Governed AI Co-Scientist

Machine Reasoning,Human Accountability.

Homi is a council of specialist agents, named for the physicist Dr. Homi Jehangir Bhabha. It reads the literature, forms conjectures, designs and analyzes experiments, and drafts results. Every step is logged, evidence-bound, and approval-gated.

Meet Homi's CouncilBack to overview
Runtime Capabilities

Model-Agnostic, Workflow-Native, Governed.

Any model behind an OpenAI-compatible or Model Context Protocol interface can back Homi. Structured-output validation, tool-schema enforcement, evidence requirements, and named-approval gates run inside ArkLab, not inside the model. Agents assist at speed while accountability stays with people.

10 specialized agents

Literature, hypothesis, methodology, experiment design, analysis, and drafting, plus integrity reviewer, citation verifier, replay verifier, and reproducibility inspector.

Authority gradient

Every agent action runs at high-autonomy, suggested-and-reviewed, or approval-gated. Model snapshot, system-prompt hash, sampler, and seed are recorded per step.

Multi-agent debates

Structured debate transcripts with proposer, challenger, and moderator agents, plus named human jurors voting on consensus.

Citation-to-claim verification

Anti-fabrication by default: confirm each citation actually supports the claim, with a confidence score and a support quote.

MCP-style tool registry

Tools register with JSON Schema parameters and outputs, an allowed-agent allowlist, rate limits, and approval-required flags.

Governed memory

Event-sourced project memory with verifiable revert. Rebuild state from any prior point, and never soft-delete.
Specialist Roster

Ten Agents, Each With an Enforced Role.

Every agent carries an enforced authority level, an allowed-tools list, and a versioned system-prompt hash.

Literature Co-ScientistHypothesis Co-ScientistMethodology Co-ScientistExperiment Design Co-ScientistAnalysis Co-ScientistDrafting Co-ScientistIntegrity ReviewerCitation VerifierReplay VerifierReproducibility Inspector
Evaluation

Benchmarks Attached, Not Bolted On.

Every model plugged into Homi's runtime is evaluated against PaperBench (research replication), LAB-Bench (chemistry and biology Q&A), and BioMysteryBench (bioinformatics task solvability), extended with ArkLab citation-grounding accuracy and replay-to-equivalence checks. Results are first-class versioned artifacts.

Homi Runtime

Put a Governed Council on Your Research.

Meet the specialist agents in the workspace, or join the institutional beta to bring them onto your projects.

Meet Homi's CouncilJoin institutional beta