Skip to content

execution

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

execution

README.md

Execution

  • Status: Active (v0.1.0 release; T0 complete, T1 in progress)

This directory holds MedScale's execution and phase-planning documents — the concrete tickets and specifications that translate the strategic foundation into work. T0 is complete; T1 has produced the literature database, screening workflow, and AI triage module, and new execution artifacts are authored as phases proceed.

Intended contents (to be authored when each phase begins; not yet present)

Document Phase Purpose Status
search_strategy.md T1 Per-source queries, PRISMA workflow, and reproducible ingestion design ✅ Authored (v1)
benchmark_spec.md T3 Task and metric definitions for MedScale-Bench Missing — to author
constrained_decoding_hypothesis.md T2 The form-vs-content experiment design (referenced by the vision) Missing — to author
Phase tickets (T0–T7) all One-ticket-per-session execution records Missing — to author

Phase map (see the Strategic Blueprint §6–§8 and §17)

  • T0 — Repository foundation (✅ complete — ADR-0003/0004)
  • T1 — Literature database & evidence foundation (🟡 in progress — medscale.litdb, ADR-0009)

Screening the corpus (medscale screen)

Human title/abstract screening turns the deduplicated corpus into a validated evidence foundation. No model is involved — decisions are the operator's, recorded in an append-only audit trail (data/litdb/screening/review_log.jsonl).

uv run medscale screen status                    # PRISMA counts (reproducible from the log)
uv run medscale screen duplicates --reviewer <you>  # FIRST: resolve uncertain duplicate groups
uv run medscale screen next --reviewer <you>     # then screen the pending queue; q to stop
uv run medscale screen resume --reviewer <you>   # same as next; picks up where you left off
uv run medscale screen next --query Q2 --limit 50   # topic-batched, capped session

Order matters: resolve the uncertain-duplicate groups before title/abstract screening (ADR-0017 ordering discipline; screen status gates on this). Recommended topic order: Q2 → Q6 → Q4 (serves the novelty verdicts first).

Per record: [1] Include [2] Exclude [3] Maybe [4] Duplicate [5] Skip [q] Quit. Exclude prompts for a machine-readable reason. Decisions map to PRISMA stages (include→screened, exclude/duplicate→excluded); corrections are new events (history is never edited). Interrupt any time — progress is saved after every decision.

  • T2 — FHIR toolkit (fhirkit)
  • T3 — MedScale-Bench
  • T4–T7 — Model landscape, training pipeline, MESC-v0 adapter, evaluation

This is a documentation placeholder. No engineering content lives here. Nothing in this directory authorizes starting a phase; phases begin only under explicit operator approval.