Anchors and uses strong VLMs for long-range reasoning, producing a cross-domain refinement dataset spanning five arenas.Anchors capture low-level grounding such as object states, temporal boundaries, OCR text, and GUI events; the arenas are daily life, education, embodied intelligence, world models, and GUI agents.Source: §1 · pp. 2–3
• DataClaw0 combines SFT with GRPO, using rewards for intent satisfaction, information density, factual consistency, and structural correctness across 4B, 9B, and 27B scales.The paper compares a jointly trained five-domain model with domain-specific experts and a router at each scale.Source: §1 · p. 3
• At 27B, DataClaw0 surpasses larger proprietary annotators on all three DataClaw0-val axes, and tailored data beats downstream-model and proprietary-annotator data across three downstream tasks.The downstream comparisons use a fixed data budget, and the joint-versus-sharded ordering on DataClaw0-val recurs across the three downstream tasks.Source: §1 · p. 3
• At 4B and 9B, DataClaw0-E experts win, while at 27B joint training wins; matched-data comparisons attribute the reversal to cross-domain transfer, with crossover near 18B.The introduction reports experiments at 4B, 9B, and 27B and identifies expert routing as preferable below the threshold and joint training above it.Source: §1 · p. 3
• The 9B variant costs $0.009 per sample versus $0.038 for the teacher API and breaks even after about 80K samples; the 27B variant costs $0.027 and breaks even after about 324K.At a fixed backbone, the joint configuration is cheaper to deploy because it serves one model rather than five; tailoring has a one-off distillation cost followed by local-inference marginal costs.Source: §1 · p. 3

Multimodal Large Language Models

• MLLMs have expanded from image assistants to general-purpose systems, but costly supervision and weak passive annotation limit learning from noisy, high-entropy multimodal streams.Existing heuristic sampling, coarse captioning, and direct VLM question-answer generation struggle with dense, reliable supervision, especially for temporal reasoning, spatial consistency, and hallucination-sensitive settings [24] [26] [34] [35] [32] [33].Source: §2 · p. 4

Agentic AI

• Agentic workflows can organize complex data-generation pipelines, inspiring DataClaw0 to coordinate multimodal models and tools to produce structured, intent-aligned supervision [39] [40] [85].Such workflows can decompose objectives, inspect intermediate outputs, correct errors, and verify results.Source: §2 · p. 4

3.1 Problem Formulation: Intent-Conditioned Data Tailoring

• Unlike passive captioning or question answering, tailoring filters and reorganizes noisy streams for an intent into semantically accurate outputs that conform to a predefined schema.The formulation requires both information filtering and focusing, and structural reorganization.Source: §3.1 · pp. 4–5

3.2 Data Construction Pipeline

• DataClaw0 constructs training triplets by extracting timestamped, spatially located factual anchors, then synthesizing structured supervision conditioned on the raw input, anchors, and domain intent.Gemini-3.1-Pro supplies synthesis once per training example for SFT corpus construction; it is not invoked at tailoring time.Source: §3.2 · pp. 5–6

3.3 Rule-Driven Reinforcement Learning via GRPO

• After SFT, rule-driven GRPO uses deterministic rewards for schema compliance, anchor alignment, and concise reasoning to improve formatting reliability, reduce hallucinations, and strengthen spatio-temporal grounding.The policy update uses group-normalized relative advantages, a clipped objective, and a KL regularizer against the reference model.Source: §3.3 · p. 6

3.4 Inference and Deployment Paradigms

• Inference ingests multimodal input and parses intent, applies schema-constrained policy inference, then verifies grounding; deployment offers a unified omni model or scenario-routed domain experts.The omni configuration favors simplicity and generality, while experts provide domain specialization and modular scalability.Source: §3.4 · p. 6

3.5 Benchmark Construction and Evaluation

• DataClaw0-val samples 200 diverse high-quality examples, while DataClaw0-Intent stress-tests colloquial, ambiguous, or incomplete intents using hierarchical JSON-validity, schema, semantic-alignment, and trajectory-shape evaluation.The validation examples cover diverse multimodal inputs, target schemas, and long-tail cases.Source: §3.5 · pp. 6–7

4.1 Experimental Setup

• Experiments compare joint and expert configurations at 4B, 9B, and 27B, with reward weights and learning rate shared across scales and tuned only on 9B.The models use 34K cleaned instruction-refinement examples for one SFT epoch; the recipe also includes a GRPO phase.Source: §4.1 · p. 7

4.2 Main Results: Joint vs. Sharded Tailoring

• DataClaw0-27B-O leads all three benchmark axes: Field 99.40 vs. Gemini-3.1-Pro-Preview 98.12, Semantic 81.14 vs. GPT-4o 75.15, and Sequence 62.15 vs. Gemini 58.50.On the Fuzzy subset, it scores Semantic 81.27 and Sequence 63.48, versus Gemini’s 79.63 and 47.25, respectively.Source: §4.2 · p. 7, Table 1 · p. 8
• Sharding leads by 14.6 Overall points at 4B and 11.0 at 9B, but joint training leads by 5.9 at 27B, placing the crossover near 18B parameters.At 27B, joint training gains most on AIGC (Semantic 80.36 vs. 71.83), Education (82.51 vs. 74.96), and Fuzzy (81.27 vs. 71.86), while Daily Life slightly favors experts (61.24 vs. 61.82).Source: §4.2 · pp. 7–8, Figure 3 · p. 9

4.3 Downstream Application: Targeted Refinement & Efficiency

• DataClaw0-27B-O leads every downstream metric, including GUI Step Success Rate (43.1 vs. Gemini’s 39.5) and VQA Partial Accuracy (56.7 vs. Gemini’s 53.4).The joint-versus-sharded ordering also flips with scale: experts lead on TSR at 4B (9.8 vs. 8.1) and 9B (15.6 vs. 12.7), while joint training leads at 27B (19.8 vs. 17.4).Source: Table 2 · p. 10, §4.3 · pp. 10–11
• Benchmark scores are strongly rank-correlated with downstream gain (Pearson r = 0.97; Spearman ρ = 0.98 over 13 points), but verbose outputs from the 9B Omni checkpoint and GPT-4o score below the trend.Feature-space analysis and fuzzy-intent evaluation also indicate broader intent coverage and robustness to ambiguous, colloquial, or incomplete requests.Source: §4.3 · p. 11, §4.4 · p. 13

4.4 Is Data Tailoring One Capability or Five?

• With matched in-domain data, the joint model leads on four of five domains, with its advantage largest in data-poor domains; on AIGC it leads by 9.7 points at 10% and 9.4 at 100% of domain data.Only Daily Life favors the expert, slightly, throughout; cross-domain transfers are positive in every off-diagonal pairing, averaging 8.2 points versus 41.0 for in-domain training.Source: §4.4 · pp. 11–12, Figure 5 · p. 11

4.5 Ablation Studies

• Reward ablation’s summed metric change from SFT Only to GRPO w/ Ranchor shifts from −0.4 at 4B to +0.9 at 9B and +6.7 at 27B, while experts nearly collapse when routed outside their own domain.Field is omitted because it saturates at 100.00 for every variant above minimal initialization.Source: Table 4 · p. 12

4.4 Is Data Tailoring One Capability or Five?

• Without seeing any examples from a held-out domain, joint training recovers 59% of the zero-shot-to-fully-trained performance gap on average, including 69% on Daily and 67% on Education.The experiments interpret joint training as more accurate and cheaper to deploy above the capacity threshold, while experts remain an engineering remedy at 9B; experts also collapse outside their own domain.Source: §4.4 · pp. 12–13, Table 3 · p. 12

5 Conclusion

• Above roughly 18B parameters, one joint model beats five domain experts; matched-data results attribute the reversal to transfer, and it recovers 59% of in-domain performance on withheld domains.The joint model receives exactly the same data per domain as each expert in the matched-data comparison.Source: §5 · p. 14
• Tailored data improves downstream GUI navigation, video generation, and spatio-temporal VQA under a fixed budget, reproducing the joint-versus-sharded ordering.At a fixed backbone, the joint model is cheaper to serve.Source: §5 · p. 14
• Reinforcement is harmful at 4B but beneficial at 27B; the larger accuracy-winning backbone needs roughly 4× the production volume to break even.The capacity threshold also governs whether reinforcement helps, as complementary objectives compete for representation below it.Source: §5 · p. 14
• The results do not establish capability beyond the proprietary teacher, generalization beyond the five studied domains and schema family, or the threshold’s shape.The three backbone sizes locate the threshold but do not characterize its shape.Source: §5 · p. 14

A DataClaw0 Dataset and Benchmark

• The appendix documents the corpus, construction pipeline, DataClaw0-val benchmark, fuzzy-intent subset, and structured evaluation metrics to support transparent, reproducible protocols.Source: A DataClaw0 Dataset and Benchmark · p. 22

A.1 Dataset Overview

• The corpus spans five domains—Daily Life, Education, Embodied AI, World Models/AIGC, and GUI Agents—with raw streams, extracted factual anchors, and verified intent-conditioned outputs.Source: §A.1 · p. 22
• Domain-specific intents specify downstream uses and output schemas, including GUI navigation, spatio-temporal VQA, embodied action trajectories, and video generation supervision.Source: §A.1 · p. 22
• Validation videos, GUI sessions, and embodied trajectories are excluded from training pools when identifiable, with repeated GUI and embodied task templates and near-duplicate action sequences also removed.Source: §A.1 · p. 22

A.2 A Complete Data Construction Case

• In embodied manipulation, the constructed supervision predicts the next subtask from temporal evidence and object-state transitions, rather than describing the current scene.After the robot places yellow beads in a blue bowl, the example identifies picking up the remaining green block as the next step.Source: §A.2 · p. 23
• For GUI composition, DataClaw0 aligns semantically compatible fragments, orders them by task progress, and rejects distractors from unrelated applications.In the example, spreadsheet fragments B, A, and C form the task sequence, while fragments D, E, and F are discarded.Source: §A.2 · pp. 23–25
• For daily-life video, DataClaw0 constructs grounded contextual-reasoning questions that infer the surrounding environment from scene layout and visible objects.The example uses a wooden table, books, and household objects to identify a living area or study space.Source: §A.2 · p. 25
• Across three domains, DataClaw0 grounds intent-conditioned structured outputs in factual anchors rather than merely captioning raw streams.The examples emphasize temporal prediction and object-state reasoning, long-horizon task composition with distractor rejection, and contextual scene understanding.Source: §A.2 · pp. 22, 26

A.3 Data Construction Pipeline

• The pipeline separates deterministic preprocessing into compact, verifiable construction units from VLM synthesis under explicit schema constraints.Lightweight scripts identify candidate windows, event fragments, and media resources; the VLM handles higher-level reasoning such as subtask prediction and long-horizon reconstruction.Source: §A.3 · p. 26
• The pipeline discovers candidate segments, packages sampled frames or action and fragment context, then invokes intent-conditioned VLM synthesis with a strict output schema.These stages turn raw videos or GUI trajectories into compact requests tailored to the target construction intent.Source: §A.3 · p. 26
• After synthesis, the pipeline parses generated JSON, validates temporal ranges and required fields, materializes media, and exports normalized samples with artifact mappings.The mappings support downstream training and inspection.Source: §A.3 · p. 26

A.3.1 Robot Data Pipeline

• Robot pipelines select informative windows or keyframes, then use VLM synthesis to create fault-recovery or operation-understanding samples grounded in episode evidence.The fault branch creates counterfactual stalled inputs while grounding recovery in successful future trajectories; the understanding branch converts long episodes into temporally grounded reasoning examples.Source: §A.3.1 · pp. 26–28

A.3.2 GUI Multi-event Construction Pipeline

• GUI processing filters likely multi-event trajectories, refines continuous non-overlapping subtask fragments, then shuffles them for task-order reconstruction.The reconstruction examples target fragment ordering, cross-fragment state tracking, and task-level composition while conservative filtering avoids over-segmenting single-goal workflows.Source: §A.3.2 · pp. 28–29

A.3.3 Unified Output Format and Artifact Management

• Robot and GUI samples share a lightweight semantic JSONL format, while media assets are stored separately and referenced through artifact metadata.Answers use placeholders such as <video_gen> and <image_recall> where generated or recalled media should be inserted.Source: §A.3.3 · pp. 29–30

A.3.4 Validation and Quality Control

• Validation checks JSON schemas, temporal ranges and materialized media, GUI fragment consistency, and resumable, auditable processing.Deterministic preprocessing provides grounding, compactness, and reproducibility, while VLM synthesis supplies high-level reasoning for diverse supervision formats.Source: §A.3.4 · p. 30

A.4 DataClaw0-val

• DataClaw0-val evaluates whether models transform high-entropy multimodal inputs and explicit tailoring intents into schema-aligned training instances.Each example pairs multimodal input, a tailoring intent, a required JSON schema, and a reference output selecting relevant frames or fragments and writing supervision.Source: §A.4 · pp. 30–31

A.4.1 Benchmark Composition

• DataClaw0-val has 200 examples across AIGC/World, Daily Life, Education, Embodied AI, and GUI Agents, while DataClaw0-Intent has 120 underspecified-instruction examples.The scenarios cover generative-video segment selection, navigation and procedural understanding, educational-material reconstruction, robot-operation data construction, and GUI task reconstruction.Source: §A.4.1 · pp. 31–32

A.4.2 Reference Output Construction

• Reference outputs conform to task JSON schemas and select only input-supported frame indices or GUI fragment labels, grounding selections in multimodal evidence.Grounding criteria include plausible robot recovery continuations and GUI fragments that compose a coherent long-horizon task.Source: §A.4.2 · p. 32
• Reference construction follows the stated intent, and DataClaw0-val streams are separated from training to avoid direct leakage.Outputs must serve the intended training use rather than merely describe the media.Source: §A.4.2 · p. 32
• AIGC/World references identify a video-generation training segment, while Education references become interleaved teaching material.Source: §A.4.2 · p. 32
• GUI references reconstruct a plausible original long-horizon task rather than listing every fragment’s contents.Source: §A.4.2 · p. 32

A.4.3 DataClaw0-Intent: Fuzzy-intent Stress Test

• DataClaw0-Intent tests whether a model can infer a useful data-construction objective from vague instructions and produce a structured output from raw media.Unlike DataClaw0-val, it does not specify the target task or output schema; the model must choose suitable training data and informative evidence.Source: §A.4.3 · p. 32
• Because fuzzy-intent videos can support multiple valid training instances, 100 raters judge usefulness, media grounding, structure, and alignment with a reasonable inferred goal instead of exact-match scoring.The 120 samples were each judged by five raters on four five-point Likert dimensions; agreement was Krippendorff’s α = 0.71, so scores are treated as indicative rather than a precise ranking.Source: §A.4.3 · p. 33

A.4.4 Metrics

• The hierarchical score uses a hard JSON-validity gate and weights λ1 = 0.20, λ2 = 0.35, and λ3 = 0.45.Its field score measures ground-truth field coverage and truncated extraneous-field ratio, with robust key matching for minor formatting differences.Source: §A.4.4 · p. 34
• For textual fields, semantic scoring uses embedding-based cosine similarity; spatio-temporal visual scoring uses trajectory-shape similarity.The visual metric normalizes and resamples trajectories to K = 50 and uses trajectory MAE with τs = 0.10.Source: §A.4.4 · p. 34

B Training and Deployment Details

• Training and deployment details cover GRPO training, reward design, and Omni/Expert deployment paradigms; the reported training approach combines SFT initialization with rule-reward-driven GRPO.The paper states that this paradigm stimulates agentic tailoring of structured knowledge, particularly spatio-temporal grounding through Ranchor.Source: B Training and Deployment Details · p. 34, §B.1 · p. 35

B.1 Rule-Driven Reinforcement Learning via GRPO

• GRPO avoids a separate Critic model, while deterministic rule-based rewards assess schema compliance, spatio-temporal alignment, and reasoning efficiency.The anchor reward compares normalized, interpolated trajectories resampled to K = 50; the reasoning penalty uses top-half rollouts as group-wise CoT-length references.Source: §B.1 · pp. 34–35

B.2 Inference Architecture & Deployment Paradigms

• The inference architecture takes raw multimodal streams and user-defined intents to generate schema-aligned structured outputs, emphasizing validity, grounding reliability, and deployment flexibility.Source: §B.2 · p. 35

B.2.1 Core System Architecture

• Three decoupled modules handle ingestion and intent parsing, schema-constrained inference, and structural verification and grounding.Together, they produce outputs intended to be semantically grounded and structurally reliable.Source: §B.2.1 · pp. 35–36
• Raw multimodal streams and user-defined intents are converted into structured context sequences; rule-based scripts currently segment long videos into semantically coherent clips using dataset annotations.The authors plan to extend this preprocessing stage with a long-video model that adaptively identifies event boundaries and intent-relevant subsequences.Source: §B.2.1 · pp. 35–36
• A lightweight schema-aware decoding constraint can suppress continuations that violate the target output schema.Source: §B.2.1 · p. 36
• Verification checks UI components, coordinates, timestamps, and action trajectories against factual anchors, using schema correctness, semantic alignment, and trajectory-shape similarity.These checks are intended to ensure semantic grounding and structural reliability.Source: §B.2.1 · p. 36

B.2.2 Omni vs. Expert Deployment Paradigms

• DataClaw0-O is a unified model jointly trained across domains to serve as a generalist for diverse cross-domain intents.Source: §B.2.2 · p. 36
• DataClaw0-E selects a domain-specific agent for the target scenario or configuration, enabling domain-specific robustness and modular updates.Source: §B.2.2 · p. 36

C.1 Detailed Analysis of Downstream Application

• Across GUI navigation, action video generation, and spatio-temporal VQA, Gemini- and DataClaw0-processed data yield substantial gains over self-refinement, which offers only limited improvement over zero-shot baselines.The comparison uses the same raw streams and equal numbers of valid instances from each data pool.Source: §C.1 · p. 36
• Gemini is slightly better on GUI SSR and VQA Partial Accuracy, whereas DataClaw0 leads on GUI TSR, VQA Overall Accuracy, FVD, and Contact mAP.These results indicate a trade-off between partial or step-level metrics and end-to-end task completion or video-generation metrics.Source: §C.1 · pp. 36–37
• DataClaw0 extracts compact, schema-aligned, task-relevant supervision, which can improve end-to-end performance over verbose general-purpose annotations under the same data budget.The paper presents this targeted supervision as a practical, controllable alternative to proprietary annotators for downstream data construction.Source: §C.1 · p. 37

C.2 Unified Refinement Protocol

• The unified protocol processes identical raw streams with each annotator under the same intent and schema, applies shared filters, samples equal valid-instance counts, and fine-tunes the same base model with identical hyperparameters.Evaluation is on held-out task-specific benchmarks.Source: §C.2 · p. 37

C.3 GUI Navigation

• For GUI navigation, Qwen3.5-4B predicts actions from screenshots, task instructions, and interaction history; evaluation reports Step Success Rate (SSR) and Task Success Rate (TSR).SSR measures matching reference actions or reaching the correct UI state; TSR measures full-task completion within a step limit.Source: §C.3 · p. 37

C.4 Action Video Generation

• For action video generation, Wan2.2-I2V-5B generates short videos from an input image and text prompt, and Figure 8 compares outputs from models trained on different refined datasets.The video fine-tuning hyperparameters are listed in Table 6.Source: §C.4 · pp. 37–38, Table 6 · p. 37

C.5 Spatio-temporal VQA

• For spatio-temporal VQA, Qwen3.5-4B answers questions from video frames, with evaluation using Partial Accuracy and Overall Accuracy.Partial Accuracy credits partially correct structured answers, while Overall Accuracy requires the full answer to match the reference.Source: §C.5 · p. 38

Task interference at 9B

• At 9B, DataClaw0-O’s overall score moves from 53.60 to 47.23, rebounds to 57.84, and oscillates; DataClaw0-E scales monotonically to 68.86.The passage attributes the fluctuations to task interference in the shared weights of the 9B model.Source: §D.1 · pp. 38–39
• At 9B, five extraction tasks with different distributions create gradient conflict in shared weights, while domain-specific experts fit local distributions and avoid that conflict.The passage characterizes the smaller-scale instability as a capacity limitation, not evidence that the domains are fundamentally incompatible.Source: §D.1 · p. 39

Why this does not generalise to larger backbones

• At 4B, the Expert ensemble leads by 14.6 points of Overall; at 9B by 10.5, but at 27B the unified model leads by 5.9.At 27B, the unified model also has a monotone rather than oscillating scaling curve.Source: §D.1 · p. 39

Emergent Diversity and Intent Comprehension

• In t-SNE visualizations, DataClaw0 refinements cover semantic space more broadly and evenly than raw data or Qwen3.5-9B refinements, revealing new clusters and long-tail patterns.The base model’s outputs expand only slightly beyond the raw data, indicating limited diversification.Source: §D.1 · p. 39
• On a subset of vague, high-level instructions, DataClaw0 significantly surpasses Qwen3.5-9B and approaches Gemini 3.1-pro-preview.The passage describes this as a challenging fuzzy-instruction benchmark but provides no performance values.Source: §D.1 · p. 39

E Qualitative Analysis

• The qualitative analysis showcases representative domains through end-to-end cases pairing raw inputs and user intents with DataClaw0 outputs.These cases are presented to exemplify the tailoring process.Source: §E.1 · p. 40

E.1 Successful Case Studies

• Each successful case presents the raw input, user intent, and resulting tailored output.This case-study format makes the end-to-end process visible.Source: §E.1 · p. 40

E.2 Failure Cases

• The sample’s answer correctly describes navigation from the workspace to the bed, but its chain of thought reverses the actual frame order.The generated narrative says movement is from workspace to bed, whereas frames 0 →20 show bed to workspace.Source: §E.2 · p. 40
• Temporal hallucination occurs when multimodal models infer a contextually likely event flow from spatial cues instead of frame chronology.The passage attributes this tendency largely to the base model’s reasoning prior.Source: §E.2 · p. 40
• Strict temporal grounding in long-horizon multimodal reasoning remains an open challenge for current LLM-based architectures.The paper characterizes the observed cases as minor but indicative of this intrinsic difficulty.Source: §E.2 · p. 40

E.1 Successful Case Studies

• Successful case studies include a world-model or video-generation tailoring example.Figure 11 is captioned as a successful world-model or video-generation tailoring case.Source: Figure 11 · p. 41

Cost model

• The break-even model compares a one-off tailoring investment plus local per-sample serving cost with API billing per sample, making production volume central to the cost comparison.Beyond break-even, self-hosting is strictly cheaper and the gap grows linearly with sample volume.Source: F Amortised Cost Analysis · pp. 41–42

Serving cost

• Though 27B costs $3,740 upfront and $0.027/sample versus 9B’s $2,360 and $0.009, joint serving replaces five experts with one model and no router.The single-model figures understate the joint configuration’s deployment advantage: one 27B model occupies less accelerator memory than five 9B experts.Source: F Amortised Cost Analysis · p. 43

Sensitivity

• For 9B, break-even remains 54K–149K samples at marketplace or median accelerator prices; 27B pays off only with competitively priced accelerators and list-tier API billing.At hyperscaler list accelerator prices or batch-tier API billing, 27B does not break even.Source: F Amortised Cost Analysis · p. 43, Table 8 · p. 44

Break-even

• Against Gemini-3.1-Pro at $0.038/sample, break-even is roughly 80K samples for 9B and 324K for 27B; against Kimi-K2.6, 9B needs 327K and 27B never breaks even.Figure 14 plots the break-even volumes; 27B's marginal cost exceeds Kimi-K2.6's per-sample price.Source: F Amortised Cost Analysis · p. 43, Figure 14 · p. 44

Scope and limitations

• The analysis supports self-hosting for repeated production on competitively priced accelerators, not the blanket claim that self-hosting is cheaper.The cost estimates are engineering estimates derived from stated assumptions rather than instrumented measurements.Source: F Amortised Cost Analysis · pp. 42, 44

Non-monetary considerations

• Self-hosting keeps sensitive raw streams on-premise and removes rate limits and provider dependence that can constrain continuous large-scale data production.The privacy consideration is relevant to GUI logs and robot trajectories that may contain screen text, credentials, or workspace imagery.Source: F Amortised Cost Analysis · p. 44

Dependence on a proprietary teacher

• DataClaw0 distills Gemini-3.1-Pro rather than claiming new capability, amortizing teacher-provided capability into a 9B open-weight model with local-inference marginal cost.Its rewards check formatting, anchor consistency, and conciseness, not question informativeness or semantic correctness; human-annotated verifiers and simulator feedback are proposed as teacher-independent signals.Source: G Limitations and Future Work · pp. 44–45

Evaluation designed alongside the method

• DataClaw0-val shares the method’s schema, anchor definitions, and intent taxonomy, so it measures compatibility with the task formulation rather than neutral data-refinement ability.The authors have not established that its scores predict downstream gains or validated the hierarchical scoring weights against human judgment; they call for checkpoint correlation and protocol-complete human evaluation.Source: G Limitations and Future Work · p. 45

Single-reference sequence matching

• Single-reference sequence matching can score valid alternative GUI or embodied action paths as errors, making reported sequence numbers lower bounds on tailoring quality.A goal-conditioned criterion or reference set allowing multiple valid orderings could give a fairer measure.Source: G Limitations and Future Work · p. 45

Scope of the empirical study

• The study’s capacity threshold and downstream gains remain bounded by three model sizes, one backbone family, three tasks, five domains, and fixed schemas.The sweep locates the threshold between 9B and 27B but does not characterize its shape or distinguish parameter count, pretraining data, and architecture; video generation uses a small fine-tuning set, and robustness to substantially different requested schemas is untested.Source: G Limitations and Future Work · p. 45

Reproducibility and data provenance

• Some implementation details are not yet released as code, and raw streams may contain personally identifying content despite use of research-licensed sources and no redistribution.The authors intend to release the construction pipeline and evaluation harness; they did not conduct an independent privacy audit and say consent and de-identification are prerequisites for deployments on newly collected streams.Source: G Limitations and Future Work · p. 45

From refinement toward autonomous construction

• DataClaw0 currently refines user-provided streams rather than creating data from intent alone, with autonomous simulation-based exploration proposed as a future extension.The proposed direction combines embodied, game, or GUI exploration with execution-grounded rewards to construct supervision from the model’s own interactions.Source: G Limitations and Future Work · p. 45