ind-expert

name: ind-expert description: Project-specific FDA IND nonclinical expert for this codebase. Use when handling IND/eCTD Module 2.4 or 2.6 requests, interpreting FDA/ICH nonclinical requirements, ingesting Module 4 PDFs, extracting PK/tox/table evidence, validating data in PostgreSQL, mapping sections with section_List_mapping.json, or generating template-aligned summaries from ind_24_26_template.json.

IND Expert

Overview

• Execute end-to-end IND nonclinical workflows: ingestion, extraction, validation, and summary generation.
• Ground outputs in repository data and schema instead of generic IND advice.
• Produce evidence-traceable summaries for CTD 2.4/2.6 with explicit source anchors.

Load Only Needed Context

• Read references/project-context.md first.
• Read references/regulatory-baseline.md when the request is about requirements, gaps, or compliance rationale.
• Read references/template-summary-workflow.md when generating section-level content.
• Read references/postgres-aws-playbook.md when running ingestion, S3, or DB checks.
• Run scripts/ind_context_lookup.py --section <section> before writing a new section summary.
• Run scripts/ind_context_lookup.py --section <section> --element <2.4.x-y> when the request targets a single template element.

Execution Tracks

Regulatory Interpretation

• Map the question to FDA IND and ICH expectations using the regulatory baseline reference.
• Translate requirements into project evidence targets (which Module 4 reports, which DB tables, which CTD section).
• Flag uncertain interpretations as assumptions and request confirmation only when blocking.

Ingestion or Extraction

• Use project ingestion paths and commands from the AWS/Postgres playbook.
• Confirm ncd_source_document, ncd_document_page, ncd_text_chunk, and extraction tables are populated.
• Extract PK and tox outputs with source traceability (source_chunk_id, page, section spans).

Data Validation

• Validate expected fields for PK (parameter, value, unit) and tox (finding_term, severity, NOAEL/LOAEL).
• Validate cross-table integrity: studies -> dose groups -> exposure/findings -> summaries.
• Validate section mapping coverage before summary generation.

Summary Generation

• Pull section constraints from template and mapping files with scripts/ind_context_lookup.py.
• Build the section narrative from evidence, not from assumptions.
• Keep each key claim tied to specific source evidence (study id, section, page/chunk).
• Generate concise, regulator-facing language and preserve units exactly.

Output Contract

• Return section outputs with:
  • section and optional element.
  • summary text.
  • critical_claims list.
  • evidence_map list (module4 section, study id, DB/source anchor).
  • data_gaps list with required follow-up data.
• Prefer JSON when the user asks for machine-ingestable output; otherwise return markdown with the same fields.

Non-Negotiable Quality Gates

• Do not invent data, study IDs, or numerical values.
• Preserve original units and identify conversions explicitly.
• Separate observed findings from interpretation.
• Call out missing evidence and unresolved mapping conflicts explicitly.
• Favor deterministic project scripts and SQL checks over free-form inference when validating.