Move public docs under docs folder

Adds schema, release notes, and release manifest under docs/.

docs/release_manifest.md

@@ -0,0 +1,47 @@
+# Hugging Face Release Manifest
+
+This package is staged for upload to a Hugging Face dataset repository.
+
+## Public Files
+
+```text
+huggingface_release/
++-- README.md
++-- data/
+|   +-- synthetic_bh_cases_v1_mdp_180.json
++-- docs/
+|   +-- schema.md
+|   +-- release_notes.md
+|   +-- release_manifest.md
+|   +-- huggingface_upload.md
+```
+
+## Source Data
+
+The public data file is copied from:
+
+```text
+clin_auth_bench/data/release/synthetic_bh_cases_v1_mdp_180.json
+```
+
+## Exclusion Policy
+
+The package must not include:
+
+- `clin_auth_bench/data/review_batches/`
+- `clin_auth_bench/data/archive/`
+- `clin_auth_bench/data/pilot/`
+- local audit result JSON files
+- `__pycache__/`
+- `.DS_Store`
+- generated QA output directories
+
+## Release Framing
+
+Use the phrase "synthetic inpatient health authorization benchmark."
+
+V1 should be described as focusing on adult inpatient psychiatric authorization.
+
+Avoid implying that the benchmark is derived from real patient data, a real EHR, a real facility, a payer policy, or a provider workflow.
+
+All chart dates should be described as synthetic timeline artifacts that do not correspond to real patient encounters or real facility operations.

docs/release_notes.md

@@ -0,0 +1,51 @@
+# ClinAuthBench V1 Release Notes
+
+ClinAuthBench v1 is a synthetic inpatient health authorization benchmark. V1 focuses on adult inpatient psychiatric authorization over dense 72-hour chart packets.
+
+## Included
+
+- `README.md`: Hugging Face dataset card.
+- `data/synthetic_bh_cases_v1_mdp_180.json`: 180-case public release file.
+- `docs/schema.md`: public schema summary.
+- `docs/release_notes.md`: this release note.
+- `docs/release_manifest.md`: packaging manifest and exclusions.
+- `docs/huggingface_upload.md`: upload and loading notes.
+
+## Not Included
+
+The Hugging Face release package intentionally excludes:
+
+- review batch files
+- archived pilot files
+- earlier pilot files
+- local audit result JSON files
+- `__pycache__`
+- `.DS_Store`
+- generated QA output directories
+
+## Dataset Composition
+
+- Total cases: 180.
+- Continued-stay cases: 108.
+- Safe/LLOC-ready cases: 72.
+- Contradiction/conflicting-documentation cases: 22.
+- Rule-based MDP cases: 120.
+- Probabilistic MDP cases: 60.
+- Probabilistic MDP trace range: cases 121-180.
+
+## QA Summary
+
+- Deterministic QA passed: PASS=180, REVIEW=0, FAIL=0.
+- Claude PHI audit sampled 30 full-content cases: PASS=30, FAIL=0.
+
+## Scope
+
+V1 is evaluation-scale and should be used for benchmark evaluation, error analysis, prompt testing, and small calibration studies. It is not training-from-scratch scale.
+
+The dataset uses generalized synthetic authorization criteria and does not represent any payer, provider, facility, EHR, contract, or proprietary utilization-management rule set.
+
+All chart dates are synthetic timeline artifacts and do not correspond to real patient encounters or real facility operations.
+
+## License
+
+ClinAuthBench v1 is released under the Creative Commons Attribution 4.0 International license (`CC BY 4.0`).

docs/schema.md

@@ -0,0 +1,73 @@
+# ClinAuthBench Schema
+
+This document summarizes the public v1 schema for the Hugging Face release package.
+
+## Top-Level Record
+
+```text
+{
+  "id": string,
+  "title": string,
+  "metadata": object,
+  "content": string
+}
+```
+
+## Top-Level Fields
+
+- `id`: stable synthetic case identifier, such as `clin_auth_bench_v1_0001`.
+- `title`: short case title.
+- `content`: model-facing synthetic chart packet containing repeated 72-hour inpatient documentation forms.
+- `metadata`: structured benchmark metadata, generation metadata, quality checks, and gold labels.
+
+## Metadata Fields
+
+- `benchmark`: benchmark name.
+- `version`: dataset version.
+- `synthetic`: boolean synthetic-data marker.
+- `privacy_design`: short privacy-design description.
+- `level_of_care`: v1 level-of-care scope.
+- `documentation_window_hours`: documentation window size.
+- `diagnosis_category`: synthetic diagnosis family.
+- `trajectory`: synthetic trajectory descriptor.
+- `documentation_challenge`: primary documentation challenge.
+- `documentation_challenge_tags`: more specific challenge tags.
+- `mdp_model`: `rule_based_v1` or `probabilistic_v1`.
+- `mdp_trajectory`: synthetic transition trajectory metadata.
+- `gold`: structured labels for evaluation.
+- `quality_checks`: deterministic generation quality checks.
+- `content_gold_checks`: deterministic content/gold consistency checks.
+
+## Gold Fields
+
+- `current_suicide_risk`: text label for current risk reasoning.
+- `historical_suicide_risk`: text label for historical risk reasoning.
+- `primary_diagnosis`: synthetic primary diagnosis.
+- `psychosis_present`: boolean.
+- `substance_use`: list of synthetic substance-use categories.
+- `safe_for_lloc`: boolean lower-level-of-care readiness label.
+- `expected_los_recommendation`: expected length-of-stay recommendation label.
+- `lower_level_of_care_barriers`: list of unresolved barriers.
+- `documentation_challenge`: primary challenge category.
+- `documentation_challenge_tags`: challenge tags.
+- `key_evidence_forms`: forms expected to contain important evidence.
+- `evidence_anchors`: structured hints linking gold fields to supporting forms.
+- `do_not_claim`: explicit unsupported claims the model should avoid.
+- `form_count`: number of rendered forms in the chart packet.
+
+## Documentation Challenge Values
+
+- `current_vs_historical_risk`
+- `contradiction`
+- `lower_level_of_care_barrier_reasoning`
+- `missing_invalid_or_stale_evidence`
+
+## MDP Trajectory
+
+Cases 1-120 use `rule_based_v1`. Cases 121-180 use `probabilistic_v1`.
+
+Rule-based trajectory entries include the day, action, and resulting synthetic patient state.
+
+Probabilistic trajectory entries include transition metadata such as transition options, conditioned probability mass, sample roll, selected outcome, selected probability, and resulting synthetic state.
+
+The MDP trajectory is synthetic generation metadata. It is not an observed clinical trajectory and should not be treated as real patient-course data.