[feedback][clinical-trial-ipd-sim]Administrative Censoring Lacks Enrollment Stagger Guidance

[lengning]

Gap in Skill Specification

The endpoint derivation in §7 specifies:

PFS_DAY = min(progression_day, death_day, ADMIN_CENSOR_DAY)

But ADMIN_CENSOR_DAY is treated as a single constant (from CTGov data cutoff). The skill does not specify that this should be per-patient based on enrollment timing.

Consequence

All censored patients have identical follow-up time — a point mass in the censoring distribution. This is:

1. Immediately detectable as synthetic data
2. Structurally incorrect (violates the trial's actual design where enrollment spans years)
3. Distorts KM curve behavior at the tail (all censoring happens simultaneously)

Proposed Addition to Skill

Add to scm_spec_template.md §7 or SKILL.md Step 5:

§7.2 Administrative Censoring with Enrollment Stagger

ADMIN_CENSOR_DAY is patient-specific, derived from:
    enrollment_offset_i ~ f(enrollment_pattern)
    ADMIN_CENSOR_DAY_i = data_cutoff_date - enrollment_start - enrollment_offset_i

Enrollment pattern options (from protocol/CTGov):
    - Uniform(0, enrollment_window) — simplest; adequate for most
    - Piecewise-linear (ramp-up, plateau, wind-down) — more realistic
    - Read directly from CTGov "Study Start" and "Primary Completion" dates

The median of ADMIN_CENSOR_DAY_i should equal the published median follow-up.
Range should span [min_followup, max_followup] from the publication.