Datasets:

NYU-OLAB
/

MedDistractNotes

+Creative Commons Attribution 4.0 International (CC BY 4.0)
+https://creativecommons.org/licenses/by/4.0/legalcode
+See README.md for source attribution and modifications.

README.md ADDED Viewed

	@@ -0,0 +1,98 @@

+---
+language:
+- en
+license: cc-by-4.0
+tags:
+- medical
+- robustness
+- benchmark
+- arxiv:2610.08585
+pretty_name: MedDistractNotes
+configs:
+- config_name: pairs
+  data_files:
+  - split: test
+    path: data/ambient/pairs_v3.parquet
+  default: true
+- config_name: insertions
+  data_files:
+  - split: test
+    path: data/ambient/insertions_v3.parquet
+- config_name: organ_systems
+  data_files:
+  - split: test
+    path: data/ambient/encounter_systems.parquet
+- config_name: notes
+  data_files:
+  - split: test
+    path: results/raw_v3/amb_notes.parquet
+- config_name: judgments
+  data_files:
+  - split: test
+    path: results/raw_v3/amb_judgments.parquet
+- config_name: attribution
+  data_files:
+  - split: test
+    path: results/raw_v3/amb_attribution.parquet
+- config_name: single_note_judgments
+  data_files:
+  - split: test
+    path: results/judge_control/judgments.parquet
+- config_name: failure_modes
+  data_files:
+  - split: test
+    path: results/v3/amb_failure_notes.parquet
+---
+# MedDistractNotes
+MedDistractNotes tests whether incidental conversation enters clinical notes. It contains **1,152 clean–distracted transcript pairs from 576 encounters**, with one bystander and one nonliteral aside per encounter. The clean and distracted transcripts differ only by the inserted exchange. Use the paired inputs to evaluate your own note-generation model; the additional configurations contain the frozen study outputs.
+```python
+from datasets import load_dataset
+pairs = load_dataset("NYU-OLAB/MedDistractNotes", "pairs", split="test")
+row = pairs[0]
+print(row["clean_transcript"], row["distracted_transcript"])
+```
+The `test` split is the entire evaluation benchmark. `source_dataset` preserves the original corpus split; these are not new training/development partitions. The same encounter appears in both perturbation families, so split by encounter if making your own partitions.
+## Contents and keys
+| Configuration | Rows | Contents |
+|---|---:|---|
+| pairs | 1,152 | Clean/distracted transcripts, reference note, inserted exchange, target summary and insertion position |
+| insertions | 1,152 | Frozen GPT-generated exchanges, assigned content, organ system and generation metadata |
+| organ_systems | 577 | Labels/rationales for the original cohort, including the one excluded single-speaker encounter |
+| notes | 18,432 | Saved clean/distracted notes from eight models |
+| judgments | 18,432 | Paired contamination/severity judgments and note-quality scores |
+| attribution | 3,921 | Attribution and clinical-use classifications for flagged notes |
+| single_note_judgments | 18,432 | Matched single-note control scores and explanations |
+| failure_modes | 9,216 | Frozen outcome, failure-mode and section-location annotations for distracted notes |
+Pair keys are `source_dataset`, `item_id`, `distractor_type`. Add `model` and `condition` to join per-note results. `model` retains the identifiers used in the frozen runs. Judge explanations are model-generated, not clinical adjudications. Two empty saved note records remain in the primary denominator; absent and unresolved attribution classifications are retained. Missing scores must not be treated as zero.
+`protocol/prompts.json` contains the generation and scoring templates. Source data, outputs and scores are unchanged from the checked submission archive; only provider bookkeeping was removed from the single-note control export. `release_manifest.json` records content hashes and `data_schema.json` describes every column. The toolkit supports downloading these files into the expected analysis paths.
+## Sources and license
+Derived from [ACI-Bench](https://github.com/wyim/aci-bench) (Yim et al., 2023) and [MTS-Dialog](https://github.com/abachaa/MTS-Dialog) (Ben Abacha et al., 2023), both under CC BY 4.0. The released cohort uses 77 ACI-Bench encounters and 499 MTS-Dialog excerpts. The original transcripts and reference notes are retained; the changes are generated incidental exchanges and the associated model outputs/annotations. Cite the source corpora as well as this benchmark. This release uses CC BY 4.0 with source attribution retained.
+## Intended use and limitations
+For research on contamination, attribution and robustness in note generation. The asides are synthetic and the reference notes come from the source corpora. The scores are automated judgments, not clinician-validated safety labels. High general note-quality scores do not establish that a note is free of contamination. Do not use the benchmark as patient-care guidance.
+Code: https://github.com/nyuolab/llm_distract (the public toolkit release is being prepared).
+## Citation
+[Paper](https://arxiv.org/abs/2610.08585). The citation below uses the publicly posted v1 title; the authors have submitted a revised title. The scientific data are frozen to the submitted study.
+```bibtex
+@misc{vishwanath2026incidental,
+  title={Incidental information contaminates patient notes and disrupts clinical reasoning in large language models},
+  author={Vishwanath, Krithik and Ye, Brandon and Alyakin, Anton and Markert, John E. and Hsieh, Aaron and Mańkowski, Michał and Oermann, Eric K.},
+  year={2026}, eprint={2610.08585}, archivePrefix={arXiv}, primaryClass={cs.CL},
+  url={https://arxiv.org/abs/2610.08585}
+}
+```

data/ambient/encounter_systems.parquet ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:30465d8b6e0f073ee069f815b543617303caacd547bac3385fbda83890bf1652
+size 24579

data/ambient/insertions_v3.parquet ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:03b908525decf30395ba660f7162f9f014d4eb6b9f69c15469440fd9d5feb20c
+size 146785

data/ambient/pairs_v3.parquet ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:c406efc2534829a4b1bc78b7a469e9c69f92cb0aec1fa33f23376ce60a4d2d45
+size 1126980

data_schema.json ADDED Viewed

	@@ -0,0 +1,433 @@

+{
+  "missing_values": "Retained from frozen inputs; missing does not mean negative or zero.",
+  "configs": {
+    "pairs": {
+      "rows": 1152,
+      "split": "test",
+      "columns": {
+        "source_dataset": {
+          "type": "large_string",
+          "description": "Original corpus and split identifier."
+        },
+        "family": {
+          "type": "large_string",
+          "description": "Source corpus family."
+        },
+        "item_id": {
+          "type": "large_string",
+          "description": "Encounter identifier within source_dataset."
+        },
+        "distractor_type": {
+          "type": "large_string",
+          "description": "Perturbation family or audio donor/level identifier; clean for unmodified audio."
+        },
+        "clean_transcript": {
+          "type": "large_string",
+          "description": "Unmodified source transcript (Notes) or saved ASR of clean recording (Audio)."
+        },
+        "distracted_transcript": {
+          "type": "large_string",
+          "description": "Transcript with inserted conversation (Notes) or saved ASR of mixed recording (Audio)."
+        },
+        "reference_note": {
+          "type": "large_string",
+          "description": "Reference clinical note supplied by original source corpus."
+        },
+        "distractor_conversation": {
+          "type": "large_string",
+          "description": "Inserted exchange or human-transcribed donor speech."
+        },
+        "insert_after_turn": {
+          "type": "int64",
+          "description": "One-based speaker-turn insertion position."
+        },
+        "distractor_summary": {
+          "type": "large_string",
+          "description": "Target content supplied to the contamination judge."
+        },
+        "distractor_topic": {
+          "type": "large_string",
+          "description": "Generated aside topic or audio donor-content label."
+        },
+        "generation_model": {
+          "type": "large_string",
+          "description": "Identifier of insertion-generation model."
+        }
+      }
+    },
+    "insertions": {
+      "rows": 1152,
+      "split": "test",
+      "columns": {
+        "source_dataset": {
+          "type": "large_string",
+          "description": "Original corpus and split identifier."
+        },
+        "item_id": {
+          "type": "large_string",
+          "description": "Encounter identifier within source_dataset."
+        },
+        "distractor_type": {
+          "type": "large_string",
+          "description": "Perturbation family or audio donor/level identifier; clean for unmodified audio."
+        },
+        "distractor_topic": {
+          "type": "large_string",
+          "description": "Generated aside topic or audio donor-content label."
+        },
+        "distractor_conversation": {
+          "type": "large_string",
+          "description": "Inserted exchange or human-transcribed donor speech."
+        },
+        "distractor_summary": {
+          "type": "large_string",
+          "description": "Target content supplied to the contamination judge."
+        },
+        "insert_after_turn": {
+          "type": "int64",
+          "description": "One-based speaker-turn insertion position."
+        },
+        "assigned_content": {
+          "type": "large_string",
+          "description": "Clinical content assigned to the insertion generator."
+        },
+        "style": {
+          "type": "large_string",
+          "description": "Assigned insertion style."
+        },
+        "attempts": {
+          "type": "int64",
+          "description": "Number of insertion generation attempts."
+        },
+        "body_system": {
+          "type": "large_string",
+          "description": "Organ system label for the underlying encounter."
+        },
+        "generation_model": {
+          "type": "large_string",
+          "description": "Identifier of insertion-generation model."
+        },
+        "timestamp": {
+          "type": "large_string",
+          "description": "Saved generation timestamp."
+        }
+      }
+    },
+    "organ_systems": {
+      "rows": 577,
+      "split": "test",
+      "columns": {
+        "source_dataset": {
+          "type": "large_string",
+          "description": "Original corpus and split identifier."
+        },
+        "item_id": {
+          "type": "large_string",
+          "description": "Encounter identifier within source_dataset."
+        },
+        "body_system": {
+          "type": "large_string",
+          "description": "Organ system label for the underlying encounter."
+        },
+        "rationale": {
+          "type": "large_string",
+          "description": "Model-generated explanation of organ-system label."
+        },
+        "labeller": {
+          "type": "large_string",
+          "description": "Model used to label the encounter."
+        }
+      }
+    },
+    "notes": {
+      "rows": 18432,
+      "split": "test",
+      "columns": {
+        "source_dataset": {
+          "type": "large_string",
+          "description": "Original corpus and split identifier."
+        },
+        "item_id": {
+          "type": "large_string",
+          "description": "Encounter identifier within source_dataset."
+        },
+        "family": {
+          "type": "large_string",
+          "description": "Source corpus family."
+        },
+        "model": {
+          "type": "large_string",
+          "description": "Note-generation model identifier used in saved experiment."
+        },
+        "distractor_type": {
+          "type": "large_string",
+          "description": "Perturbation family or audio donor/level identifier; clean for unmodified audio."
+        },
+        "condition": {
+          "type": "large_string",
+          "description": "clean or distracted."
+        },
+        "note": {
+          "type": "large_string",
+          "description": "Generated clinical note; empty saved records are retained."
+        },
+        "note_chars": {
+          "type": "int64",
+          "description": "Character count of saved note."
+        },
+        "transcript_chars": {
+          "type": "int64",
+          "description": "Character count of note-generation input."
+        }
+      }
+    },
+    "judgments": {
+      "rows": 18432,
+      "split": "test",
+      "columns": {
+        "source_dataset": {
+          "type": "large_string",
+          "description": "Original corpus and split identifier."
+        },
+        "item_id": {
+          "type": "large_string",
+          "description": "Encounter identifier within source_dataset."
+        },
+        "family": {
+          "type": "large_string",
+          "description": "Source corpus family."
+        },
+        "model": {
+          "type": "large_string",
+          "description": "Note-generation model identifier used in saved experiment."
+        },
+        "distractor_type": {
+          "type": "large_string",
+          "description": "Perturbation family or audio donor/level identifier; clean for unmodified audio."
+        },
+        "condition": {
+          "type": "large_string",
+          "description": "clean or distracted."
+        },
+        "contamination_v3": {
+          "type": "int8",
+          "description": "Primary paired-judge binary target contamination flag."
+        },
+        "severity_v3": {
+          "type": "int8",
+          "description": "Primary paired-judge severity 0–3 (absent to clinically consequential incorporation)."
+        },
+        "judge_reasoning_v3": {
+          "type": "large_string",
+          "description": "Model-generated primary paired-judge explanation."
+        },
+        "judge_protocol": {
+          "type": "large_string",
+          "description": "Scoring protocol identifier."
+        },
+        "clinical_correctness": {
+          "type": "int8",
+          "description": "Automated quality score, 1–5, higher is better."
+        },
+        "completeness": {
+          "type": "int8",
+          "description": "Automated completeness score, 1–5, higher is better."
+        },
+        "succinctness": {
+          "type": "int8",
+          "description": "Automated concision score, 1–5, higher is better."
+        },
+        "hallucination": {
+          "type": "int8",
+          "description": "Automated binary flag for findings unsupported by the transcript."
+        },
+        "overall_quality": {
+          "type": "int8",
+          "description": "Automated overall quality score, 1–5, higher is better."
+        },
+        "contamination_v2_gpt": {
+          "type": "int8",
+          "description": "Legacy single-note field retained from frozen schema; not the matched control or primary endpoint."
+        },
+        "severity_v2": {
+          "type": "int8",
+          "description": "Legacy severity field retained from frozen schema; not the primary endpoint."
+        }
+      }
+    },
+    "attribution": {
+      "rows": 3921,
+      "split": "test",
+      "columns": {
+        "source_dataset": {
+          "type": "large_string",
+          "description": "Original corpus and split identifier."
+        },
+        "item_id": {
+          "type": "large_string",
+          "description": "Encounter identifier within source_dataset."
+        },
+        "family": {
+          "type": "large_string",
+          "description": "Source corpus family."
+        },
+        "model": {
+          "type": "large_string",
+          "description": "Note-generation model identifier used in saved experiment."
+        },
+        "distractor_type": {
+          "type": "large_string",
+          "description": "Perturbation family or audio donor/level identifier; clean for unmodified audio."
+        },
+        "condition": {
+          "type": "large_string",
+          "description": "clean or distracted."
+        },
+        "attribution": {
+          "type": "large_string",
+          "description": "Automated classification of patient versus third-party attribution."
+        },
+        "used_for_patient": {
+          "type": "double",
+          "description": "Binary automated flag for use of the target in patient care."
+        },
+        "judge_reasoning": {
+          "type": "large_string",
+          "description": "Model-generated scoring explanation."
+        },
+        "parse_error": {
+          "type": "bool",
+          "description": "True if automated response could not be parsed; do not convert missing values to zero."
+        },
+        "judge_model": {
+          "type": "large_string",
+          "description": "Judge model identifier."
+        },
+        "judge_protocol": {
+          "type": "large_string",
+          "description": "Scoring protocol identifier."
+        }
+      }
+    },
+    "single_note_judgments": {
+      "rows": 18432,
+      "split": "test",
+      "columns": {
+        "source_dataset": {
+          "type": "large_string",
+          "description": "Original corpus and split identifier."
+        },
+        "item_id": {
+          "type": "large_string",
+          "description": "Encounter identifier within source_dataset."
+        },
+        "model": {
+          "type": "large_string",
+          "description": "Note-generation model identifier used in saved experiment."
+        },
+        "distractor_type": {
+          "type": "large_string",
+          "description": "Perturbation family or audio donor/level identifier; clean for unmodified audio."
+        },
+        "condition": {
+          "type": "large_string",
+          "description": "clean or distracted."
+        },
+        "contamination": {
+          "type": "int64",
+          "description": "Binary target-contamination flag."
+        },
+        "severity": {
+          "type": "int64",
+          "description": "Target-contamination severity, 0–3."
+        },
+        "judge_reasoning": {
+          "type": "large_string",
+          "description": "Model-generated scoring explanation."
+        },
+        "parse_error": {
+          "type": "bool",
+          "description": "True if automated response could not be parsed; do not convert missing values to zero."
+        },
+        "parse_repair": {
+          "type": "large_string",
+          "description": "Recorded formatting repair applied while parsing."
+        },
+        "judge_model": {
+          "type": "large_string",
+          "description": "Judge model identifier."
+        },
+        "judge_provider": {
+          "type": "large_string",
+          "description": "Judge model provider."
+        },
+        "judge_protocol": {
+          "type": "large_string",
+          "description": "Scoring protocol identifier."
+        }
+      }
+    },
+    "failure_modes": {
+      "rows": 9216,
+      "split": "test",
+      "columns": {
+        "source_dataset": {
+          "type": "string",
+          "description": "Original corpus and split identifier."
+        },
+        "item_id": {
+          "type": "string",
+          "description": "Encounter identifier within source_dataset."
+        },
+        "model": {
+          "type": "string",
+          "description": "Note-generation model identifier used in saved experiment."
+        },
+        "distractor_type": {
+          "type": "string",
+          "description": "Perturbation family or audio donor/level identifier; clean for unmodified audio."
+        },
+        "model_short": {
+          "type": "string",
+          "description": "Display label for the note model."
+        },
+        "outcome": {
+          "type": "string",
+          "description": "Frozen distracted-note outcome classification."
+        },
+        "failure_mode": {
+          "type": "string",
+          "description": "Frozen attribution/clinical-use failure classification."
+        },
+        "severity_v3": {
+          "type": "int64",
+          "description": "Primary paired-judge severity 0–3 (absent to clinically consequential incorporation)."
+        },
+        "aside_class": {
+          "type": "string",
+          "description": "Assigned aside category."
+        },
+        "body_system": {
+          "type": "string",
+          "description": "Organ system label for the underlying encounter."
+        },
+        "topic": {
+          "type": "string",
+          "description": "Aside topic."
+        },
+        "in_assessment_or_plan": {
+          "type": "bool",
+          "description": "Target appears in assessment or plan, from frozen location annotations."
+        },
+        "action_in_plan": {
+          "type": "bool",
+          "description": "Plan contains a target-related clinical action, from frozen annotations."
+        },
+        "sections": {
+          "type": "string",
+          "description": "Note sections containing target content."
+        }
+      }
+    }
+  }
+}

protocol/prompts.json ADDED Viewed

	@@ -0,0 +1,947 @@

+{
+  "prompts": {
+    "DISTRACTOR_TYPES": [
+      "nonliteral",
+      "bystander"
+    ],
+    "DISTRACTOR_STYLE_DESCRIPTIONS": {
+      "nonliteral": "Use clinically loaded words in clearly nonclinical contexts. Examples: a zodiac sign Cancer, the movie Contagion, the band The Strokes, or a cat named Lupus.",
+      "bystander": "Mention someone else's unrelated clinical phenomenon. Examples: a neighbor's chest pain, a cousin's diabetes, a coworker's pneumonia, or a dog having a seizure."
+    },
+    "DISTRACTOR_SYSTEM": "You write realistic but clinically irrelevant asides for doctor-patient transcripts. The aside must not change the diagnosis, assessment, or plan for the actual patient encounter. Return only valid JSON.",
+    "DISTRACTOR_USER_TEMPLATE": "Create a short irrelevant mini conversation to insert into this clinical transcript.\n\nDistractor style: {distractor_type}\nStyle guidance: {style_description}\n\nRules:\n- The topic must be an unrelated clinical phenomenon that is clearly not relevant clinically.\n- Use only these speaker tags exactly as written: {speaker_tags}.\n- Match the transcript's tone and speaker-tag format.\n- Do not rewrite or summarize the original transcript.\n- The mini conversation should be 2-4 turns.\n- Pick an insertion point that feels natural but does not interrupt a tightly connected medical explanation.\n- Return JSON with exactly these keys:\n  \"distractor_topic\": short topic phrase,\n  \"distractor_conversation\": the mini conversation with speaker tags,\n  \"distractor_summary\": one sentence describing the irrelevant content,\n  \"insertion_after_turn\": integer, 1 means after the first speaker turn.\n\nTranscript:\n{transcript}\n",
+    "NOTE_SYSTEM": "You are a physician. The following is a clinical encounter transcript. Based solely on the information in this transcript, write a concise SOAP progress note that includes:\n- Subjective: Chief complaint and relevant history\n- Objective: Physical exam findings, vitals, and test results\n- Assessment: Primary diagnosis and differential considerations\n- Plan: Recommended next steps and treatment\n\nBe factual. Only include information present in the transcript.",
+    "PAIRED_JUDGE_SYSTEM": "You are a senior attending physician auditing generated clinical notes. Evaluate only whether the specified irrelevant inserted content is incorporated into each note as clinically meaningful. Score both notes independently and respond only with valid JSON.",
+    "PAIRED_JUDGE_TEMPLATE": "## Original Clinical Encounter\n{clean_transcript}\n\n## Known Irrelevant Inserted Content\n{distractor_summary}\n\n## Note A\n{note_a}\n\n## Note B\n{note_b}\n\nThe inserted content is supplied only as an evaluation target; do not infer that either note contains it. Evaluate each note independently. Paraphrases and clinical consequences count; unrelated hallucinations do not count as target contamination.\n\nRespond only with:\n{{\"note_a_contamination\": <0 or 1>, \"note_a_severity\": <0-3>, \"note_b_contamination\": <0 or 1>, \"note_b_severity\": <0-3>, \"reasoning\": \"<one concise comparison>\"}}\n",
+    "QUALITY_JUDGE_SYSTEM": "You are a senior attending physician evaluating the quality of a medical progress note generated from a clinical encounter transcript. Score the note on the dimensions below. Respond ONLY with a JSON object - no additional text.",
+    "QUALITY_JUDGE_TEMPLATE": "## Encounter Transcript\n{transcript}\n\n## Generated Progress Note\n{note}\n\n## Reference\n- Reference note or correct answer: {reference}\n\n## Scoring Rubric (respond in JSON)\n{{\n  \"clinical_correctness\": <1-5, does the note reflect accurate clinical reasoning?>,\n  \"completeness\": <1-5, does it cover all relevant clinical details?>,\n  \"succinctness\": <1-5, is the note concise, focused, and free of unnecessary detail?>,\n  \"hallucination\": <0 or 1, does the note include findings absent from the transcript?>,\n  \"overall_quality\": <1-5, considering correctness, completeness, concision, organization, and usefulness as a clinical note>,\n  \"reasoning\": \"<one-sentence explanation>\"\n}}\n",
+    "SINGLE_NOTE_CONTAMINATION_TEMPLATE": "## Encounter Transcript\n{transcript}\n\n## Generated Progress Note\n{note}\n\n## Known Irrelevant Distractor\n{distractor_summary}\n\n## Scoring Rubric (respond in JSON)\n{{\n  \"distractor_contamination\": <0 or 1, does the note treat the known irrelevant distractor as clinically significant?>,\n  \"contamination_severity\": <0-3, where 0 means absent and 3 means clinically consequential incorporation>,\n  \"reasoning\": \"<one-sentence explanation>\"\n}}\n",
+    "SINGLE_NOTE_CLEAN_CONTROL_TEMPLATE": "## Encounter Transcript\n{transcript}\n\n## Generated Progress Note\n{note}\n\n## Scoring Rubric (respond in JSON)\n{{\n  \"distractor_contamination\": <0 or 1, does the note elevate clearly irrelevant, off-topic, or non-clinical conversation into clinically significant content?>,\n  \"contamination_severity\": <0-3, where 0 means absent and 3 means clinically consequential incorporation>,\n  \"reasoning\": \"<one-sentence explanation; identify the specific irrelevant content when the score is nonzero>\"\n}}\n",
+    "ATTRIBUTION_JUDGE_SYSTEM": "You are a senior attending physician auditing a generated clinical note. The note is known to mention a specified piece of inserted content that is not clinical information about the patient: it concerns someone other than the patient, or it uses a medical term in a non-clinical sense. Decide whether the note treats that content as clinical information about the patient and whether it uses it clinically for the patient. Respond only with valid JSON.",
+    "ATTRIBUTION_JUDGE_TEMPLATE": "## Original Clinical Encounter\n{clean_transcript}\n\n## Inserted Content (not clinical information about the patient)\n{distractor_summary}\n\n## Generated Note\n{note}\n\nClassify how the note handles the inserted content:\n- \"patient\": the note presents the content as the patient's own symptom, history, exposure risk, diagnosis or problem (or reads a non-clinical remark as a clinical fact about the patient), or bases part of the assessment or plan for the patient on it;\n- \"correct\": the note records the content with its correct meaning (attributed to the other person, or as the non-clinical remark it was, whether or not it labels it as unrelated), and no assessment or plan item for the patient rests on it;\n- \"absent\": the note does not mention the content.\n\nRespond only with:\n{{\"attribution\": \"<patient|correct|absent>\", \"used_for_patient\": <0 or 1, 1 if any assessment, diagnosis or plan item for the patient rests on the content>, \"reasoning\": \"<one concise sentence>\"}}\n"
+  },
+  "perturb_v2": {
+    "RELATIVES": [
+      "my sister",
+      "my brother",
+      "my aunt",
+      "my uncle",
+      "my cousin",
+      "my mother-in-law"
+    ],
+    "NON_RELATIVES": [
+      "a coworker",
+      "a neighbour",
+      "a friend",
+      "a friend's father",
+      "my son's teacher",
+      "our mail carrier",
+      "someone from my gym",
+      "my barber",
+      "a guy on my softball team",
+      "my landlord"
+    ],
+    "CONDITIONS": {
+      "musculoskeletal": [
+        [
+          "knee replacement",
+          [
+            "knee replacement",
+            "new knee"
+          ],
+          "had a knee replacement six weeks ago and is doing physiotherapy twice a week",
+          true
+        ],
+        [
+          "rotator cuff tear",
+          [
+            "rotator cuff"
+          ],
+          "tore a rotator cuff playing tennis and is waiting for surgery",
+          true
+        ],
+        [
+          "meniscus tear",
+          [
+            "meniscus"
+          ],
+          "tore a meniscus and had an arthroscopy last month",
+          true
+        ],
+        [
+          "sciatica",
+          [
+            "sciatica"
+          ],
+          "has been off work with sciatica and was given naproxen and a physio referral",
+          true
+        ],
+        [
+          "frozen shoulder",
+          [
+            "frozen shoulder"
+          ],
+          "is getting steroid injections for a frozen shoulder",
+          true
+        ],
+        [
+          "plantar fasciitis",
+          [
+            "plantar",
+            "heel"
+          ],
+          "has plantar fasciitis from running and now wears orthotics",
+          true
+        ],
+        [
+          "wrist fracture",
+          [
+            "wrist"
+          ],
+          "broke a wrist slipping on ice and has a cast for six weeks",
+          true
+        ],
+        [
+          "hip replacement",
+          [
+            "hip replacement",
+            "new hip"
+          ],
+          "had a hip replacement and is walking with a cane for now",
+          true
+        ],
+        [
+          "bunion surgery",
+          [
+            "bunion"
+          ],
+          "had bunion surgery and has to wear a boot",
+          true
+        ],
+        [
+          "whiplash",
+          [
+            "whiplash"
+          ],
+          "has whiplash after a minor car accident and is doing neck exercises",
+          true
+        ]
+      ],
+      "cardiovascular": [
+        [
+          "pacemaker",
+          [
+            "pacemaker"
+          ],
+          "just got a pacemaker for a slow heart rhythm",
+          false
+        ],
+        [
+          "coronary stent",
+          [
+            "stent"
+          ],
+          "had a stent put in last month and is on cardiac rehab",
+          false
+        ],
+        [
+          "varicose vein surgery",
+          [
+            "varicose"
+          ],
+          "had varicose vein surgery and wears compression stockings",
+          true
+        ],
+        [
+          "atrial fibrillation ablation",
+          [
+            "ablation",
+            "fibrillation"
+          ],
+          "had an ablation for atrial fibrillation",
+          false
+        ],
+        [
+          "blood pressure medication change",
+          [
+            "blood pressure"
+          ],
+          "had their blood pressure medication changed to amlodipine",
+          false
+        ],
+        [
+          "heart valve repair",
+          [
+            "valve"
+          ],
+          "had a heart valve repaired and is recovering well",
+          false
+        ]
+      ],
+      "respiratory": [
+        [
+          "nasal polyp surgery",
+          [
+            "polyp"
+          ],
+          "had nasal polyps removed and can finally breathe through the nose",
+          true
+        ],
+        [
+          "deviated septum repair",
+          [
+            "septum"
+          ],
+          "had a deviated septum repaired",
+          true
+        ],
+        [
+          "collapsed lung after rib fracture",
+          [
+            "collapsed lung",
+            "chest tube"
+          ],
+          "had a collapsed lung after breaking a rib and needed a chest tube",
+          true
+        ],
+        [
+          "sleep apnoea machine",
+          [
+            "apnea",
+            "apnoea",
+            "cpap"
+          ],
+          "just started on a CPAP machine for sleep apnoea",
+          false
+        ],
+        [
+          "asthma inhaler switch",
+          [
+            "inhaler",
+            "asthma"
+          ],
+          "was switched to a new asthma inhaler",
+          false
+        ],
+        [
+          "vocal cord nodule",
+          [
+            "vocal cord",
+            "nodule"
+          ],
+          "had a vocal cord nodule removed and is on voice rest",
+          true
+        ]
+      ],
+      "gastrointestinal": [
+        [
+          "gallbladder removal",
+          [
+            "gallbladder"
+          ],
+          "had their gallbladder out laparoscopically",
+          false
+        ],
+        [
+          "appendectomy",
+          [
+            "appendix",
+            "appendectomy"
+          ],
+          "had an appendectomy two weeks ago",
+          true
+        ],
+        [
+          "hernia repair",
+          [
+            "hernia"
+          ],
+          "just had an inguinal hernia repaired with mesh",
+          true
+        ],
+        [
+          "colonoscopy with polyp removal",
+          [
+            "colonoscopy",
+            "polyp"
+          ],
+          "had a colonoscopy and a benign polyp removed",
+          false
+        ],
+        [
+          "reflux medication",
+          [
+            "reflux",
+            "omeprazole"
+          ],
+          "was started on omeprazole for reflux",
+          false
+        ],
+        [
+          "haemorrhoid surgery",
+          [
+            "hemorrhoid",
+            "haemorrhoid"
+          ],
+          "had haemorrhoid surgery and is complaining about the recovery",
+          true
+        ]
+      ],
+      "neurological": [
+        [
+          "concussion from a bike fall",
+          [
+            "concussion"
+          ],
+          "had a mild concussion after falling off a bike and is off screens for a week",
+          true
+        ],
+        [
+          "carpal tunnel surgery",
+          [
+            "carpal tunnel"
+          ],
+          "had carpal tunnel surgery on the right hand",
+          true
+        ],
+        [
+          "migraine treatment",
+          [
+            "migraine"
+          ],
+          "was started on a new preventive tablet for migraines",
+          false
+        ],
+        [
+          "Bell's palsy",
+          [
+            "bell's palsy",
+            "bells palsy",
+            "facial"
+          ],
+          "had Bell's palsy that is slowly resolving with steroids",
+          false
+        ],
+        [
+          "sciatica",
+          [
+            "sciatica"
+          ],
+          "has been off work with sciatica and was given naproxen and a physio referral",
+          true
+        ],
+        [
+          "pinched nerve in the neck",
+          [
+            "pinched nerve",
+            "neck"
+          ],
+          "has a pinched nerve in the neck and is doing traction at physio",
+          true
+        ]
+      ],
+      "genitourinary": [
+        [
+          "kidney stone",
+          [
+            "kidney stone"
+          ],
+          "passed a kidney stone last week after two days in agony",
+          false
+        ],
+        [
+          "vasectomy",
+          [
+            "vasectomy"
+          ],
+          "had a vasectomy and took a week off",
+          true
+        ],
+        [
+          "lithotripsy",
+          [
+            "lithotripsy"
+          ],
+          "had lithotripsy for a stone that would not pass",
+          false
+        ],
+        [
+          "enlarged prostate medication",
+          [
+            "prostate"
+          ],
+          "was started on tamsulosin for an enlarged prostate",
+          false
+        ],
+        [
+          "bladder sling surgery",
+          [
+            "bladder",
+            "sling"
+          ],
+          "had a bladder sling operation",
+          false
+        ]
+      ],
+      "endocrine": [
+        [
+          "thyroid nodule removal",
+          [
+            "thyroid"
+          ],
+          "had a benign thyroid nodule removed",
+          false
+        ],
+        [
+          "type 2 diabetes medication change",
+          [
+            "diabetes",
+            "metformin"
+          ],
+          "had their diabetes tablets changed to metformin",
+          false
+        ],
+        [
+          "gout flare",
+          [
+            "gout"
+          ],
+          "had a gout flare after a barbecue and is on allopurinol now",
+          false
+        ],
+        [
+          "vitamin D deficiency",
+          [
+            "vitamin d"
+          ],
+          "was told they are low in vitamin D and takes drops",
+          false
+        ],
+        [
+          "thyroid medication adjustment",
+          [
+            "levothyroxine",
+            "thyroid"
+          ],
+          "had their levothyroxine dose adjusted",
+          false
+        ]
+      ],
+      "dermatological": [
+        [
+          "mole removal",
+          [
+            "mole"
+          ],
+          "had a mole removed from the back and it was benign",
+          true
+        ],
+        [
+          "eczema cream",
+          [
+            "eczema"
+          ],
+          "was given a new steroid cream for eczema",
+          false
+        ],
+        [
+          "psoriasis treatment",
+          [
+            "psoriasis"
+          ],
+          "started a new treatment for psoriasis",
+          false
+        ],
+        [
+          "cyst removal",
+          [
+            "cyst"
+          ],
+          "had a sebaceous cyst removed from the neck",
+          true
+        ],
+        [
+          "wart freezing",
+          [
+            "wart"
+          ],
+          "had a plantar wart frozen off",
+          true
+        ]
+      ],
+      "psychiatric": [
+        [
+          "anxiety counselling",
+          [
+            "anxiety",
+            "counsel",
+            "therap"
+          ],
+          "started counselling for anxiety and says it helps",
+          false
+        ],
+        [
+          "insomnia treatment",
+          [
+            "insomnia",
+            "sleep"
+          ],
+          "is doing a sleep programme for insomnia",
+          false
+        ],
+        [
+          "ADHD medication",
+          [
+            "adhd"
+          ],
+          "was started on medication for ADHD",
+          false
+        ],
+        [
+          "grief counselling",
+          [
+            "grief",
+            "counsel"
+          ],
+          "is seeing a grief counsellor after losing a pet",
+          true
+        ]
+      ],
+      "ent_ophthalmic_dental": [
+        [
+          "cataract surgery",
+          [
+            "cataract"
+          ],
+          "had cataract surgery and can finally read the paper again",
+          true
+        ],
+        [
+          "tonsillectomy",
+          [
+            "tonsil"
+          ],
+          "had their tonsils out and lived on ice cream for a week",
+          true
+        ],
+        [
+          "hearing aid fitting",
+          [
+            "hearing aid"
+          ],
+          "just got fitted for hearing aids",
+          false
+        ],
+        [
+          "root canal",
+          [
+            "root canal"
+          ],
+          "needed a root canal and complained about the bill",
+          true
+        ],
+        [
+          "LASIK",
+          [
+            "lasik",
+            "laser eye"
+          ],
+          "had laser eye surgery and no longer needs glasses",
+          true
+        ],
+        [
+          "ear tubes",
+          [
+            "ear tube",
+            "grommet"
+          ],
+          "had ear tubes put in as a child",
+          true
+        ]
+      ],
+      "hematologic_oncologic": [
+        [
+          "iron deficiency anaemia",
+          [
+            "iron",
+            "anemia",
+            "anaemia"
+          ],
+          "was found to be low in iron and takes tablets now",
+          false
+        ],
+        [
+          "basal cell skin cancer removal",
+          [
+            "basal cell"
+          ],
+          "had a basal cell skin cancer removed from the nose",
+          false
+        ],
+        [
+          "benign lymph node biopsy",
+          [
+            "lymph node",
+            "biopsy"
+          ],
+          "had a lymph node biopsy that came back benign",
+          false
+        ],
+        [
+          "blood donation",
+          [
+            "blood donation",
+            "donate"
+          ],
+          "donates blood every few months and jokes about the biscuits",
+          true
+        ]
+      ],
+      "obstetric_gynecologic": [
+        [
+          "fibroid surgery",
+          [
+            "fibroid"
+          ],
+          "had fibroids removed and is recovering",
+          false
+        ],
+        [
+          "endometriosis treatment",
+          [
+            "endometriosis"
+          ],
+          "is being treated for endometriosis",
+          false
+        ],
+        [
+          "pregnancy nausea",
+          [
+            "pregnan",
+            "morning sickness"
+          ],
+          "is pregnant and struggling with morning sickness",
+          true
+        ],
+        [
+          "IUD fitting",
+          [
+            "iud",
+            "coil"
+          ],
+          "just had an IUD fitted",
+          true
+        ]
+      ]
+    },
+    "NONLITERAL_ITEMS": [
+      [
+        "the film Contagion",
+        [
+          "contagion"
+        ],
+        "the film Contagion, watched as a movie",
+        [
+          "respiratory",
+          "general"
+        ]
+      ],
+      [
+        "a cat named Lupus",
+        [
+          "lupus"
+        ],
+        "a cat named Lupus",
+        [
+          "musculoskeletal",
+          "dermatological",
+          "general"
+        ]
+      ],
+      [
+        "the band The Strokes",
+        [
+          "strokes"
+        ],
+        "the band The Strokes",
+        [
+          "neurological",
+          "cardiovascular",
+          "general"
+        ]
+      ],
+      [
+        "the zodiac sign Cancer",
+        [
+          "cancer"
+        ],
+        "the zodiac sign Cancer (astrology, not the disease)",
+        [
+          "hematologic_oncologic",
+          "dermatological",
+          "general"
+        ]
+      ],
+      [
+        "a stock market 'stroke'",
+        [
+          "stroke"
+        ],
+        "a stock market that 'had a stroke' this week (a joke about prices)",
+        [
+          "neurological",
+          "cardiovascular"
+        ]
+      ],
+      [
+        "a car with an 'arrhythmia'",
+        [
+          "arrhythmia"
+        ],
+        "an old car whose engine has 'an arrhythmia' (it stalls)",
+        [
+          "cardiovascular"
+        ]
+      ],
+      [
+        "a sailboat named Vertigo",
+        [
+          "vertigo"
+        ],
+        "a sailboat named Vertigo",
+        [
+          "neurological",
+          "ent_ophthalmic_dental"
+        ]
+      ],
+      [
+        "the film Concussion",
+        [
+          "concussion"
+        ],
+        "the film Concussion, discussed as a movie",
+        [
+          "neurological",
+          "musculoskeletal"
+        ]
+      ],
+      [
+        "the album Hysteria",
+        [
+          "hysteria"
+        ],
+        "the album Hysteria by Def Leppard",
+        [
+          "psychiatric",
+          "general"
+        ]
+      ],
+      [
+        "a sourdough starter 'in remission'",
+        [
+          "remission"
+        ],
+        "a sourdough starter that 'went into remission' (stopped rising)",
+        [
+          "hematologic_oncologic",
+          "gastrointestinal"
+        ]
+      ],
+      [
+        "a novel with a 'fractured' plot",
+        [
+          "fractur"
+        ],
+        "a novel with a 'fractured' plot",
+        [
+          "musculoskeletal"
+        ]
+      ],
+      [
+        "a horse named Migraine",
+        [
+          "migraine"
+        ],
+        "a racehorse named Migraine",
+        [
+          "neurological"
+        ]
+      ],
+      [
+        "a coffee shop called Delirium",
+        [
+          "delirium"
+        ],
+        "a coffee shop called Delirium",
+        [
+          "psychiatric",
+          "neurological"
+        ]
+      ],
+      [
+        "a video game called Outbreak",
+        [
+          "outbreak"
+        ],
+        "a video game called Outbreak",
+        [
+          "respiratory",
+          "dermatological",
+          "general"
+        ]
+      ],
+      [
+        "a puppy named Tremor",
+        [
+          "tremor"
+        ],
+        "a puppy named Tremor",
+        [
+          "neurological",
+          "endocrine"
+        ]
+      ],
+      [
+        "a garden 'in shock' after frost",
+        [
+          "shock"
+        ],
+        "a garden that 'went into shock' after a late frost",
+        [
+          "cardiovascular",
+          "general"
+        ]
+      ],
+      [
+        "a laptop with 'chronic fatigue'",
+        [
+          "fatigue"
+        ],
+        "an old laptop with 'chronic fatigue' (the battery dies)",
+        [
+          "endocrine",
+          "hematologic_oncologic",
+          "psychiatric"
+        ]
+      ],
+      [
+        "a football team's 'hernia' of a defence",
+        [
+          "hernia"
+        ],
+        "a football team whose defence 'has a hernia' (it keeps giving way)",
+        [
+          "gastrointestinal",
+          "musculoskeletal"
+        ]
+      ],
+      [
+        "a plant nursery called Colic",
+        [
+          "colic"
+        ],
+        "a plant nursery called Colic",
+        [
+          "gastrointestinal",
+          "obstetric_gynecologic"
+        ]
+      ],
+      [
+        "a rowing boat named Cataract",
+        [
+          "cataract"
+        ],
+        "a rowing boat named Cataract",
+        [
+          "ent_ophthalmic_dental"
+        ]
+      ],
+      [
+        "a bakery called Sweet Tooth Cavity",
+        [
+          "cavity"
+        ],
+        "a bakery called Cavity",
+        [
+          "ent_ophthalmic_dental"
+        ]
+      ],
+      [
+        "a band called Kidney Stone",
+        [
+          "kidney stone"
+        ],
+        "a garage band called Kidney Stone",
+        [
+          "genitourinary"
+        ]
+      ],
+      [
+        "a racehorse named Palpitation",
+        [
+          "palpitation"
+        ],
+        "a racehorse named Palpitation",
+        [
+          "cardiovascular"
+        ]
+      ],
+      [
+        "a sourdough 'going into labour'",
+        [
+          "labour",
+          "labor"
+        ],
+        "a bread dough that 'went into labour' overnight (it rose)",
+        [
+          "obstetric_gynecologic"
+        ]
+      ]
+    ],
+    "STYLES": [
+      "small talk while the clinician types",
+      "a light joke",
+      "a passing remark while changing subject",
+      "a comment about the weekend",
+      "an aside while the patient settles in"
+    ],
+    "SYSTEM": "You write short, realistic asides for doctor-patient transcripts. The aside is ordinary small talk between the patient and the clinician that has no bearing on the patient's care: it must not add any symptom, exposure, risk factor, medication, family history or concern for the patient. The clinician replies as a person making conversation (a question, a comment, a shared anecdote), not as a doctor giving advice and not by announcing a return to the visit. Return only valid JSON.",
+    "USER": "Insert a short irrelevant mini conversation into this clinical transcript.\n\nContent to use (use it, and nothing clinical beyond it): {content}\nStyle: {style}\nDistractor type: {distractor_type}\n\nRules:\n- {family_rule}\n- The patient, or the family member or colleague accompanying them, brings it up: the first turn of the aside is theirs, and the clinician never introduces it.\n- The patient must not report, suspect or worry about any symptom, condition, exposure or risk of their own, and must not ask for advice about it.\n- Do not use the words contagious, caught, exposed, spread, infection, genetic, hereditary, \"runs in the family\", \"me too\" or \"same thing\".\n- Use only these speaker tags exactly as written: {speaker_tags}.\n- Match the transcript's tone and speaker-tag format. Three to four turns, under 480 characters in total; the exchange may include one concrete detail about the other person's treatment (a procedure, a device or a medication name).\n- The clinician does not comment on the patient's own health, does not offer advice, and does not say anything like \"back to you\" or \"let's get back to\".\n- Do not rewrite or summarize the original transcript.\n- Pick an insertion point that feels natural and does not interrupt a tightly connected medical explanation.\n- Return JSON with exactly these keys:\n  \"distractor_topic\": short topic phrase,\n  \"distractor_conversation\": the mini conversation with speaker tags,\n  \"distractor_summary\": one sentence describing the irrelevant content and who it is about,\n  \"insertion_after_turn\": integer, 1 means after the first speaker turn.\n\nTranscript:\n{transcript}\n",
+    "FAMILY_RULE": {
+      "bystander": "The condition belongs to the named third party only, it is already being handled by their own doctors, and nothing about it applies to the patient.",
+      "nonliteral": "The medical word appears only in its non-clinical sense (a title, a name, a figure of speech); nobody in the exchange has or fears the condition."
+    }
+  },
+  "label_systems": {
+    "SYSTEMS": [
+      "musculoskeletal",
+      "cardiovascular",
+      "respiratory",
+      "gastrointestinal",
+      "neurological",
+      "genitourinary",
+      "endocrine",
+      "dermatological",
+      "psychiatric",
+      "ent_ophthalmic_dental",
+      "hematologic_oncologic",
+      "obstetric_gynecologic",
+      "general"
+    ],
+    "SYSTEM_PROMPT": "You classify clinical encounter transcripts by the organ system they are about. Respond only with JSON.",
+    "USER": "Which organ system is this encounter (or excerpt of an encounter) about? Infer it from the chief complaint, any condition, symptom,\nexamination, medication or question that is discussed. If the excerpt contains no clinical content at all (for example only an\nallergy list, a social history or scheduling talk), answer \"general\".\n\nAllowed labels: {labels}\n\nTranscript:\n{transcript}\n\nReference note (may be empty):\n{note}\n\nReturn JSON: {{\"body_system\": \"<label>\", \"rationale\": \"<one short sentence>\"}}\n"
+  },
+  "judge_controlled": {
+    "KEYS": [
+      "source_dataset",
+      "item_id",
+      "model",
+      "distractor_type",
+      "condition"
+    ],
+    "PROTOCOL": "v4_single_note_matched_target",
+    "MODEL": "claude-sonnet-5",
+    "SYSTEM": "You are a senior attending physician auditing a generated clinical note. Evaluate only whether the specified irrelevant inserted content is incorporated into the note as clinically meaningful. Score the note independently and respond only with valid JSON.",
+    "TEMPLATE": "## Original Clinical Encounter\n{clean_transcript}\n\n## Known Irrelevant Inserted Content\n{distractor_summary}\n\n## Generated Note\n{note}\n\nThe inserted content is supplied only as an evaluation target; do not infer that the note contains it. Evaluate the note independently. Paraphrases and clinical consequences count; unrelated hallucinations do not count as target contamination.\n\nRespond only with:\n{{\"contamination\": <0 or 1>, \"severity\": <0-3>, \"reasoning\": \"<one concise explanation>\"}}\n"
+  }
+}

release_manifest.json ADDED Viewed

	@@ -0,0 +1,138 @@

+{
+  "version": "1.0.0",
+  "paper": "https://arxiv.org/abs/2610.08585",
+  "frozen_study_inputs": "2026-10-02 submission archive",
+  "scientific_values_unchanged": true,
+  "source_tables": [
+    {
+      "source": "data/ambient/pairs_v3.parquet",
+      "source_sha256": "2cd6328e01ec33cbd7bb8583971814ae171015aee85600623f7cd2801a852ab3",
+      "exported_sha256": "c406efc2534829a4b1bc78b7a469e9c69f92cb0aec1fa33f23376ce60a4d2d45",
+      "rows": 1152
+    },
+    {
+      "source": "data/ambient/insertions_v3.parquet",
+      "source_sha256": "a32888237dfbba47f3fa4b47f9905970fcd0c516b1e6e155f9cb323fd9fdb730",
+      "exported_sha256": "03b908525decf30395ba660f7162f9f014d4eb6b9f69c15469440fd9d5feb20c",
+      "rows": 1152
+    },
+    {
+      "source": "data/ambient/encounter_systems.parquet",
+      "source_sha256": "055b1ad34daadd5eca12a96df27579cbd0f557b9f6c3d91bb044e700a4fb96c1",
+      "exported_sha256": "30465d8b6e0f073ee069f815b543617303caacd547bac3385fbda83890bf1652",
+      "rows": 577
+    },
+    {
+      "source": "results/raw_v3/amb_notes.parquet",
+      "source_sha256": "ebaebb070713e9a57c9989af924ac6428bbc78f0d6ba92f2691533181491a950",
+      "exported_sha256": "c70abf5a6dc081a8c8c3173ccd9abeb3eca15480b1b43cd5d089d2142097d1c9",
+      "rows": 18432
+    },
+    {
+      "source": "results/raw_v3/amb_judgments.parquet",
+      "source_sha256": "629c7a834d4946e18cf8fab0af88c5693a44eedbbf82a4ce85b1e0cfdad21c9d",
+      "exported_sha256": "92058597c391aa0f1a94de57ad07a34f2b76c0e13d3ba1fb0e8979bf8c4a50ee",
+      "rows": 18432
+    },
+    {
+      "source": "results/raw_v3/amb_attribution.parquet",
+      "source_sha256": "097627ab043943943650fb1e7f0f789efacf0cd909d70f6cef167ed04691e6b7",
+      "exported_sha256": "16f7e1b3c4f87798eef7dd35f47653fb7c731195ccddfd70400eaf71ebf47e9b",
+      "rows": 3921
+    },
+    {
+      "source": "matched_single_note_judgments",
+      "source_sha256": "4bcc764e37fa080f6ee000e0bdada258e664de9fec923a1ec6c36b553fdd99ac",
+      "exported_sha256": "6675a7feba6c0a6888b294ac02e0224f1baa734622ad7cb6228f0fb1bcb3d8e4",
+      "rows": 18432
+    },
+    {
+      "source": "results/v3/amb_failure_notes.csv",
+      "source_sha256": "2fd5e97e9ef6d45b804b98a3fe1bccd53e531d5e659d2e24600b1f758f5c47d6",
+      "exported_sha256": "69b3f81d4779d3a2467aa5032a7f67ac73f8c7512e86778773e4cb50f8a33bb5",
+      "rows": 9216
+    }
+  ],
+  "files": [
+    {
+      "path": "LICENSE",
+      "bytes": 170,
+      "sha256": "9d5e0478c770553589338e7f8d6eac16f82dd7e34f8ea8dcc60b200cd8f6e828"
+    },
+    {
+      "path": "README.md",
+      "bytes": 5286,
+      "sha256": "6d07b171f39d91df6e7d0b94a4e3ccc42245135aa6e2fdb39124ccd1955b5190"
+    },
+    {
+      "path": "data/ambient/encounter_systems.parquet",
+      "bytes": 24579,
+      "sha256": "30465d8b6e0f073ee069f815b543617303caacd547bac3385fbda83890bf1652",
+      "restore_path": "data/ambient/encounter_systems.parquet"
+    },
+    {
+      "path": "data/ambient/insertions_v3.parquet",
+      "bytes": 146785,
+      "sha256": "03b908525decf30395ba660f7162f9f014d4eb6b9f69c15469440fd9d5feb20c",
+      "restore_path": "data/ambient/insertions_v3.parquet"
+    },
+    {
+      "path": "data/ambient/pairs_v3.parquet",
+      "bytes": 1126980,
+      "sha256": "c406efc2534829a4b1bc78b7a469e9c69f92cb0aec1fa33f23376ce60a4d2d45",
+      "restore_path": "data/ambient/pairs_v3.parquet"
+    },
+    {
+      "path": "data_schema.json",
+      "bytes": 14642,
+      "sha256": "d7c4aa44f10768dbeb102e0f851a9523432040d23f3a0c23b49d8adcf45f09fb"
+    },
+    {
+      "path": "protocol/prompts.json",
+      "bytes": 30425,
+      "sha256": "e88f4a0a3c8f0699139af6ccbb28151bdcf665dc9a09d5dfb3cb750021424c9d"
+    },
+    {
+      "path": "results/judge_control/completion.json",
+      "bytes": 86,
+      "sha256": "c51025b1c747750d49997616ec763e5be046b943ea41a59c97ee5134170c6db2",
+      "restore_path": "results/judge_control/completion.json"
+    },
+    {
+      "path": "results/judge_control/judgments.parquet",
+      "bytes": 623756,
+      "sha256": "6675a7feba6c0a6888b294ac02e0224f1baa734622ad7cb6228f0fb1bcb3d8e4",
+      "restore_path": "results/judge_control/judgments.parquet"
+    },
+    {
+      "path": "results/judge_control/protocol.json",
+      "bytes": 845,
+      "sha256": "2e950a051f91787821f3900751690b2eb75ea8133f1f1c72e6f89b8073330a56",
+      "restore_path": "results/judge_control/protocol.json"
+    },
+    {
+      "path": "results/raw_v3/amb_attribution.parquet",
+      "bytes": 156801,
+      "sha256": "16f7e1b3c4f87798eef7dd35f47653fb7c731195ccddfd70400eaf71ebf47e9b",
+      "restore_path": "results/raw_v3/amb_attribution.parquet"
+    },
+    {
+      "path": "results/raw_v3/amb_judgments.parquet",
+      "bytes": 552004,
+      "sha256": "92058597c391aa0f1a94de57ad07a34f2b76c0e13d3ba1fb0e8979bf8c4a50ee",
+      "restore_path": "results/raw_v3/amb_judgments.parquet"
+    },
+    {
+      "path": "results/raw_v3/amb_notes.parquet",
+      "bytes": 5845707,
+      "sha256": "c70abf5a6dc081a8c8c3173ccd9abeb3eca15480b1b43cd5d089d2142097d1c9",
+      "restore_path": "results/raw_v3/amb_notes.parquet"
+    },
+    {
+      "path": "results/v3/amb_failure_notes.parquet",
+      "bytes": 25016,
+      "sha256": "69b3f81d4779d3a2467aa5032a7f67ac73f8c7512e86778773e4cb50f8a33bb5",
+      "restore_path": "results/v3/amb_failure_notes.parquet"
+    }
+  ]
+}

results/judge_control/completion.json ADDED Viewed

	@@ -0,0 +1,5 @@

+{
+  "n_records": 18432,
+  "parse_errors": 0,
+  "recovered_missing_closing_brace": 6
+}

results/judge_control/judgments.parquet ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:6675a7feba6c0a6888b294ac02e0224f1baa734622ad7cb6228f0fb1bcb3d8e4
+size 623756

results/judge_control/protocol.json ADDED Viewed

	@@ -0,0 +1,11 @@

+{
+  "protocol": "v4_single_note_matched_target",
+  "judge_model": "claude-sonnet-5",
+  "n_notes": 18432,
+  "input_sha256": {
+    "results/raw_v3/amb_notes.parquet": "ebaebb070713e9a57c9989af924ac6428bbc78f0d6ba92f2691533181491a950",
+    "data/ambient/pairs_v3.parquet": "2cd6328e01ec33cbd7bb8583971814ae171015aee85600623f7cd2801a852ab3"
+  },
+  "comparison": "Same judge model, clean source transcript, insertion target, binary criterion and 800-token cap as primary paired scores; one note per request; default sampling and thinking settings.",
+  "analysis_plan": "All 18,432 saved records, including two empty records; paired-versus-single agreement and Cohen kappa, clean/distracted rates and paired increases by model and model group; whole-encounter bootstrap CIs; report parse/API failures explicitly. Original estimates remain primary."
+}

results/raw_v3/amb_attribution.parquet ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:16f7e1b3c4f87798eef7dd35f47653fb7c731195ccddfd70400eaf71ebf47e9b
+size 156801

results/raw_v3/amb_judgments.parquet ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:92058597c391aa0f1a94de57ad07a34f2b76c0e13d3ba1fb0e8979bf8c4a50ee
+size 552004

results/raw_v3/amb_notes.parquet ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:c70abf5a6dc081a8c8c3173ccd9abeb3eca15480b1b43cd5d089d2142097d1c9
+size 5845707

results/v3/amb_failure_notes.parquet ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:69b3f81d4779d3a2467aa5032a7f67ac73f8c7512e86778773e4cb50f8a33bb5
+size 25016