# Drug-Program Orchestration — how to use it

You now have a disease-agnostic pipeline that runs a full therapeutic program end to end, with
you in the loop at every consequential decision. It comes in two pieces:

1. **`drug-program-orchestration`** — the conductor **skill** (published, in the live catalog).
2. **Drug Program Architect** (`DRUG_PROGRAM_ARCHITECT`) — a **specialist agent** that preloads
   the skill and has full access to the skill catalog and all data/regulatory connectors.

## How to invoke it

**Easiest:** switch the session to the Drug Program Architect specialist (session config
popover, or accept the switch prompt), then say:

> "Run the drug program for **<disease>**."

**Or** from any agent:

> "Load the drug-program-orchestration skill and run the pipeline for **<disease>**."

That's the whole entry point. The orchestrator initializes a program dossier and walks the 10
stages, stopping to ask you at each gate.

## The 10 stages

1. **Indication & unmet need** — epidemiology, biology, standard of care, precedent.
2. **Patient priorities / PFDD** — what patients actually want fixed (+ optional survey).
3. **Literature + competitive-pipeline map** — mechanism + who's developing what, and why they
   succeeded/failed.
4. **Data mining → targets** — omics meta-analysis + single-cell → ranked druggable targets.
5. **Target & modality selection** — the pivotal fork; scored rubric → **you confirm**.
6. **Design & pressure-test** — modality-routed AI/ML design + in-silico validation.
7. **Scientific development plan** — AI-to-first-in-human assay cascade, CMC, biomarkers.
8. **Regulatory strategy** — FDA/EMA path, designations, endpoints.
9. **Commercial & financing** — TAM/SAM/SOM, whitespace, financial model, the Ask.
10. **Deliverables & synthesis** — business plan, scientific plan, deck, manuscript, peer review.

Full per-stage detail is in `stage_spec.md`; the composition map (which existing skill each
stage uses) is in the skill's `SKILL.md`.

## How the human checkpoints work

The pipeline **never silently** commits a target, modality, endpoint, price, or go/no-go. It
stops at named gates and asks you a structured question with its recommendation + options:

- **Hard gates** (block until you answer): SCOPE LOCK, WHITESPACE CONFIRM, **TARGET/MODALITY
  LOCK**, **DESIGN GO/NO-GO**, **FINANCING ASSUMPTIONS**, FINAL REVIEW.
- **Soft gates** (it proposes a default, you can veto): PRIORITY RANK, TARGET EVIDENCE,
  REG ASSUMPTIONS, and any LOW-CONFIDENCE flag.

It also stops whenever it's uncertain or a claim needs clinical validation — those become
recorded open questions, not confident-sounding assertions. The full rule set is in
`checkpoint_protocol.md`; the gate table is `assets/checkpoint_map.csv`.

## The program dossier

One JSON file (`program_dossier.json`) is the source of truth for the whole run: disease, scope,
every decision and who made it, per-stage artifacts, gate status, and open questions. It lets a
program be paused and resumed, and makes every decision auditable. Schema:
`assets/program_dossier.schema.json`.

## What's disease- and modality-agnostic

The disease, target, and modality are all parameters resolved from data + your decisions — not
hard-coded. Validated on IPF (see `dryrun_IPF.md`): the same modality router that produced a
tolerizing **vaccine** for EoE produces a **ligand-trap**, **antibody**, or **small molecule**
for IPF depending on the target biology.

## Requirements & caveats

- **Stage 6 (design) needs a GPU compute target.** Add one in the Compute panel; the
  orchestrator checks and will run the CPU-feasible parts or defer design if none is configured.
- Connector calls (ClinicalTrials.gov, Open Targets, FDA, GEO/PRIDE, ChEMBL) run inside the
  orchestrator automatically — the specialist has them all.
- Forward-looking scientific, regulatory, and commercial statements are development hypotheses,
  not established results, medical advice, or guaranteed outcomes.

## Files in the skill
- `SKILL.md` — the conductor + composition map.
- `kernel.py` — dossier/gate/routing helpers (auto-load with the skill).
- `stage_spec.md` — full per-stage blueprint.
- `checkpoint_protocol.md` — human-in-the-loop rules.
- `assets/checkpoint_map.csv` — every gate: question, go-criteria, no-go action.
- `assets/program_dossier.schema.json` — dossier state schema.
- `assets/target_modality_rubric.csv` — Stage-5 scoring rubric.
- `assets/scientific_plan_template.md`, `assets/program_summary_template.md` — fill-in templates.