01
Role-aware programming
Separates developer workflows from independent QC workflows so the agent does not reuse the wrong macros, assumptions, or validation depth.
Why it exists
Prompt-only AI is not enough for clinical programming.
General coding agents can write R or SAS, but clinical work needs controlled behavior: correct study context, role separation, source verification, execution location, validation evidence, and reviewer-facing rationale. CSP Agent Harness turns those expectations into repeatable workflow gates and artifacts.
RoleDeveloper vs independent QC behavior
ContextRegistry, specs, shells, source evidence
RunLocal or remote R/SAS execution policy
AuditMismatch ledger and dashboard
Core capabilities
Clinical programming guardrails for AI agents.
02
Study context packages
Builds evidence maps from registry paths, specs, SAP summaries, source inventories, codelists, shells, parser audits, and known quirks.
03
Source evidence before code
Requires the agent to verify datasets, variables, types, codelists, filters, and missingness before implementing derivation or reporting logic.
04
ADaM and TLG support
Supports dataset generation, independent QC, TLG code briefs, source plans, program skeletons, and output validation workflows.
05
Risk-based QC levels
Defines Level 1, Level 2, and Level 3 QC scope, including independent double programming for high-value ADaM datasets and unique TLGs.
06
Audit dashboards
Produces reviewer-facing audit indexes and batch dashboards linking plans, programs, logs, compares, QC summaries, and mismatch decisions.
Adapter layer
Works with the AI coding agents your team already uses.
The harness is not a replacement coding model. It supplies clinical workflow discipline around tools such as Codex, GitHub Copilot, Claude Code, and AGENTS.md-aware assistants.
CLI
Python command layer
Doctor checks, registry inspection, work-order creation, source probes, QC planning, TLG briefs, and audit indexes.
R
R utility tools
Source evidence probes and dataset comparison utilities for controlled clinical programming workflows.
ENV
Execution profiles
Local or remote R/SAS runtime profiles with data location, write boundaries, and clinical execution policy.
AI
Agent adapters
Instruction templates for Copilot, Claude Code, and generic agent environments.
Want to pilot AI-assisted ADaM or TLG work without losing clinical control?
Snowbird can help install the harness, configure execution profiles, define study registries, and run a first controlled work-order pilot.
Plan a CSP Harness Pilot