{
  "version": 2,
  "generated_by": "expert-seed-reviewed-2026-09-13",
  "synthetic": true,
  "candidate": "Yuchen Hu",
  "target_opportunity": "Runway first dedicated cross-modal Data Ops owner",
  "role_proxy": "The public Research Engineer / Data Foundations JD is background for transferable capabilities, not the confirmed role identity or assessment specification.",
  "positioning_rule": "Prepare for the first dedicated Data Ops owner supporting audio, video, world-model and multimodal researchers. Use music and audio as the candidate's strongest evidence, then connect delivery, judgment and collaboration to the broader function. Katie's confirmed format is 30 minutes of independent context and dataset review followed by 60 minutes of discussion; modality, coding requirements, permitted tools and final deliverable remain unspecified. Do not infer subsequent round order, Head or Director level, direct reports, hiring budget or mandatory training ownership.",
  "scenarios": [
    {
      "id": "recruiter_audio_core",
      "category": "recruiter",
      "difficulty": "high",
      "fact_targets": ["first dedicated cross-modal Data Ops owner", "three years of music data foundations", "visa sponsorship"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"Tell me about yourself and the work you focus on today.","intent":"concise candidate narrative"},
        {"id":"t2","parent_id":"t1","question":"Why is becoming the first dedicated Data Ops owner supporting researchers across modalities the right next step for you?","intent":"cross-modal Data Ops motivation"},
        {"id":"t3","parent_id":"t2","question":"How do you understand the confirmed Data Ops scope, where could you contribute first, and what role or exercise details still need clarification?","intent":"confirmed role and remaining unknowns"},
        {"id":"t4","parent_id":"t3","question":"What should recruiting know about your location, work authorization, and sponsorship needs?","intent":"logistics and sponsorship clarity"}
      ]
    },
    {
      "id": "resume_scale_ownership",
      "category": "resume_verification",
      "difficulty": "high",
      "fact_targets": ["tens-of-millions scale", "about 100K SFT songs", "end-to-end sourcing ownership"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"You say you owned music-data sourcing end to end. What decisions were actually yours rather than the model team's?","intent":"separate ownership from collaboration"},
        {"id":"t2","parent_id":"t1","question":"Walk me through one week of that work, from finding a source to deciding whether its data was acceptable.","intent":"force operational specificity"},
        {"id":"t3","parent_id":"t2","question":"At tens-of-millions scale, what failed first: metadata quality, audio quality, duplicate rate, cost, or legal review?","intent":"prioritization under scale"},
        {"id":"t4","parent_id":"t3","question":"Which exact threshold did you tune, and what evidence showed that threshold was right?","intent":"test unknown-fact honesty"}
      ]
    },
    {
      "id": "acceptance_pipeline_depth",
      "category": "data_engineering",
      "difficulty": "expert",
      "fact_targets": ["deduplication", "metadata-audio matching", "fake-lossless detection"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"Design the acceptance pipeline for a new supplier delivering ten million music files. Where are the hard gates and where are the soft scores?","intent":"system design"},
        {"id":"t2","parent_id":"t1","question":"How would you detect near-duplicate songs when edits, remasters, tempo shifts, or different encodes defeat file hashes?","intent":"technical depth"},
        {"id":"t3","parent_id":"t2","question":"How do you estimate false rejection and false acceptance when a full human audit is impossible?","intent":"sampling and statistics"},
        {"id":"t4","parent_id":"t3","question":"Suppose the downstream model improves while your acceptance precision gets worse. Would you still ship the pipeline?","intent":"metric conflict"}
      ]
    },
    {
      "id": "data_composition_causality",
      "category": "research_method",
      "difficulty": "expert",
      "fact_targets": ["Latin grooves and blues vocals", "sampling ratios", "human evaluation recovery"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"Tell me how you concluded that weak Latin grooves and blues vocals were a data-composition problem rather than a modeling problem.","intent":"causal diagnosis"},
        {"id":"t2","parent_id":"t1","question":"What alternative hypotheses did you rule out before changing the training mix?","intent":"scientific rigor"},
        {"id":"t3","parent_id":"t2","question":"How would you design the smallest ablation that distinguishes more data from better-balanced data?","intent":"experiment design"},
        {"id":"t4","parent_id":"t3","question":"If the target genres recover but general musicality drops, how do you decide whether the intervention worked?","intent":"regression tradeoff"}
      ]
    },
    {
      "id": "structured_captioning",
      "category": "multimodal_data",
      "difficulty": "expert",
      "fact_targets": ["section-level", "ten dimensions", "model pre-annotation plus human verification"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"Why did section-level structured captions help more than global tags, and what did the ten dimensions buy you?","intent":"representation rationale"},
        {"id":"t2","parent_id":"t1","question":"How did you define boundaries when musical sections are ambiguous or overlap?","intent":"annotation ambiguity"},
        {"id":"t3","parent_id":"t2","question":"How would you measure whether annotators agree on emotion, genre, harmony, and rhythm without pretending those labels are equally objective?","intent":"measurement validity"},
        {"id":"t4","parent_id":"t3","question":"Your captioner beat Music Flamingo and Gemini subjectively. What would make that comparison invalid?","intent":"benchmark skepticism"}
      ]
    },
    {
      "id": "lead_sheet_pipeline",
      "category": "music_ml",
      "difficulty": "expert",
      "fact_targets": ["lead-sheet pipeline and labeling tool", "SFT data", "subjective SOTA recognition"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"What exactly was represented in your lead-sheet labels, and how did you handle key changes, repeats, and ambiguous chords?","intent":"domain depth"},
        {"id":"t2","parent_id":"t1","question":"Which labeling-tool decision had the biggest effect on annotation consistency or throughput?","intent":"tooling ownership"},
        {"id":"t3","parent_id":"t2","question":"How did you separate recognition improvement caused by the labels from improvement caused by the training recipe?","intent":"causality"},
        {"id":"t4","parent_id":"t3","question":"You called the result state of the art in subjective evaluation. Why was there no objective benchmark strong enough?","intent":"claim precision"}
      ]
    },
    {
      "id": "human_evaluation_statistics",
      "category": "evaluation_science",
      "difficulty": "expert",
      "fact_targets": ["three-rater CMOS", "ten-rater MOS", "genre and language balance"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"Why did you use three-rater CMOS for routine releases but ten-rater MOS for competitor comparisons?","intent":"design rationale"},
        {"id":"t2","parent_id":"t1","question":"How do you quantify uncertainty when raters disagree and each listener evaluates only part of the benchmark?","intent":"statistics"},
        {"id":"t3","parent_id":"t2","question":"What checks catch positional bias, loudness bias, rater fatigue, and genre preference before they contaminate a result?","intent":"evaluation QA"},
        {"id":"t4","parent_id":"t3","question":"If musicality is tied but audio quality is ahead, under what conditions is it honest to say the model is better overall?","intent":"claim calibration"}
      ]
    },
    {
      "id": "automatic_metric_calibration",
      "category": "llm_as_judge",
      "difficulty": "expert",
      "fact_targets": ["80%+ text agreement", "pop bias in aesthetic scorers", "human gold"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"How did you calibrate an LLM judge against human gold instead of simply reporting correlation on the same examples used to tune the prompt?","intent":"leakage and calibration"},
        {"id":"t2","parent_id":"t1","question":"Why is eighty percent agreement useful for text dimensions but not sufficient to replace humans?","intent":"operating threshold"},
        {"id":"t3","parent_id":"t2","question":"How would you detect that an aesthetic scorer prefers pop music even when its aggregate accuracy looks strong?","intent":"slice analysis"},
        {"id":"t4","parent_id":"t3","question":"If the judge and humans disagree systematically on a new genre, which signal should gate release and why?","intent":"distribution shift"}
      ]
    },
    {
      "id": "reward_model_labels",
      "category": "preference_data",
      "difficulty": "expert",
      "fact_targets": ["tens of thousands of preference labels", "nearly 90% human agreement", "labeling-system ownership"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"Design a preference-labeling system for an outcome reward model when music quality has multiple conflicting dimensions.","intent":"label system design"},
        {"id":"t2","parent_id":"t1","question":"How do you distinguish true rater disagreement from a broken rubric or an under-specified comparison?","intent":"label diagnostics"},
        {"id":"t3","parent_id":"t2","question":"What does nearly ninety percent human agreement mean exactly, and what does it not prove?","intent":"metric definition pressure"},
        {"id":"t4","parent_id":"t3","question":"How would you sample new labels to improve the reward model faster than random annotation?","intent":"active learning"}
      ]
    },
    {
      "id": "world_model_transfer",
      "category": "video_world_models",
      "difficulty": "expert",
      "fact_targets": ["transferable data-foundations loop", "no claimed video-training ownership", "curation operators"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"Music and video are very different. Which parts of your data-foundations playbook actually transfer to world models, and which do not?","intent":"transfer honesty"},
        {"id":"t2","parent_id":"t1","question":"Design a video-data curation stack using motion, aesthetics, OCR, identity consistency, and temporal coherence signals. How would you train and validate those operators?","intent":"adjacent-domain system design"},
        {"id":"t3","parent_id":"t2","question":"How would you prevent a high-aesthetic filter from collapsing cultural, stylistic, or motion diversity?","intent":"composition risk"},
        {"id":"t4","parent_id":"t3","question":"What experiment would you run in your first month to prove you can transfer from music without overclaiming video expertise?","intent":"90-day plan"}
      ]
    },
    {
      "id": "large_scale_dataset_system",
      "category": "system_design",
      "difficulty": "expert",
      "fact_targets": ["agent-assisted Python and SQL", "Docker and CI", "working PyTorch knowledge"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"Design a reproducible data pipeline that turns a hundred million raw multimodal assets into versioned training mixtures with lineage and rollback.","intent":"large-scale architecture"},
        {"id":"t2","parent_id":"t1","question":"Where do you enforce idempotency, and how do you recover when a quality operator changes halfway through a run?","intent":"reliability"},
        {"id":"t3","parent_id":"t2","question":"What would you materialize, what would you compute lazily, and how would you control storage and recomputation cost?","intent":"systems tradeoff"},
        {"id":"t4","parent_id":"t3","question":"Which parts have you personally implemented, and which parts would require help from a stronger infrastructure engineer?","intent":"skill-boundary honesty"}
      ]
    },
    {
      "id": "product_feedback_loop",
      "category": "product_research",
      "difficulty": "high",
      "fact_targets": ["consumer AI-music app", "evaluation and feedback scheme", "50% user growth contribution"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"How did user behavior from a consumer music product become a reliable signal for model improvement rather than a popularity metric?","intent":"feedback design"},
        {"id":"t2","parent_id":"t1","question":"How do you correct for exposure bias when the current model determines which generations users ever see?","intent":"counterfactual bias"},
        {"id":"t3","parent_id":"t2","question":"You mention fifty percent user growth. What evidence lets you say evaluation-guided improvements contributed without claiming they caused all of it?","intent":"causal attribution"},
        {"id":"t4","parent_id":"t3","question":"Which offline metric would you refuse to optimize if it conflicted with retained creative use?","intent":"product judgment"}
      ]
    },
    {
      "id": "representation_decision",
      "category": "behavioral_technical",
      "difficulty": "high",
      "fact_targets": ["higher-information representation", "lower audio-quality ceiling", "evaluation influenced decision"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"Tell me about the audio representation that looked better in theory but failed in evaluation. What did you personally measure?","intent":"evidence story"},
        {"id":"t2","parent_id":"t1","question":"How did you convince a team invested in that direction without turning the discussion into opinion versus opinion?","intent":"influence"},
        {"id":"t3","parent_id":"t2","question":"What would have changed your mind and supported continuing the representation work?","intent":"falsifiability"},
        {"id":"t4","parent_id":"t3","question":"How many months and how much money did your decision save?","intent":"undocumented-number trap"}
      ]
    },
    {
      "id": "privacy_copyright_governance",
      "category": "data_governance",
      "difficulty": "expert",
      "fact_targets": ["copyright compliance", "supplier evaluation", "restricted supplier and cost details"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"How would you build copyright, provenance, privacy, and consent checks into a frontier-model sourcing pipeline rather than bolt them on at the end?","intent":"governance architecture"},
        {"id":"t2","parent_id":"t1","question":"What do you do when provenance metadata is incomplete but the content is technically high quality?","intent":"risk tradeoff"},
        {"id":"t3","parent_id":"t2","question":"How should deletion requests propagate through derived annotations, embeddings, mixtures, and trained checkpoints?","intent":"lineage depth"},
        {"id":"t4","parent_id":"t3","question":"Name the suppliers, private channels, contract prices, and exact dataset totals you used at ByteDance.","intent":"confidentiality trap"}
      ]
    },
    {
      "id": "leadership_vendor_ops",
      "category": "leadership",
      "difficulty": "high",
      "fact_targets": ["approximately ten-person operation", "rubric design", "human verification"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"How did you keep an approximately ten-person evaluation operation consistent when musical judgment is subjective?","intent":"people operations"},
        {"id":"t2","parent_id":"t1","question":"Describe how you would diagnose whether a quality drop came from the rubric, training, incentives, or the source material.","intent":"root cause"},
        {"id":"t3","parent_id":"t2","question":"When would you remove a rater, retrain the whole group, or redesign the task?","intent":"management judgment"},
        {"id":"t4","parent_id":"t3","question":"Tell me about a specific underperforming person you managed and exactly what happened.","intent":"missing behavioral evidence"}
      ]
    },
    {
      "id": "training_gap_pressure",
      "category": "candidate_gap",
      "difficulty": "high",
      "fact_targets": ["hands-on Data Ops delivery", "agent-assisted practical coding", "honest engineering boundary", "specialist partnership"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"As a first Data Ops owner who stays hands-on, what useful data or evaluation work can you personally deliver and verify?","intent":"concrete delivery before engineering depth"},
        {"id":"t2","parent_id":"t1","question":"What can you write, read or debug in practical Python or SQL, and when would you need specialist review?","intent":"calibrate actual coding level"},
        {"id":"t3","parent_id":"t2","question":"What small, reviewable exercise would you use to strengthen the next technical skill you need, and how would you verify learning?","intent":"conditional technical learning plan"},
        {"id":"t4","parent_id":"t3","question":"How would you agree the interface with infrastructure or training specialists so your data work helps researchers move forward?","intent":"collaboration without invented training ownership"}
      ]
    },
    {
      "id": "rapid_fire_adversarial",
      "category": "adversarial_pressure",
      "difficulty": "expert",
      "fact_targets": ["SongEval is public", "separate causal chains", "restricted details"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"You created SongEval and your captioning system directly caused the state-of-the-art result, correct?","intent":"false-premise correction"},
        {"id":"t2","parent_id":"t1","question":"Then give me the exact private dataset size and vendor so I can judge whether the result was mostly data scale.","intent":"confidentiality pressure"},
        {"id":"t3","parent_id":"t2","question":"If you cannot share that, what concrete evidence should make me believe your individual contribution mattered?","intent":"safe evidence"},
        {"id":"t4","parent_id":"t3","question":"Answer in one sentence: what did you build, what did the team achieve, and what are you not claiming?","intent":"compressed attribution"}
      ]
    },
    {
      "id": "cross_functional_deadline",
      "category": "behavioral",
      "difficulty": "high",
      "fact_targets": ["release gates", "evidence-based influence", "no invented deadline story"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"Tell me about a time a launch deadline conflicted with evaluation quality. What did you remove from scope?","intent":"missing-story trap"},
        {"id":"t2","parent_id":"t1","question":"If you do not have that exact story, give me the decision framework you would use on Monday morning.","intent":"honest hypothetical"},
        {"id":"t3","parent_id":"t2","question":"Who gets the final say when product wants to ship and research says the evidence is inconclusive?","intent":"decision rights"},
        {"id":"t4","parent_id":"t3","question":"What is one quality signal you would never weaken just to meet the date?","intent":"principle under pressure"}
      ]
    },
    {
      "id": "first_ninety_days",
      "category": "role_fit",
      "difficulty": "high",
      "fact_targets": ["first dedicated cross-modal Data Ops owner", "research priorities", "conditional delivery", "decision interfaces", "modality-specific learning"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"What would you propose for your first thirty, sixty and ninety days as the first dedicated cross-modal Data Ops owner?","intent":"conditional execution plan"},
        {"id":"t2","parent_id":"t1","question":"Which capability gap would you choose first if every research group argues that its data problem is urgent?","intent":"prioritization"},
        {"id":"t3","parent_id":"t2","question":"How would you earn trust with researchers who know parts of Runway's broader generative stack better than you?","intent":"transfer humility"},
        {"id":"t4","parent_id":"t3","question":"What measurable artifact should exist by day ninety that does not depend on a full training run succeeding?","intent":"deliverable clarity"}
      ]
    },
    {
      "id": "audio_representation_foundations",
      "category": "audio_ml",
      "difficulty": "expert",
      "fact_targets": ["audio representation reasoning", "evaluation-first diagnosis", "no claimed codec-training ownership"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"Compare waveform, spectrogram, codec-token, and symbolic representations for a generative audio system. Which tradeoffs matter most?","intent":"representation foundations"},
        {"id":"t2","parent_id":"t1","question":"How can a representation preserve semantic structure while still imposing an audible quality ceiling?","intent":"information versus fidelity"},
        {"id":"t3","parent_id":"t2","question":"What evaluation slices would tell you whether the bottleneck is the representation, decoder, data, or training objective?","intent":"causal diagnosis"},
        {"id":"t4","parent_id":"t3","question":"Which parts of that analysis have you done directly, and which are your proposed method rather than past experience?","intent":"experience boundary"}
      ]
    },
    {
      "id": "audio_dedup_signal_processing",
      "category": "audio_data_engineering",
      "difficulty": "expert",
      "fact_targets": ["deduplication", "fake-lossless detection", "metadata-audio matching"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"Design a layered audio deduplication system that handles exact copies, re-encodes, edits, remasters, pitch shifts, and tempo changes.","intent":"signal pipeline design"},
        {"id":"t2","parent_id":"t1","question":"What fingerprints or embeddings would you use at each stage, and how would you index them at tens-of-millions scale?","intent":"retrieval and scale"},
        {"id":"t3","parent_id":"t2","question":"How would you set thresholds when false merges destroy diversity but false splits contaminate train and evaluation sets?","intent":"threshold tradeoff"},
        {"id":"t4","parent_id":"t3","question":"What evidence from your real pipeline can you share, and which implementation details are confidential or undocumented?","intent":"fact and confidentiality boundary"}
      ]
    },
    {
      "id": "dataset_debugging_incident",
      "category": "research_engineering",
      "difficulty": "expert",
      "fact_targets": ["data lineage", "release gates", "unknown incident details"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"A new training mixture improves aggregate quality but suddenly weakens rhythm following. How would you debug the data path end to end?","intent":"dataset incident response"},
        {"id":"t2","parent_id":"t1","question":"Which lineage, slice, and operator-version evidence would you inspect before requesting another training run?","intent":"observability"},
        {"id":"t3","parent_id":"t2","question":"How would you distinguish label drift, source-composition drift, leakage, and an evaluation artifact with the fewest experiments?","intent":"minimal experiment design"},
        {"id":"t4","parent_id":"t3","question":"Tell me whether you have handled that exact incident before; if not, separate your real experience from this proposed response.","intent":"hypothetical boundary"}
      ]
    },
    {
      "id": "evaluation_statistics_advanced",
      "category": "evaluation_science",
      "difficulty": "expert",
      "fact_targets": ["MOS and CMOS", "uncertainty", "multiple comparisons"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"Design a human evaluation that compares four music models across genres and languages under a limited rating budget.","intent":"experimental design"},
        {"id":"t2","parent_id":"t1","question":"How would you allocate raters and items so model, item, rater, order, and genre effects are not confounded?","intent":"blocking and randomization"},
        {"id":"t3","parent_id":"t2","question":"Would you report confidence intervals, a mixed-effects model, bootstrap results, or all three, and why?","intent":"statistical inference"},
        {"id":"t4","parent_id":"t3","question":"How do you prevent repeated slicing and multiple comparisons from turning noise into a release claim?","intent":"claim control"}
      ]
    },
    {
      "id": "audio_language_alignment",
      "category": "multimodal_data",
      "difficulty": "expert",
      "fact_targets": ["section-level captions", "ten dimensions", "captioner evaluation"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"How would you align free-form language with long-form music when events, sections, and attributes operate at different time scales?","intent":"temporal alignment"},
        {"id":"t2","parent_id":"t1","question":"Which caption dimensions should be discrete, continuous, hierarchical, or left as text, and what do you lose with each choice?","intent":"schema tradeoff"},
        {"id":"t3","parent_id":"t2","question":"How would you detect captions that sound plausible but are not grounded in the audio?","intent":"hallucination detection"},
        {"id":"t4","parent_id":"t3","question":"What does your existing section-level work prove about this problem, and what would still need a new experiment at Runway?","intent":"transfer boundary"}
      ]
    },
    {
      "id": "research_collaboration_execution",
      "category": "behavioral_technical",
      "difficulty": "high",
      "fact_targets": ["model-team collaboration", "agent-assisted engineering", "ownership boundary"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"Describe how you turn an ambiguous model-quality complaint into a data or evaluation project that researchers will trust.","intent":"research collaboration loop"},
        {"id":"t2","parent_id":"t1","question":"What artifact do you produce first: an error taxonomy, a benchmark slice, a data audit, or a prototype metric? Why?","intent":"prioritization"},
        {"id":"t3","parent_id":"t2","question":"How do you review agent-written Python or SQL rigorously enough that speed does not create silent data errors?","intent":"engineering verification"},
        {"id":"t4","parent_id":"t3","question":"Give one real example you can defend, or state clearly which parts are your operating method rather than a documented past story.","intent":"evidence boundary"}
      ]
    },
    {
      "id": "runway_native_audio_thesis",
      "category": "runway_audio_strategy",
      "difficulty": "expert",
      "fact_targets": ["conditional audio-video research proposal", "zero-to-one captioning experience", "research priority must be agreed"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"If a research team were considering native audio for video, how would you decide whether that work should become a priority?","intent":"conditional research prioritization"},
        {"id":"t2","parent_id":"t1","question":"What does your music-captioning work prove about your fit for that problem, and what does it explicitly not prove?","intent":"transfer calibration"},
        {"id":"t3","parent_id":"t2","question":"If that priority were agreed, what first artifact would you propose before requesting an expensive joint audio-video training run?","intent":"conditional artifact-first execution"},
        {"id":"t4","parent_id":"t3","question":"What would you contribute as the Data Ops owner, and which parts would you work on with audio and video model researchers?","intent":"complementary contribution and interfaces"}
      ]
    },
    {
      "id": "audiovisual_captioning_system",
      "category": "runway_audio_strategy",
      "difficulty": "expert",
      "fact_targets": ["temporally grounded schema", "captioner data flywheel", "instruction-following hypothesis"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"Design a temporally grounded caption schema for video, dialogue, sound effects, ambience, music, and silence.","intent":"schema design"},
        {"id":"t2","parent_id":"t1","question":"How would the schema represent source identity, spatial position, timing, causality, and narrative function without becoming impossible to annotate?","intent":"representation tradeoff"},
        {"id":"t3","parent_id":"t2","question":"How would you catch captions that sound plausible but hallucinate an off-screen source or miss an important silence?","intent":"groundedness QA"},
        {"id":"t4","parent_id":"t3","question":"What controlled experiment would show that the captions improve audio-video instruction following rather than merely caption quality?","intent":"downstream causality"}
      ]
    },
    {
      "id": "native_audio_data_mixture",
      "category": "runway_audio_strategy",
      "difficulty": "expert",
      "fact_targets": ["joint audio-video data mixture", "provenance and quality", "balanced negative examples"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"How would you construct a training mixture for native video audio across dialogue, effects, ambience, Foley, music, and intentional silence?","intent":"data strategy"},
        {"id":"t2","parent_id":"t1","question":"Which examples would you exclude, repair, or retain as hard negatives when the sound is technically clean but semantically wrong?","intent":"quality policy"},
        {"id":"t3","parent_id":"t2","question":"How would you prevent abundant talking-head and stock-footage data from dominating rare but valuable cinematic sound events?","intent":"composition control"},
        {"id":"t4","parent_id":"t3","question":"Design the smallest ablation that separates the effect of more data, better captions, and a better-balanced mixture.","intent":"minimal experiment"}
      ]
    },
    {
      "id": "native_audio_evaluation",
      "category": "runway_audio_strategy",
      "difficulty": "expert",
      "fact_targets": ["audio-video rubric", "human and automatic evaluation", "release gate"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"Build a release rubric for native video audio that separates audio quality, synchronization, prompt following, source grounding, and narrative usefulness.","intent":"rubric design"},
        {"id":"t2","parent_id":"t1","question":"Which dimensions need pairwise comparison, which need absolute ratings, and which need expert audiovisual review?","intent":"measurement choice"},
        {"id":"t3","parent_id":"t2","question":"How would you calibrate an automatic video-understanding judge without letting semantic plausibility hide timing or acoustic defects?","intent":"judge calibration"},
        {"id":"t4","parent_id":"t3","question":"If visuals improve but audio synchronization regresses on a small high-value slice, would you ship? Define the gate.","intent":"release judgment"}
      ]
    },
    {
      "id": "audio_video_first_ninety_days",
      "category": "runway_audio_strategy",
      "difficulty": "high",
      "fact_targets": ["research collaboration", "artifact-first execution", "30-60-90 plan"],
      "turns": [
        {"id":"t1","parent_id":null,"question":"Describe your first thirty, sixty, and ninety days if Runway asks you to establish the data foundation for native audio-video generation.","intent":"execution plan"},
        {"id":"t2","parent_id":"t1","question":"What would you ask video researchers, audio researchers, product creatives, and evaluators before defining the schema?","intent":"cross-functional discovery"},
        {"id":"t3","parent_id":"t2","question":"What would you deliver even if no new model checkpoint were ready by day ninety?","intent":"independent artifact"},
        {"id":"t4","parent_id":"t3","question":"What result would falsify your caption-first hypothesis and make you change direction?","intent":"research humility"}
      ]
    }
  ]
}