BrennerBot Lab Mode

Research Orchestration

BrennerBot: Three AI Minds. One Rigorous Method. Zero Blind Spots.

Coordinate Claude, GPT, and Gemini in structured debates via NTM and Agent Mail. Run 10-phase falsification loops with the /brennerbot-with-ntm skill that produce hypothesis slates, discriminative tests, and auditable evidence trails.

1. CLI Quick Install

curl -fsSL https://brennerbot.org/install.sh | bash

2. Skill + NTM Swarm (jsm)

jsm install brennerbot-with-ntm ntm vibing-with-ntm agent-mail

Get Started Explore /brennerbot-with-ntm

Read the Method View Distillations

Locked predictionsNTM Swarm PipelinesAuditable artifacts

NTM Agent Constellation10-phase loop

Proposer & InvestigatorClaude + Codex

Devil's-Advocate CriticGemini + Codex

Synthesizer & AdjudicatorClaude + AGY

Triangulated Artifacts

by_cc.md + by_cod.md + by_agy.md → disagreement_register.md → meta_synthesis.md

ntm pipeline run brennerbot-squad.yaml → HANDBACK.md

NEW: /brennerbot-with-ntm Skilljeffreys-skills.md (jsm)Dicklesworthstone/ntm

Native Swarm Falsification: /brennerbot-with-ntm

Operationalize Sydney Brenner's epistemology directly inside Claude Code, Codex, Gemini, and Antigravity panes orchestrated by NTM (Named Tmux Manager). Distributed via jeffreys-skills.md (jsm), this skill turns a multi-agent terminal cockpit into a disciplined machine for deleting hypothesis space cheaply.

The One Rule ObjectiveScore ≥ 2.0

Maximize expected mind-change per unit of experimental cost:

Value(T) = (MindChange × OptionValue) / (Time × Cost × Ambiguity × Infra)

Mandatory Phases

PREFLIGHT → DRIFT

Operating Modes

Including Incident Mode

15+1

Cognitive Operators

◊ ⊘ 𝓛 ≡ ✂ ⟂ ↑ ⌂ 🔧 ⊞ 🤝 ΔE † ∿ ⊙ ⤴

Marching Orders

Parameterized MO-*.md

Helper Scripts

Zero-context validators

NTM Pipelines

Executable YAML workflows

Step 1 • Install Skill Pack via jeffreys-skills.md

Installs /brennerbot-with-ntm + 36 Marching Orders + 48 scripts + 6 NTM pipelines

# 1. Install the jsm CLI (jeffreys-skills.md)
curl -fsSL https://jeffreys-skills.md/install.sh | bash
jsm login

# 2. Install /brennerbot-with-ntm + companion flywheel skills
jsm install brennerbot-with-ntm ntm vibing-with-ntm agent-mail \
  beads-br beads-bv cass cass-memory multi-model-triangulation

Step 2 • Install NTM & Run Preflight Doctor

Named Tmux Manager (Go binary) orchestrates multi-agent panes & pipelines

# Install NTM (Named Tmux Manager - Go multi-agent control plane)
curl -fsSL https://raw.githubusercontent.com/Dicklesworthstone/ntm/main/install.sh | bash

# Verify NTM + BrennerBot skill doctor
ntm --version
~/.claude/skills/brennerbot-with-ntm/scripts/brennerbot-doctor.sh

Execution Control Plane

How NTM Powers Multi-Agent Brenner Loops

Rather than fragile wrapper APIs, NTM orchestrates real CLI coding agents in persistent named tmux panes with a deterministic robot control loop:

ntm spawn <sess> --cc=2 --cod=1 --gmi=1 --agy=1

Launches persistent panes across Claude Code, Codex, Gemini, Antigravity, Grok, and Oh My Pi.

ntm pipeline run brennerbot-squad.yaml

Executes multi-stage YAML workflows using command:, template:, foreach:, foreach_pane:, and branch:.

ntm --robot-snapshot / --robot-attention

Zero-polling state inspection: surfaces only panes needing operator attention or phase adjudication.

ntm --robot-causality & checkpoint save

Tracks message-to-bead provenance and freezes full session state for zero-loss resume.

Works alongside brenner CLI, br, bv, and Agent Mail.

Open Multi-Agent Cockpit Guide

Ready to Apply the Method?

Three learning paths from quick start to multi-agent orchestration. Apply Brenner's scientific method to your own research questions.

~30 min Quick StartAgent-AssistedMulti-Agent Cockpit

Start Tutorial

What's Inside

A research toolkit for applying Brenner's epistemology to your own scientific questions.

Browse

Corpus

The complete Brenner transcript collection from Web of Stories, plus curated and .

Learn more

Compare

Distillations

Three frontier model of Brenner's methodology. Compare perspectives from GPT-5.2, Opus 4.5, and Gemini 3.

Learn more

Learn

Method

The , loop structure, and framework that operationalize Brenner's approach to scientific discovery.

Learn more

Core Workflow

From Question to Conclusion: The Brenner Loop

Research sessions follow a rigorous, reproducible path. Every step is tracked, auditable, and reversible.

1. Intake

Frame the research question

2. Sharpening

Refine hypotheses and scope

3. Level-Split

Separate program from interpreter

4. Exclusion-Test

Design discriminative tests

5. Object-Transpose

Choose the optimal system

6. Scale-Check

Validate against physics

7. Agent-Dispatch

Convene the tribunal

8. Synthesis

Merge agent outputs

9. Evidence

Gather external signals

10. Revision

Update hypotheses

11. Complete

Publish artifacts

Phase

Intake

Frame the research question

Phase

Sharpening

Refine hypotheses and scope

Phase

Level-Split

Separate program from interpreter

Phase

Exclusion-Test

Design discriminative tests

Phase

Object-Transpose

Choose the optimal system

Phase

Scale-Check

Validate against physics

Phase

Agent-Dispatch

Convene the tribunal

Phase

Synthesis

Merge agent outputs

Phase

Evidence

Gather external signals

Phase

Revision

Update hypotheses

Phase

Complete

Publish artifacts

Undo / Redo

Every action is reversible. Explore without fear.

Session Replay

Reproduce any session exactly for audit and learning.

Error Recovery

Graceful checkpoints when things go wrong.

Research Lab

Apply the Method to Your Research

The is an interactive research framework that helps you develop using four cognitive operators. Run structured sessions and track your evolving understanding.

⊘Level-Split

✂Exclusion-Test

⟂Object-Transpose

⊞Scale-Check

Explore Sessions View Operators Start Tutorial

Multi-Agent Orchestration

Your Research Team: AI Agents That Debate, Challenge, and Synthesize

Each agent has a precise mandate. Together they sharpen hypotheses, design lethal tests, and merge evidence into auditable artifacts - without surrendering control.

"What if you could have Claude, GPT, and Gemini debate your hypothesis - challenging each other until only the strongest ideas survive?"

Hypothesis Generator

Hunt paradoxes, propose hypotheses

Creative, divergent thinking

Signature stance

"What if both established models are wrong?"

Test Designer

Design discriminative tests with potency controls

Rigorous, detail-oriented

Signature stance

"This test will eliminate half our hypotheses in one observation."

Adversarial Critic

Attack framing, check scale constraints

Skeptical, thorough

Signature stance

"Have you considered that the entire premise might be wrong?"

Oxford Style

Debate Mode

Proposition vs opposition with a judge

Best for: Testing hypothesis strength

Socratic

Debate Mode

Probing questions to surface hidden assumptions

Best for: Finding weak links fast

Steelman Contest

Debate Mode

Build the strongest case, then dismantle it

Best for: Exploring the hypothesis space

Terminal

brenner-cli

# Start a debate session
brenner session start --thread-id RS-20260105 \
  --format oxford \
  --question "Does the morphogen gradient model explain cell fate?"

# Watch agents debate in real-time
brenner session status --thread-id RS-20260105 --watch

# See the merged artifact
brenner session compile --thread-id RS-20260105

Coordination Visualization

Deterministic Merge

Thread ID: RS-20260106-001Ack tracking enabled

Kickoff

Threaded prompt goes to each agent role

Deltas

Structured responses return with citations

Merge

Deterministic compiler reconciles evidence

Human

You decide what ships and what dies

Coordination Without Chaos

Agent Mail keeps every exchange auditable

Every message lands in a thread, every response is acknowledged, and every delta is preserved. You stay in the loop with human approval gates at every step.

Built on with thread IDs, ack receipts, and merge-safe deltas.

Kickoff sent3 agents live

Deltas merged1 artifact ready

Human approvalRequired

Research Hygiene

Built-In Guardrails for Rigorous Science

Catch mistakes before they become problems. Lock predictions to prevent hindsight bias. Track calibration over time. Detect confounds automatically.

Learn the Method While You Work

Coach Mode

Progressive scaffolding based on experience level. Inline explanations with Brenner quotes catch mistakes before they compound.

Three levels: beginner, intermediate, advanced
Auto-promotes based on tracked progress
Specific feedback, not generic warnings

Prevent Hindsight Bias

Prediction Lock

Predictions must be locked before test results arrive. You cannot 'predict' outcomes you already know.

Cryptographic commitment timestamps
Cannot edit after lock engages
Calibration tracks locked predictions only

Know Your Accuracy

Calibration Tracking

Brier scores for probabilistic accuracy. Overconfidence bias detection keeps you honest over time.

Domain-specific calibration curves
Overconfidence alerts when warranted
Historical accuracy visualization

See What You Might Miss

Confound Detection

Eight research domains with domain-specific confounds. Automatic detection based on hypothesis text triggers prompting questions.

Selection bias, reverse causation, omitted variables
Domain-aware prompting questions
Links to methodological literature

Domain-Specific Confound Detection

Each domain has unique methodological pitfalls

8 domains

PsychologyEpidemiologyEconomicsBiologySociologyNeuroscienceCSGeneral

Example: "Selection bias detected in hypothesis H2 — how will you ensure random sampling?"

Artifact Linting

50+ rules for research hygiene

Every hypothesis, test, and evidence item passes through automated hygiene checks. Structural integrity, citation resolution, and reasoning validity—all verified before you ship.

PassWarningError

Linter Categories52 rules

Hypothesis Hygiene

Third alternative required • Unfalsifiable language flagged

Test Design

Potency control required • Sample size power analysis

Evidence Chain

Citation anchors must resolve • Source credibility check

Structural

Missing fields detected • Orphan references flagged

Reasoning

Circular logic detection • Base rate neglect warning

Without Guardrails vs. With BrennerBot

Without Guardrails

Predictions adjusted after seeing results
Confounds discovered in peer review
Vague hypotheses survive unchallenged
Overconfidence goes unchecked

With BrennerBot

Predictions locked before execution
Confounds flagged during design
Unfalsifiable language caught early
Calibration tracked and displayed

Discovery & Intelligence

Intelligence Built In: Search, Simulate, Score

Connect to prior work instantly, model evidence impact before you test, and track which hypotheses survive pressure. This is research intelligence, not a chat log.

Hypothesis Similarity Search

Find related work across sessions with offline embeddings and clusters.

Client-side onlyDuplicate detection

What-If Scenarios

Simulate outcomes before running tests and prioritize high-impact experiments.

Info gain rankedScenario builder

Robustness Scoring

Evidence-weighted survival scores reveal fragile vs battle-tested ideas.

Support vs challengeRobustness meter

Anomaly Detection

Track contradictions and spawn new hypotheses instead of burying them.

Anomaly registerParadigm alerts

Similarity Search

Offline

Query: "morphogen gradient cell fate"

Morphogen gradient (RS-20251230)82%

Statement 0.8 / Mechanism 0.6 / Domain 0.9

Timing gate model (RS-20250112)71%

Statement 0.7 / Mechanism 0.5 / Domain 0.8

Signal relay chain (RS-20241018)64%

Statement 0.6 / Mechanism 0.4 / Domain 0.9

Runs entirely client-side - your hypotheses never leave your machine.

What-If Scenario

Info gain

Starting confidence60%

If supports78%

If challenges35%

Expected information gain: 0.42

Best next test: Perturb gradient + checkpoint timing

Robustness

Survival score

H1: Morphogen gradient72%

3 supporting / 1 challenging (survived)

H2: Timing mechanism35%

1 supporting / 2 inconclusive

Anomaly Register

Quarantine

X-001Active

Oscillating fate markers

Conflicts with H1 + H2

X-014Deferred

Late-stage inversion

Waiting on potency control

Deep Dive

The Operator Algebra: Brenner's Methods as Executable Code

Sydney Brenner's breakthrough wasn't just his discoveries - it was his method. We've encoded his cognitive patterns as composable operators that you can apply systematically.

The Brenner Method in 4 Steps

Split the levels

Separate the 'what' from the 'how'

Design killing tests

Find experiments that eliminate possibilities

Choose your system

Pick the easiest organism/model to test with

Check the physics

Make sure it's physically possible

Want the precise notation? See the operators below.

⊘

Level-Split

"Separate program from interpreter"

Message vs machine, genotype vs phenotype. Includes the 'chastity vs impotence' diagnostic.

Template

"What is the information? What is the mechanism?"

✂

Exclusion-Test

"Design tests that eliminate, not confirm"

Forbidden patterns: what cannot occur if H is true. Rated by discriminative power.

Template

"If H1 is true, we should NEVER see..."

⟂

Object-Transpose

"Change the system until the test is easy"

Choose organism or model strategically. The experimental object is a design variable.

Template

"What system would make this test cheap and unambiguous?"

⊞

Scale-Check

"Stay imprisoned in physics"

Validate against physical constraints. Calculate timescales, length scales, energy scales.

Template

"Is this physically possible at the relevant scale?"

The Core Composition

(⌂ ∘ ✂ ∘ ≡ ∘ ⊘) powered by (↑ ∘ ⟂ ∘ 🔧) constrained by (⊞) kept honest by (ΔE ∘ †)

- Start from a paradox (◊), split levels (⊘), extract invariants (≡)

- Design exclusion tests (✂), materialize as decision procedure (⌂)

- Power by amplification (↑) in well-chosen system (⟂) you build yourself (🔧)

- Constrain by physics (⊞), keep honest with exception handling (ΔE) and theory killing (†)

Extended Operators12 more cognitive & epistemic patterns

◊

Paradox-Hunt

Use contradictions and anomalies as beacons

𝓛

Recode

Change representation to expose hidden invariants

≡

Invariant-Extract

Find what survives across transformations

↑

Amplify

Use selection, dominance, and regime switches

⌂

Materialize

Compile theory into a concrete decision procedure

🔧

DIY-Bricolage

Build the minimal instrument yourself now

🤝

Adversarial-Collab

Cross-family debate before consensus

ΔE

Exception-Quarantine

Isolate anomalies without ad-hoc patching

†

Theory-Kill

Drop hypotheses immediately when falsified

∿

Dephase

Work out of phase with prevailing fashion

⊙

Self-Reflect

Audit methodology drift and cognitive bias

⤴

Quickie-Pilot

Run a cheap 10-minute probe before full test

TypeScript

brenner-loop/operators

import { pipe } from "@/lib/brenner-loop/operators/framework";

const brennerPipeline = pipe(
  levelSplit,        // Separate levels
  invariantExtract,  // Find what survives
  exclusionTest,     // Design killing experiments
  materialize,       // Compile to decision procedure
);

const result = brennerPipeline(hypothesis, context);

Technical Architecture

Built for Serious Work: Architecture That Respects Your Research

CLI-first design. Local-first storage. Deterministic outputs. Every architectural decision prioritizes reproducibility, control, and auditability.

Your Terminal, Your Subscriptions, Your Control

CLI-First Design

BrennerBot doesn't call AI APIs from code. CLI tools run in your terminal under your existing subscriptions. You see everything, control everything.

No API keys to manage or leakNo rate limits beyond your subscriptionFull session context preservedReal-time visibility into agent work

Your Data Stays With You

Local-First Storage

Hypotheses, evidence, and session artifacts are stored locally. The system works offline and syncs when you're ready.

IndexedDB for structured dataFile locks for concurrent accessBackground sync when onlineFull offline capability

Reproducible by Design

Deterministic Merging

When multiple agents produce outputs, they merge deterministically. Two runs with the same inputs produce identical outputs.

Timestamp orderingLast-write-wins for conflictsKilled items preserved in historyConflicts logged for audit

Fail-Closed, Not Fail-Open

Security Model

Lab Mode is protected by default. Unauthorized requests get 404 (no information leakage), not 401.

Environment check (BRENNER_LAB_MODE)Cloudflare Access or lab secretTiming-safe secret comparisonContent policy enforcement

The Join-Key Contract

Thread ID is the universal join key

Thread IDAgent Mail thread

ntm session

Artifact file path

Beads issue ID

brenner-cli

# Search the corpus

brenner corpus search "exclusion test"

# Start a session

brenner session start --thread-id RS-20260106 \ --question "How do cells determine position?"

# Watch session status

brenner session status --thread-id RS-20260106 --watch

# Compile artifacts

brenner session compile --thread-id RS-20260106

Performance

CLI startup

< 50ms

Artifact compile

< 50ms

Test suite

4300+ tests

Page load

< 2s

Key Principle

"Same inputs, same outputs. Every merge is deterministic. Every action is auditable. Every session is reproducible."

Interview Segments

Model Distillations

Operator Types

0k+

Words of Wisdom

“
I think many fields of science could do a great deal better if they went back to the classical approach of studying a problem, rather than following the latest fashion.
Sydney Brenner-Nobel Laureate in Physiology or Medicine, 2002

Get Started