---
description: DynamisLab JFM research, decision, and drafting workflow
alwaysApply: true
---

# JFM Research and Manuscript Workflow

## Round-aware operating mode
- Identify the active round and output before work. R1 indexes evidence; R2 designs the manuscript; R3 builds section dossiers; R4 settles author-facing arguments, ownership, displays and section freezes; R5 drafts prose only from frozen R4 packets and authorities.
- Keep evidence states distinct: `mapped` is not `read`; `read` is not independently verified or artifact-audited. Use a compact ledger (`read`, `mapped`, `not-covered`, `external-inaccessible`) proportional to the round; do not let governance overwhelm argument/design.
- In R4/5 prioritize the section question, bounded answer, paragraph jobs, display logic, ownership, interfaces, author choices and manuscript value. Do not recreate R1 inventories or rediscover the project.

## Continuation and context
- When prior decisions matter, continuation agents begin with targeted Nowledge `memory_search`; use thread recall for exact prior discussion. Then verify every operative decision against current repository authorities. Memory aids discovery/history and never overrides authorities, manifests, code/configuration or immutable artifacts.
- Before launching section agents, the coordinator creates a minimum context packet containing: exact controlling reads; accepted author feedback; available material/artifact inventory with readiness; requested output decision; exclusions; target length/section share; paragraph jobs; concrete display layout; ownership; and exact questions. Agents stay within that packet and perform only targeted authority checks.
- Cadence after the initial architecture: present two detailed section packets together; receive author feedback; launch two tightly scoped continuation agents in parallel; then perform a coordinator second audit before either section is frozen or handed to R5.

## Tool routing and execution safety
- Nowledge: decisions, preferences and discussion history. Undermind: external literature only, after `get_orientation` and `list_workspaces`; literature cannot establish internal results. Repository read/search tools: internal evidence, authorities, code, manifests and artifacts.
- For read-only derivations or scripts, discover the designated conda environment from current repository authority or environment listings and record it; never guess an environment. Do not run CFD, training, artifact generation or expensive analysis unless explicitly authorized.

## Evidence and claim boundaries
- Trace claims to current authorities and preserve solver, geometry/configuration, Reynolds convention, case/seed/role, comparator, metric/estimand, ROI/window and artifact. Keep internal Legacy and modern V5 chains separate; use physical configuration names in manuscript prose.
- Use case-specific periodic, transient and steady metrics. Distinguish sourced fact, bounded observation, interpretation, historical/negative result and open question; report gaps and conflicts without inventing citations or numbers.
- Image work inventories existing data/scripts before proposing panels; inventory is not panel readiness. Label existing derivations, independent checks and new conjectures separately.

## Fixed scientific cautions
- Bound positive SR to Kármán/cloaking. Illusion SR is historical/negative; CCD is descriptive/noncausal and uses `action-correlated correction-field modes`; OID is claim-free; steady results are contextual; experiment is qualitative/provenance-gated.
- Kan99b K2 retains the authority's exact overall `partial pass`. Its opposite mean-lift sign is explained by the CFD coordinate/force-direction convention, not unresolved physics or suspected wrong rotation; do not call the sign unresolved.
- Never promote claims across chains, alter evidence packages/archive/history, or inflate solver qualification into controller or mechanism validation.