The contract layer

Departments

Eleven departments, each with one job, a fixed tool list, and a stated hand-off. This page compiles departments/*/CONTEXT.md — read one of those files for its full contract; this is the map of how the eleven fit together.

Numbering is sequence, not importance. 00 through 50 (plus 60) are the stages a candidate actually moves through in order; 80, 90 and 95 sit outside that sequence on purpose — they reason over or render what a candidate already has, and advance no stage:, no status:, and no score. See Flow for the sequence itself, including where it loops back on itself rather than running straight through.

00_scouting

Watch for facts that changed

One job: find facts that should trigger re-scoring of ideas already logged, and write them down. Stage 0. Tier 1 — free to run unattended, on a schedule.

Stage 0 · Tier 1
Command /scan
Writes knowledge-base/signals/log.md
Hands to Flags an affected candidate's stage: for re-scoring (→ 20_scoring), and watches every killed idea's revive condition for the fact that would reopen it (→ 10_front-end).
05_macro-trends

Read the world a device would ship into

One job: write sourced, falsifiable briefs on what is moving — so Stage 1 generation has steer. Stage 0.5. Tier 1, on its own cadence.

Stage 0.5 · Tier 1
Command /seed-scan
Writes macro-trends/seeds/, macro-trends/watchlist/
Hands to Seeds steer 10_front-end whenever it runs — read as standing context, never chained to it automatically.
10_front-end

Own the opportunity, and run the loop

One job: frame an opportunity, say what it rests on, convene the chairs, and change its shape when a statement fails. Tier 1.

Stage 1, and the loop · Tier 1
Command /generate, /front-end-loop, /round-table, /impact, /pivot
Writes knowledge-base/candidates/idea-XXX.md, knowledge-base/loop/idea-XXX.md
Hands to → 20_scoring (Stage 2, first-pass triage). And, continuously, → the seven chairs, whose conditions come back as unverified claims for 30_verification — a failed one returning here as an impact review.
11_regulatory

The pathway chair

One job: say what regulatory route this opportunity is on, what the claimed indication costs in evidence, and what would have to be true for that route to hold. A chair, not a stage. Tier 1.

Stage — (chair) · Tier 1
Command /position
Writes knowledge-base/positions/idea-XXX-regulatory.md
Hands to → 10_front-end as a verdict and a set of conditions; the conditions go to 30_verification as unverified claims like any other.
12_clinical

The evidence chair

One job: say whether the mechanism does the thing, and what would have to be shown, in whom, before a clinician acts on it. A chair, not a stage. Tier 1.

Stage — (chair) · Tier 1
Command /position
Writes knowledge-base/positions/idea-XXX-clinical.md
Hands to → 10_front-end. Its abstains verdicts are usually research requests → 45_research.
13_market-access

The payment chair

One job: say whether money can reach this device at all, through which existing policy, and what would have to be true for that route to hold. A chair, not a stage. Tier 1.

Stage — (chair) · Tier 1
Command /position
Writes knowledge-base/positions/idea-XXX-market-access.md
Hands to → 10_front-end. Shares the coverage question with 80_commercial, which sizes what this department establishes can be paid for at all.
14_technology

The feasibility chair

One job: say whether this can be built with what exists, what in it is still research, and what would have to be true for the core technology to be ready. Absorbs the R&D question. A chair, not a stage. Tier 1.

Stage — (chair) · Tier 1
Command /position
Writes knowledge-base/positions/idea-XXX-technology.md
Hands to → 10_front-end, and → 15_operations-npi, which reads this position first because what the thing is bounds what making it costs.
15_operations-npi

The make-it chair

One job: say whether this can be made, at a cost and a volume that works, and what industrialising it would actually take. A chair, not a stage. Tier 1.

Stage — (chair) · Tier 1
Command /position
Writes knowledge-base/positions/idea-XXX-operations-npi.md
Hands to → 10_front-end. Argues time-to-revenue from the making end, where 80_commercial argues it from the money end.
16_legal-ip

The freedom chair

One job: say whether this design is free to operate, what in it is defensible, and what someone else already owns in this shape. A chair, not a stage. Tier 1.

Stage — (chair) · Tier 1
Command /position
Writes knowledge-base/positions/idea-XXX-legal-ip.md
Hands to → 10_front-end. Raises more research requests than any other chair, because there is no landscape primitive → 45_research.
20_scoring

Apply the rubric, twice

One job: score a candidate against the 9-factor rubric. Stages 2 and 4 — one contract, two entry points, because it is the same agent doing the same thing against different evidence. Tier 1.

Stage 2, 4 · Tier 1
Command /score, /rescore
Writes the Scores table (Stage 2); the Verified Composite (Stage 4)
Hands to Stage 2 → 30_verification, if judged worth the cost (Gate A). Stage 4 reads 30_verification's tags → 50_gate (Stage 5).
30_verification

Check every claim against a primary source

One job: independently confirm or refute each unverified claim. Stage 3. Tier 2 — the mandatory gate. Nothing else in this repo is structurally required.

Stage 3 · Tier 2 — mandatory
Command /verify
Writes confidence tags, in place
Hands to → 20_scoring (Stage 4, rescore). Runs again, unchanged, on whatever 45_research adds — the same claims, a second pass.
45_research

Build the evidence base under a candidate

One job: turn "we don't know" into sourced context, so verification has something to check. Stage 4.5, plus the research fleet that runs on its own clock. Tier 1.

Stage 4.5 · Tier 1
Command /research, /research-fleet
Writes knowledge-base/domains/, knowledge-base/research/, unverified claims
Hands to → 30_verification → 20_scoring → 50_gate — its own contract calls this “the loop that actually turns research into coverage.” Also works rows 80_commercial commissions (commercial-gap).
50_gate

Kill it, or finish the packet

One job: apply the hard kill conditions, then finish what survives so a human can read it. Stages 5 and 5.5. Tier 1.

Stage 5, 5.5 · Tier 1
Command /gate-check, /wrap
Writes the Gate Check section; the ## Wrap section
Hands to Cleared and wrapped → 60_human-review (Stage 6). Killed → the Librarian (inside 60_human-review) logs it with a revive condition 00_scouting watches for. Under 6/9 factors verified → looped back to 45_research as queued-research, never treated as a kill.
60_human-review

Where agents stop

One job: assemble what a human needs to decide, and record what they decided. Stages 6, 7 and 8. Tier 3 — the boundary this whole repo is built around.

Stage 6, 7, 8 · Tier 3 — human only
Command /shortlist, /diligence, /log-outcome
Writes the packet; human_decisions (humans only); calibration-log.md
Hands to Approved → Stage 7 diligence. A real-world outcome, whenever it lands → Stage 8, which feeds 00_scouting and 10_front-end — the flywheel's one long loop, spanning the human boundary.
80_commercial

How big it could be, how the money would work, and the routes to it

Two jobs, two artifacts, in that order. A market model — the chain that would produce a size, with every leg resting on evidence, a declared assumption, or a piece of research this department commissioned itself. Then a falsifiable business case — who pays, from which budget, how much, how often — laying out the commercial routes to first revenue without picking one.

Stage — · Tier 1
Command /market-model, /business-case
Writes knowledge-base/market-models/, knowledge-base/commercial/, research rows
Hands to An unfillable market-model leg becomes a commissioned row in 45_research (source_of_question: commercial-gap) — the only artifact in the repo that gets better on a rebuild, not just fresher.
90_artifacts

Views onto candidates, for people

One job: render what is already known into something a person can read. Artifacts, not stages. Nothing here scores, gates, or touches stage: / status:, and /flywheel-cycle runs none of it.

Stage — · Tier 1
Command /pitch-decks, /narrate
Writes knowledge-base/decks/, knowledge-base/narratives/, review/
Hands to Reads whatever the candidate, its dossier(s) and 80_commercial already state, and renders it. Nothing downstream reads an artifact back.
95_marketing

The story, and what substantiates it

One job: draft what would actually be said — to investors, to users, to payers — and bind every line to the claim that substantiates it, or mark it unsubstantiated. An artifact, not a stage. Tier 1. Nothing here scores, gates, or touches stage: / status:, and /flywheel-cycle does not run it.

Stage — · Tier 1
Command /messaging
Writes knowledge-base/messaging/
Hands to Reads everything else last — claims, decks, the business case — and writes any surfaced assumption back into ## Claims as unverified.
Generated from departments/CONTEXT.md by engine/scripts/build_review.py on 2026-09-13. This page is a view — edit the source file and rebuild, never edit here.