AgenFab Consensus · Decision support

Every expert must critique before it speaks.

A council of 3 to 20 AI expert personas debates your decision, is checked against the evidence turn by turn, and hands you a verdict with dissent, an action plan and a stress test. High-risk verdicts wait for a person to sign off.

Crudfab / AgenFab / Consensus

Ask one model for advice and you get one voice that agrees with how you asked. No dissent, no record of which source backed which claim, nothing to stop a confident guess reaching the report.

How a question becomes a verdict.

01

Input gate

Personal data is masked and leading or manipulative phrasing is flagged before anything reaches a model.

02

Council

Three to twenty expert personas, arranged in one of six debate formats.

03

Critique first

A turn is accepted only if it names a flaw in the argument before it. Flattery does not count.

04

Evidence check

Institutions and numbers that appear in no source are flagged, and a separate judge model scores each turn.

05

Verdict

Answered, degraded, review required or abstained, decided by rules, not by the model grading itself.

Measured agreement

Agreement is measured, not assumed.

After every round the council's positions are compared with each other, and the debate ends when they truly converge or stop moving.

  • The consensus score combines how close the arguments are, how spread out the stances are, how the plan survives the stress test, and how much new ground is still being covered.
  • Six formats: round-robin, cross-fire, moderator-led, opposing poles, answer alone then debate, and an investigation mode that hands the floor to whoever can close an open gap.
  • When the evidence is not there, the system can abstain instead of inventing an answer.
Stress test

A plan the council agrees on still has to survive eight crises.

After synthesis the action plan is run against eight fixed scenarios at once, and the result feeds back into the score.

  • Black swan, regulatory shock, supply-chain shock, cyber attack, loss of key people, macroeconomic crisis, an aggressive competitor and technological disruption.
  • A scenario that runs out of time counts as a failure.
  • A council that converges on a plan that breaks under crisis gets a lower score, so agreement alone cannot close the debate.
Human sign-off

High-risk verdicts go nowhere without a person.

A verdict can be pushed to your issue tracker or other systems, but only under rules you control.

  • A verdict marked review required cannot be sent anywhere until a reviewer records approve or conditional.
  • Sending is off by default and needs both a service switch and an explicit request; every attempt, dry runs included, is logged.
  • When a question carries a classification marker, the session switches to private mode and every public web-search path is closed.
Research agents

Every expert does its own research before it speaks.

The personas are agents. Before a round, each one gathers its own evidence from open sources and from your documents.

  • Open-source research brings in current reporting, records and public data, each item kept with its source.
  • Retrieval over your own documents grounds each persona in the institution's knowledge, not only in the model's memory.
  • Every claim in the debate points back to the evidence card that supports it, so a verdict can be traced line by line.

Built so a decision can be checked later.

01

Decision record

Each verdict carries a manifest: model versions, settings, a hash of the prompt, SHA-256 hashes of every document and web source, code version and seed.

02

What-if branches

Change the budget, deadline, risk or compliance limits and open a new debate linked to the original.

03

Claims tested in code

Agents can write Python and run it in a sandbox with no network and strict memory and time limits, then argue from the real output.

04

Your documents, their sources

Upload PDF, Word, Excel and other files to the whole council or to one expert, alongside 31 research source types, from academic indexes to vulnerability data.

05

Sycophancy, measured

An optional off-the-record channel asks each speaker for a private view and records how far it drifts from what they said in the room.

06

Experts that remember

Library personas keep long-term memory across debates and carry adjustable personality axes.

07

On your servers

Runs on your own Linux server, in containers or on Kubernetes, pointed at your own model by default, with outside models as an option.

08

Reports you can share

Export the verdict as Markdown or as a printable report.

Put your next hard decision in front of the council.

Request a demo
Mobile and WhatsApp
+90 540 278 33 22