Which plan are we working?
A planning session is one plan carried through the seven steps of its planning methodology (MDMP or JPP). It holds the mission, the facts and assumptions, the candidate COAs, the wargame results and the approved COA — and it persists, so the Autonomy Adjustment Function is reading a real plan's condition, not a slider.
Open a session
New session
Current session
Planning progress
Stage workspace
Stage products
Planning assistant —
Extract from documents
Scan this session's ingested documents (screen 04) for candidates — each cited to its source. Nothing merges automatically; you check each item and add only what you keep.
Document extraction — Mission Analysis
What autonomy does the plan's condition permit?
Seven candidate formulations run on the same scenario at once. The scenario is derived from the live session — task complexity C from the identified tasks, time compression T from the clock, information uncertainty U from the stated intel gaps — unless you override it below. Where the formulations disagree by a whole governance level, the allocation is not robust and the call belongs to a human.
Derived scenario
Higher-order trust factors
Governing formulation
Methods on display
—
| Method | A(s) | Governance level |
|---|
What has the plan actually been given to read?
Upload a higher order, an intel summary, or any other planning document — PDF, DOCX, or plain text. It is extracted, chunked, and made searchable within this session only. Nothing here writes itself into a stage's products yet; retrieval returns cited passages for a planner to read and copy in by hand.
Upload a document
Simulated OSINT (exercise injects)
Session documents
Search across this session's documents
What was decided, and who let the AI in?
Append-only. Every stage transition, every governed assist and whether its draft was accepted, and every human confirm or override of an autonomy level.
After-action review — human-machine teaming
Audit trail
How the tool works (and its limits)
Three components over one doctrine base. The Autonomy Adjustment Function is deterministic and model-free — it computes an autonomy level for each stage from the session's C/T/U and seven candidate formulations, clipped to the Level III ceiling. The planning assistant retrieves the governing doctrine and drafts a stage's products, but only within the governed level — and never at a human-anchored stage. The wargame assistant generates action–reaction–counteraction turns for COA Analysis.
Deterministic where it matters. The autonomy levels and the structured drafts come from inspectable rules, not a model call, so the same session state always produces the same allocation and the same draft. The model streams the explanation.
Architecture. A FastAPI backend with the reasoning model behind one swappable interface (Claude direct or via Bedrock in cloud/GovCloud, an open-weight model on vLLM/Ollama air-gapped, a deterministic mock offline) and retrieval behind another. The frontend is dependency-free static files. One process serves both.
Limits. This is a governance instrument, not a planning system of record. The autonomy levels are illustrative — they plot where a formula would place authority, not what any real system should do. The doctrine base is a seven-stage summary with references, not the field manuals themselves. A worked example is fictitious training data.