---
description: General JFM startup, authority, memory, literature, and safety routing
alwaysApply: true
---

# JFM Research Start and Tool Routing

## Start from the assigned unit
- Name the project, active round, exact unit, deliverable, output, and exclusions. R1 indexes evidence; R2 designs; R3 builds dossiers; R4 freezes arguments/displays; R5 writes admitted section prose and controls from frozen R4 packets.
- Read fresh supplied context once. Follow the coordinator packet's exact reads, ownership, limits, questions, and acceptance checks; do not rediscover the project or recreate R1 coverage.

## Recall, then verify
- Use 2–3 narrow Nowledge `memory_search` queries containing `DynamisLab JFM` plus round, section/unit, and decision/claim/term. Use `thread_search` and scoped `thread_fetch_messages` only when exact prior wording or conversation history is needed.
- Memory navigates decisions; `mem_fs` may locate source/original text only when actually stored, with source state retained. It is not a general full-text corpus.
- R1–R4 indexes, dossiers, plans, controls, and figure plans route architecture, ownership, ceilings, and fallbacks; they do not establish a new internal fact or verify/promote an artifact.
- Bind internal facts/numbers to current package authority, manifest, code/configuration, or immutable artifact. If unbound, omit, use the frozen fallback, mark an authorized placeholder, or report the gate. Current authority overrides memory and summaries.
- Search before `memory_add`; update rather than duplicate. Save only durable decisions, procedures, or lessons.

## Literature and tool routing
- For Undermind: run `get_orientation`, list/select the DynamisLab workspace, inspect existing folders/files/searches, search the library first, and broaden only if needed. Confirm identity, DOI, and availability with `get_paper_info`.
- Keep each `read_pdfs` request bounded and self-contained: identify the papers/pages or question, request only necessary passages, and preserve `full-PDF`, `full-report`, `mapped`, `snippet`, or `inaccessible` state. Mapped/snippet evidence cannot support detailed wording or structural claims.
- Literature supports external propositions and rhetorical/structural diagnosis only; it never verifies internal results, numbers, artifacts, priority, readiness, or claim ceilings.
- Before any authorized read-only script, discover and use the repository-designated environment and documented entry point; do not guess an environment or install/change dependencies.

## Safety and closeout
- Do not run CFD, training, artifact generation, or expensive/new analysis without explicit authorization. Never fill an evidence gate by inference.
- Close out whether Working Memory, durable memory, thread recall, `mem_fs`, scripts, and literature were used, distinguishing `read`, `mapped`, and independently verified/audited evidence.