🌲

Cambium

A full research institution —
from RFP to verified results

46 specialized agents across 11 councils handle grant intake, brainstorming, proposal writing, development, adversarial verification, and reporting. Every claim carries an evidence tier. Every gate requires your explicit approval. Nothing is published without you.

CI Version 1.00.0 46 agents 8 human gates MCP ready License: MIT
🌲 Open Dashboard 🎓 Enter the Academy ▶ 60-second tour
46
Agents
11
Councils
8
Human gates
Smart-Tier
Model profile
CI
Evidence-checked

What it is

One portable Claude project.
Your entire research org.

The Cambium turns a Claude project into a working research organization. You are the Director: you set direction and approve at every gate. A Orchestrator dispatches specialists; Labs, Scouts, and Verification boards do the science; Faculty advise; and a Document Office writes the final deliverable — only from verified findings, only after you approve. It works for any field and at any stage of a project.

🔬
Verification runs code
The three Opus audit boards don't read and opine — they execute your scripts and reproduce the numbers. That's what catches data leakage, unfair baselines, and mis-stated claims.
📋
One output contract, 46 agents
Every agent emits the same five fields with a severity (P0/P1/P2) and a claim tier (Proved / Code-verified / Asserted / Open). Conflicts resolve by rule: whoever ran the code wins.
⚡
Smart-Tier model profile
Opus only where deep adversarial reasoning pays — theory and the audit boards. Sonnet where it's sufficient. Your main model for synthesis and final prose. Best result per token.
🛡️
Governed by construction
Ships an AI Governance policy (authorship, FERPA, IRB, data sovereignty, dual-use), a recorded human-approval ledger, and a validate.py that fails the build on an open P0 or un-evidenced claim.

The lifecycle

RFP → Proposal → Develop → Verify → Report

Two halves, eight gates. At every ◆ gate the institute stops and presents you with the decision, options, and risks. Nothing proceeds without your explicit approval.

Pre-Award
Read RFP
RFP-Analyst extracts requirements, criteria & budget
◆ G1
→
Pre-Award
Brainstorm
Ideation + Faculty rank the idea slate
◆ G2
→
Faculty
Aims + Review
PI writes Specific Aims; faculty disciplines critique
→
Pre-Award
Proposal
Proposal-Writer drafts full submission
◆ G3
→
Post-Award
Develop
Labs + Execution build the methods and run experiments
→
Post-Award
Verify
Three Opus boards run the code; find flaws
◆ G4
→
Post-Award
Report
Reports + decks from verified findings only
◆ G5 / G6

The organization

11 councils, 46 agents — all answering to you

Smart-Tier models: Opus only on the critical reasoning path; Sonnet everywhere it's sufficient; your own model for synthesis and final prose. See the full interactive org chart in dashboard.html.

Orchestration
2 agents · inherit (your model)
Orchestrator dispatches & synthesizes; Document Office is the only writer — gated on your approval.
Pre-Award Office
8 agents · Sonnet + Opus
RFP-Radar + RFP-Analyst (intake & fit), Ideation + Idea-Tournament (Elo-ranked ideas), PI, Proposal-Writer, Budget-Officer, Grants-Compliance. Wins the project.
Partnerships Office
4 agents · Sonnet
Collaborator-Scout, Partnership-Liaison, Program-Manager, Convener. Finds verified collaborators; drafts outreach; runs the shared pre-award room.
Faculty
1 parameterized agent · Opus
Summoned per discipline — statistics, math, CS, ML, economics, or your field. Strongest peer-review objection + the fix. Unlimited fields.
Scouts
3 agents · Sonnet
Prior-Art (novelty gate), Methods, Landscape. Two modes: quick scan / deep research. Verify citations, never invent them.
Labs
4 agents · Opus + Sonnet
Theory (Opus), Methods, Domain, and Statistics (the stats doer — fits models, CIs, power, diagnostics).
Verification
5 agents · Opus + Sonnet
Rigor, Methodology, Evidence boards run the code; Domain board checks utility; Referee scores vs a venue rubric. Code-verified beats asserted.
Execution
4 agents · Sonnet
Experiments, Ablation, Iteration (budget-aware diagnose→tune→re-run tree-search), and Research-Engineer (reproducible code).
Reporting Office
2 agents · Sonnet
Progress / annual reports and slide decks from verified content only. You release at G5/G6.
Support Staff
11 agents · Sonnet + Haiku
Record-Keeper (institutional memory), Librarian (citation-resolution gate), Feedback-Router, Data-Steward, Integrity-Officer, Figures, RA, TA, Office-Manager, Outreach, Janitor.
Governance
1 agent · Opus
Research-Conduct-Officer runs the responsible-research checklist at EVERY gate — IRB/IACUC, COI, FERPA, data sovereignty, dual-use, reproducibility → GO / CONDITIONS / STOP.

Full interactive roster in dashboard.html · charter in INSTITUTE.md.

Governance + CI-enforced honesty

Honesty is structural, not optional

Every claim carries a tier. Every number must trace to a script that was run. The Document Office cannot strengthen a claim beyond its proof. governance/validate.py fails the CI build on any open P0 or un-evidenced claim.

Claim tiers — on every finding

ProvedTheorem with a checked proof
Code-verifiedScript ran it; cite command + output
AssertedStated in a doc, not run or proved
OpenAcknowledged, unresolved

When two agents conflict, code-verified beats asserted — the agent who ran the code wins.

Eight human gates — non-negotiable

  • G0Know the PI & mandate (Director)
  • G1Pursue this RFP? (Director)
  • G2Which idea advances? (Director + Co-PIs)
  • G3aWho do we contact? (Director sends)
  • G3Finalize & submit proposal? (Director only)
  • G4Apply P0 fixes? (Area Lead per workstream)
  • G5Release the report? (Director or Area Lead)
  • G6Publish / external send? (Director only)

Standing principles

  • Evidence-or-silence: no claim without file:line or command + output
  • Never fabricate a citation — flag unverifiable references
  • Single writer per file; no clobbering peers
  • Read-only during review; only Document Office edits final deliverables
  • AI drafts; the human sends, submits, and publishes
  • Every deliverable ships an AI Use Statement

Policy coverage

  • Authorship & AI disclosure (ICMJE/COPE)
  • Research integrity — no fabrication, reproducible runs
  • IRB / ethics gate before human-subjects data work
  • Indigenous data sovereignty (CARE principles)
  • Student-data privacy (FERPA) — no autonomous grading
  • Dual-use & export controls — Integrity Officer flags, work stops

Read the full governance policy →

Proving it — not just claiming it

  • Enforcement A/B harness (evals/enforcement_study/) — measures enforced-gates vs soft-prompt on seeded defects. Result stays Open until real runs land — no premature "beats prompting" claim.
  • Per-funder governance corpus (governance/funders/) — NIH + NSF AI rules mapped to gates, dated and guarded by a hard-failing freshness check. Guidance, not certification.
  • Self-grade A (graded 2026-06-28) — doctor --grade across roster, governance, tooling & tests; 168 tests pass / 1 skipped on every push.

Getting started

Start where you are — three doors

Install once: copy .claude/agents/ into your Claude project, keep the protocol files, and say one sentence. The institute meets you wherever your project is.

📋
Door A
You have a new grant opportunity
Hand the institute an RFP. It reads it, brainstorms ideas with faculty, writes the proposal, and stops for your approval before submission. Gates G1–G3.
read rfp <file/link>
🔬
Door B
You have approved research in progress
Run the full development cycle: generate → verify → synthesize → you approve → revise → validate. Three Opus boards catch what reading misses. Gate G4.
run lab
📊
Door C
You need to report to your sponsor
Reporting-Officer writes a progress or annual report from verified findings. Deck-Builder builds the slides. You release at Gate G5/G6.
progress report

Full detail in GETTING_STARTED.md · 10-minute test drive with bundled demo examples

Install

Two ways to get started

Both paths give you the same 46 agents. Pick the one that fits your workflow.

A

Template repo (simplest)

Click "Use this template" on GitHub, then copy .claude/agents/ into your project. Copy config.example.yml → config.yml and set your lab name.

  • Clone or "Use this template" on GitHub
  • Copy .claude/agents/ into your Claude project
  • Copy config.example.yml → config.yml
  • Open dashboard.html to see the org chart
  • Say new project: <name> to begin
B

Plugin

Install via the Claude plugin marketplace and get auto-updates when the institute ships improvements.

  • Add the marketplace: /plugin marketplace add
  • Install: /plugin install cambium-institute
  • See plugin/README.md for the manifest and options

Plugin manifest: plugin/marketplace.json

🌲 Open Dashboard    ▶ 60-second tour    📄 README

Learn by doing

Cambium Academy

A research institute should also teach. The Academy turns the ideas behind Cambium into something you actually practice, not just read. It is built on how people really learn: you predict an answer before you see it, flip spaced-repetition flashcards, click through the architecture instead of staring at a wall of text, do a small change yourself, and explain each idea back in your own words. Ten interactive modules across three tiers, for students, researchers, and faculty.

The Cambium way
How one sentence becomes verified results.
Evidence tiers
Say exactly as much as the evidence supports.
Human gates
Why stopping to ask a person beats autonomy.
Verify without fooling yourself
Reproduce, then check for leakage and weak baselines.
Ethics & data stewardship
The duties that come before the science.

Every Cambium run also produces a per-run Learning Lab: the same active walkthrough, generated for the exact thing you just built, so you leave understanding it, not just approving it.

🎓 Enter the Academy    ▶ See a Learning Lab