Files
DynamisLab/docs/JFM_WYQ/ROUND4_TOP_DOWN/00_QUALITY_GATE.md
T
Frank14fandCursor 0ade812864 docs(jfm): establish traceable manuscript planning baseline
Track the research dossiers, section freezes, supporting manuscript materials, and round-aware agent controls so future drafting decisions can be reviewed across both repository mirrors.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-08-10 18:51:53 +08:00

3.6 KiB

Round 4/5 Quality Gate

Round 4 freezes author-facing arguments and section designs; Round 5 drafts from those freezes. Neither round creates evidence. Decision value and manuscript design outrank procedural volume.

Entry gate — coordinator packet

  • Exact unit, round, output decision and file are named.
  • Exact controlling reads and accepted dated author feedback are listed.
  • Available material/artifact inventory distinguishes text-, data-, script-, panel- and dependency-readiness.
  • Target words/pages and section/manuscript share are stated.
  • Paragraph jobs and closing bridge form an argument, not an inventory.
  • Concrete display layout gives count, panel geometry, placement, job, owner, source, readiness and fallback.
  • Imports, exports, exclusions and exact bounded questions are explicit.

A subagent failing this gate is not launched to rediscover the project.

Process gate — continuation and routing

  • When prior decisions matter, agent begins with targeted Nowledge memory/thread recall, then verifies against repository authorities; memory never overrides artifacts.
  • Repository tools handle internal evidence. Undermind is literature-only after get_orientation and list_workspaces.
  • Read-only scripts/derivations use a designated conda environment discovered from authority/listing; no environment is guessed.
  • No CFD, training, artifact generation or expensive analysis ran without authorization.
  • Coverage ledger is compact and limited to controlling reads/gaps (read, mapped, not-covered, external-inaccessible).

Section design/freeze gate

  • Recommendation answers the exact author-facing question and explains consequence/trade-off.
  • Length/share, paragraph progression, display cadence and evidence density are credible.
  • Claims retain owner, chain/configuration, Reynolds convention, role, comparator, metric/estimand, ROI/window and authority.
  • Imports stay brief; exports and deletion owners prevent duplication.
  • Displays/equations state what they establish and do not establish; text readiness is separate from artifact readiness.
  • Author choices remain discuss/blocked or receive an accepted decision; safe defaults are not falsely frozen.
  • Kan99b K2 stays exactly partial pass; the mean-lift sign difference is attributed to CFD coordinate/force-direction convention, never unresolved physics or suspected wrong rotation.
  • Fixed SR/CCD/OID/steady/experiment and cross-chain/estimand cautions pass.

Two-section cadence and second audit

  1. Coordinator presents two detailed packets together.
  2. Author responds to both; coordinator records accepted, provisional and open feedback.
  3. Coordinator launches two tightly scoped continuation agents in parallel.
  4. Coordinator re-reads both outputs and controlling authorities, checks interfaces/duplication/claims/displays/choices, and records a second audit.
  5. Only then may R4 become frozen or R5 prose become drafting-ready.

Statuses

  • researching: a named controlling read is incomplete.
  • discuss: a bounded author choice is decision-ready.
  • revise: targeted argument/design/interface correction is needed.
  • frozen: coordinator second audit locked the R4 text contract.
  • blocked: required authority, material, role, comparator, decision or conflict resolution is absent.

Final Round-4 audit checks the paired sections in the whole-manuscript order, ownership, bridges, claim ladder, display/equation inventory, accepted author decisions and Round-5 handoff. Optional omitted panels do not block text when a safe fallback preserves the argument.