docs/phase_workflow.md

Clinical Trial Phase-Aware Workflow & Scoring

10-Phase Clinical Workflow

The agent learns this ordering through reward signal, not hard-coding.

#	Phase	Description	Actions
0	literature_review	Understand disease and constraints	Review scenario
1	hypothesis	Form hypothesis about drug mechanism	Estimate expected effect
2	phase_i_design	Phase I safety/dose-finding	run_dose_escalation, observe_safety_signal
3	phase_i_analysis	Analyze Phase I results	estimate_effect_size
4	phase_ii_design	Design Phase II efficacy trial	set_primary_endpoint, set_sample_size, set_inclusion_criteria, etc.
5	regulatory	FDA review	submit_to_fda_review, request_protocol_amendment
6	enrollment	Enroll patients	(implicit after FDA approval)
7	monitoring	Interim analysis, adaptation	run_interim_analysis, modify_sample_size, add_biomarker_stratification
8	analysis	Final statistical test	run_primary_analysis
9	conclusion	Synthesize results	synthesize_conclusion

---

Phase-Order Scoring

Condition	Reward
Action in correct or next phase	+0.2
Action stays in current phase	+0.2
Action skips N phases ahead	−0.3 × N

Judge persona scaling by tier:

Tier	Persona	Forward Bonus	Skip Penalty	Extras
Warmup	Junior	+0.20	−0.30/skip	Allows 1 skip free, gives hints
Beginner	Junior→Senior	+0.20	−0.30/skip	Standard
Intermediate	Senior	+0.15	−0.30/skip	Expects correct ordering
Advanced	Senior→Principal	+0.10	−0.50/skip	Redundancy penalty −0.10
Expert	Principal	+0.05	−0.50/skip	Redundancy −0.15, efficiency penalty

---

Hard Prerequisites

These block the action entirely (not a reward signal — returns error):

Action	Requires
estimate_effect_size	≥1 run_dose_escalation
set_sample_size	estimate_effect_size
submit_to_fda_review	set_primary_endpoint + set_sample_size
run_interim_analysis	submit_to_fda_review passed
run_primary_analysis	submit_to_fda_review passed
synthesize_conclusion	run_primary_analysis
modify_sample_size	run_interim_analysis
add_biomarker_stratification	estimate_effect_size

---

Protocol Amendment

• request_protocol_amendment allows recovery from FDA review failure
• Costs time and budget (realistic consequence)
• Successful recovery: +0.3 recovery bonus
• Maximum 2 amendments per episode