MAGELLAN

Multi-Agent Generative Exploration of Latent Links Across kNowledge

Autonomous AI experiment in cross-disciplinary scientific discovery. Can a multi-agent system autonomously find real scientific connections that humans haven't made yet? This project tests that question.

No domain expertise required from the user. Type /discover, walk away. Built entirely on Claude Code — 15 agents, 7 slash commands, deterministic hooks, and 2 MCP servers, all configured via .claude/ and executed through Claude Code's Agent tool. Come back to find testable hypothesis cards, then validate cross-model. Optimized for cross-disciplinary life sciences discovery (retrieval tools, scoring weights, hypothesis format); other domains are supported but scores reflect infrastructure asymmetry, not hypothesis quality. Impact-aware prioritization steers the pipeline toward high-impact directions via tiebreakers, decomposed scoring, and meta-learning — without sacrificing novelty or rigor.

Prerequisites

1. Claude Code: Anthropic's terminal-based AI tool (the CLI version, not the web chat or desktop app). Requires v2.1.257 or newer, the version where claude-fable-5-1 (released September 1 2026) becomes the default Fable model; every agent pins that full model ID in its frontmatter, so older builds fail loudly instead of silently running the wrong model. Also needs canonical hook output, terminalSequence notifications, and hook exec-form. Run claude --version and update if /model does not offer Fable 5.1. Requires a Claude subscription (Pro minimum, Max/Team recommended). Install docs

   Before running /discover, set the session model with /model claude-fable-5-1. The 14 sub-agents pin that model ID in their own frontmatter, but the orchestrator runs on the session model, so leaving the session on another model gives you a weaker model coordinating Fable 5.1 sub-agents.

   Zero data retention orgs cannot run this pipeline as configured. Fable 5.1 is a Covered Model requiring 30-day retention; under ZDR every call returns 400 invalid_request_error unless Anthropic has expressly authorized it. Roll the fleet back to claude-opus-5, which is available under ZDR.

2. Node.js 20+ — Required for cross-model validation scripts and website upload. Run npm install after cloning to install dependencies.

3. API keys for cross-model validation (optional) — For GPT-5.5 Pro and Gemini Deep Research Max independent review. Create a .env.local file in the project root:

   OPENAI_API_KEY=sk-...
   GEMINI_API_KEY=AI...
   

   Get your keys: OpenAI · Google AI Studio Without these, the pipeline generates export files for manual copy-paste validation. Note: Gemini Deep Research Max is paid-tier only (not available on the free Gemini API tier). Typical per-session Gemini-side cost ~$4.80.

Quick Start

git clone https://github.com/kakashi-ventures/magellan-cli.git
cd magellan-cli
npm install
claude --enable-auto-mode

Primary mode (fully autonomous — the whole point)

/discover

Alternative modes (for targeted testing)

/discover circadian biology × tumor immune evasion
/discover solve: antibiotic resistance

What Happens When You Run /discover

Phase 0a: Scout generates 5-6 candidates (10 strategies + bridge concepts
            + strategy diversification + exploration slot
            + rotating creativity constraint)
Phase 0b: Literature Scout verifies disjointness for ALL candidates
            (domain-aware retrieval + bridge validation)
Phase 0c: Orchestrator narrows to 3 (DISJOINT priority + strategy diversity)
Phase 0d: Target Evaluator — adversarial challenge on 4 axes
Phase 1:  Orchestrator selects best target
          (hard constraint: DISJOINT preferred over PARTIALLY_EXPLORED)
          → Computational Validator — KEGG, STRING, PubMed co-occurrence,
            back-of-envelope physics checks
Phase 2:  Generator builds Structured Relationship Map, then creates
          6-8 hypotheses (parametric + literature + computational validation
            + bisociation + multi-level abstraction)
          → SELF-CRITIQUE + claim-level verification before output
Phase 3:  Critic attacks all hypotheses (9 vectors + web search)
          → META-CRITIQUE reflection + critic_questions for Generator
          → Groundedness reinforcement (if majority LOW/SPECULATIVE)
Phase 4:  Ranker scores on 6 dimensions + diversity check + Elo sanity check
          → ADAPTIVE CYCLE DECISION: early-complete | standard | extended
Phase 5:  Evolver recombines top candidates (conditionally skippable)
          ── Cycle 2: Phases 2-5 repeat with evolved + fresh hypotheses ──
          ── Critic questions forwarded to Generator in cycle 2 ──
          ── Conditional: skip Evolver if cycle 2 top-3 >= 6.5 ──
Phase 6:  Quality Gate — 10-point rubric + web grounding + per-claim verification
          → META-VALIDATION reflection before output
          → Session Analyst — meta-learning metrics → knowledge/meta-insights.md
Phase 7:  Cross-Model Validation — GPT-5.5 Pro (background poll, web search + code interpreter) +
          Gemini Deep Research Max (Interactions API agent: google_search + url_context
          + code_execution, ~80-160 searches/task) → consensus report
          (automatic if API keys set, export files only otherwise)
Phase 7b: Convergence Scanner — ClinicalTrials.gov, NIH Reporter, patents (non-blocking)
          Dataset Evidence Miner — HPA, GWAS, ChEMBL, UniProt, PDB queries + follow-up suggestions (non-blocking)
Phase 8:  Deliverables verification → check all JSON + markdown artifacts exist
          Post-QG Amendments → arithmetic/citation corrections appended to final-hypotheses.md
          Session summary (written AFTER post-QG agents + verification) → results/{session-id}/
Phase 9:  Knowledge persistence → knowledge/discovery-log.json + knowledge/meta-insights.md

Typical runtime: 20-55 minutes. Check progress with /status.

After Discovery: Cross-Model Validation

Automatic: If OPENAI_API_KEY and/or GEMINI_API_KEY are set in .env.local, the pipeline automatically calls GPT-5.5 Pro (background submit + 30s poll, with web search + code interpreter) and Gemini Deep Research Max (agent deep-research-max-preview-04-2026 on the Interactions API, with google_search + url_context + code_execution, running an autonomous research loop of ~80-160 web searches + URL reads + code execution) for independent validation and generates a consensus report.

# Create .env.local in the project root (one-time setup)
OPENAI_API_KEY=sk-...
GEMINI_API_KEY=AI...

Get your keys: OpenAI · Google AI Studio

Coming soon: OpenAI Codex CLI support as an alternative runtime to Claude Code.

Manual fallback: If no API keys are set, export files are generated:

/export gpt       → produces results/{session-id}/export-gpt.md
/export gemini    → produces results/{session-id}/export-gemini.md

Copy into ChatGPT / Gemini AI Studio. See prompts/orchestration-guide.md for step-by-step instructions.

Commands

Command	What it does
/discover	Full autonomous (Scout finds targets)
/discover [A] × [C]	Targeted discovery between two fields
/discover [topic]	Open exploration from one domain
/discover solve: [problem]	Problem-driven discovery
/discover --context "text"	Provide domain expertise as context for Scout/Generator
/discover --papers DOI1,DOI2	Provide seed papers for Literature Scout
/discover --interactive	Pause after Scout for target approval before proceeding
/connect <key>	Link CLI to your MAGELLAN web profile for attribution
/validate [hypothesis]	Deep novelty + counter-evidence check
/evolve	Another evolutionary cycle on current results
/export gpt	Self-contained prompt for GPT-5.5 Pro validation
/export gemini	Self-contained prompt for Gemini Deep Think
/status	Check pipeline progress mid-run
/validate-holdout	Run holdout validation test (rediscovery check)

Flags can be combined: /discover ferroptosis × serpentinization --context "I study lipid peroxidation in hepatocytes" --papers 10.1038/s41586-024-xxxxx --interactive

Contribute Your Discoveries

Connect to the MAGELLAN website so your discoveries are publicly attributed to you:

1. Create an account at magellan-discover.ai/sign-in
2. Go to your profile and generate a Contributor Key
3. In the CLI: /connect mgln_your_key_here
4. All subsequent /discover sessions are attributed to your profile

Your discoveries appear on the public leaderboard with your name, institution, and rank (Cabin Boy → Navigator → Cartographer → Captain → Admiral).

Scientists with domain expertise can use --context and --papers to direct discovery toward their areas of interest, and review hypotheses on the website as expert reviewers.

Project Structure

CLAUDE.md                                    ← Project instructions for Claude Code
.mcp.json                                   ← MCP servers (Semantic Scholar, PubMed)
.claude/
  settings.json                              ← Permissions, hooks, Agent Teams
  agents/
    discovery-orchestrator.md                 ← Dispatches to agents, guard logic [Fable 5.1, max]
    scout.md                                 ← Finds WHERE (10 strategies) [Fable 5.1, max]
    target-evaluator.md                      ← Adversarial target challenge [Fable 5.1, max]
    literature-scout.md                      ← Retrieves literature context [Fable 5.1, high]
    computational-validator.md               ← Programmatic bridge checks [Fable 5.1, high]
    generator.md                             ← Creates hypotheses [Fable 5.1, max]
    critic.md                                ← Attacks hypotheses (9 attack vectors) [Fable 5.1, max]
    ranker.md                                ← 6-dimension scoring + Elo sanity check [Fable 5.1, high]
    evolver.md                               ← Recombines with diversity constraint [Fable 5.1, high]
    quality-gate.md                          ← 10-point rubric + web grounding [Fable 5.1, max]
    session-analyst.md                       ← Post-pipeline meta-learning [Fable 5.1, high]
  commands/
    discover.md                              ← /discover (main entry point)
    connect.md                               ← /connect contributor key
    validate.md                              ← /validate deep check
    evolve.md                                ← /evolve refinement cycle
    export.md                                ← /export for cross-model validation
    status.md                                ← /status pipeline progress
  skills/
    discovery-engine/SKILL.md                ← Core methodology + hypothesis card format
    hypothesis-validation/SKILL.md           ← Validation protocol + groundedness
    literature-retrieval/SKILL.md            ← Search patterns for scientific DBs
    domain-life-sciences/SKILL.md            ← Bio/med domain knowledge
    domain-physics-math/SKILL.md             ← Physics/math domain knowledge
prompts/
  validation-prompt-gpt.md                    ← GPT-5.5 Pro validation prompt
  validation-prompt-gemini.md                 ← Gemini mathematical structure prompt
  orchestration-guide.md                     ← Cross-model validation step-by-step
  session-summary-format.md                  ← Session summary formatting per status type
  ingest-schema.json                         ← Schema for website ingest manifest
  knowledge-schema.json                      ← Schema for discovery-log entries
docs/
  methodology-v5.md                          ← Full methodology with evidence
  CHANGELOG.md                               ← Pipeline evolution history
scripts/                                     ← Hook scripts + orchestrator support scripts
  init-session.sh                            ← Session initialization (state + results dir)
  upload-session.mjs                         ← Website upload (ingest → API)
state/                                       ← Coordination state (machine-readable)
  session.json                               ← Slim coordination index (~3KB)
  dispatch-log.json                          ← Agent dispatch log with timestamps
results/                                     ← All session outputs (markdown + JSON)
  {session-id}/                              ← Session-scoped directory
    papers/                                  ← Full-text papers retrieved by Literature Scout
    scout-targets.md                         ← Scout output
    scout.json                               ← Scout targets + quality scores (structured)
    target-evaluation.md                     ← Target Evaluator output
    target-evaluation.json                   ← Axis scores + recommendations (structured)
    literature-landscape.md                  ← Literature Scout output (all modes)
    literature.json                          ← Literature context + paper metadata
    computational-validation.md              ← Computational Validator output
    computational.json                       ← Computational readiness checks
    raw-hypotheses-cycle{N}.md               ← Generation output
    cycle{N}-raw.json                        ← Raw hypotheses (IDs, titles, scores)
    critiqued-cycle{N}.md                    ← Critique output
    cycle{N}-critiqued.json                  ← Critique verdicts + critic_questions
    ranked-cycle{N}.md                       ← Ranking output
    cycle{N}-ranked.json                     ← Rankings + composite scores
    evolved-cycle{N}.md                      ← Evolution output
    cycle{N}-evolved.json                    ← Evolved hypotheses with lineage
    quality-gate.md                          ← Quality Gate rubric
    quality-gate.json                        ← Quality gate verdicts + session_status
    final-hypotheses.md                      ← Final hypothesis cards
    final.json                               ← PASS/CONDITIONAL_PASS only (created by orchestrator from quality-gate.json)
    session-analysis.md                      ← Session Analyst output
    meta-insights.json                       ← Session analyst structured output
    export-gpt.md                            ← GPT validation prompt
    export-gemini.md                         ← Gemini validation prompt
    validation-gpt.md                        ← GPT-5.5 Pro response (if API key set)
    validation-gemini.md                     ← Gemini Deep Research Max report (if API key set)
    cross-model-consensus.md                 ← Consensus report (if any API key set)
    cross-model.json                         ← Cross-model validation consensus
    convergence-report.md                    ← Convergence Scanner output
    convergence.json                         ← Trials, grants, patents, partial confirmations
    dataset-evidence-report.md               ← Dataset Evidence Miner output
    dataset-evidence.json                    ← Per-claim database verification results
    session-summary.md                       ← Session overview
knowledge/                                   ← Persistent data across sessions
  discovery-log.json                         ← Explored pairs, productive bridges, kill reasons
  meta-insights.md                           ← Cumulative meta-learning insights
validation/
  holdout-discoveries.json                   ← Curated post-cutoff holdout targets
  results/                                   ← Holdout and retrospective validation reports
verification/
  {slug}/                                    ← One directory per computational verification
    manifest.json                            ← Links verification to website DB (required)
    analyze_{slug}.py                        ← Reproducible analysis script
    results/
      {SLUG}_REPORT.md                       ← Full verification report (markdown)
      fig1_*.png ... figN_*.png              ← Figures (PNG for web)

Architecture

15 specialized agents, all running on Claude Fable 5.1 (claude-fable-5-1, released September 1 2026), Anthropic's most capable widely released model. The deep/structured distinction is carried by effort, not model tier: effort max for deep reasoning agents, high for structured/search agents, pinned per agent to guarantee quality regardless of the user's session-level effort setting. Fable 5.1 costs $10/$50 per MTok, double Opus 5, and is the tier Anthropic points to for demanding reasoning and long-horizon agentic work. Its safety classifiers cover a broader set than Opus 5's cybersecurity-only ones, including biology, so each agent has a documented rollback to claude-opus-5 and the orchestrator auto-falls-back on a decline; see docs/CHANGELOG.md and CLAUDE.md:

• Scout [Fable 5.1, max] — 10 strategies to find WHERE undiscovered connections hide (incl. structural isomorphism + serendipity). TARGET QUALITY CHECK + strategy diversification + exploration slot + rotating creativity constraint
• Target Evaluator [Fable 5.1, max] — Adversarial challenge of Scout targets on 4 axes (popularity, vagueness, impossibility, local-optima)
• Literature Scout [Fable 5.1, high] — MCP servers (mandatory first step) + WebSearch fallback + full-text paper retrieval + RETRIEVAL QUALITY CHECK reflection
• Computational Validator [Fable 5.1 + Bash, high] — Programmatic bridge verification: KEGG, STRING, PubMed co-occurrence, back-of-envelope physics
• Generator [Fable 5.1, max] — Parametric creativity + literature + computational validation → 6-8 hypotheses per cycle. SELF-CRITIQUE + claim-level verification reflection
• Critic [Fable 5.1, max] — 9 adversarial attack vectors (incl. claim-level fact verification) + META-CRITIQUE reflection + critic_questions feedback
• Ranker [Fable 5.1, high] — 6-dimension scoring (Impact decomposed into Paradigm 5% + Translational 5%) + diversity check + Elo tournament sanity check
• Evolver [Fable 5.1, high] — Crossover, mutation, specification with diversity constraint + EVOLUTION QUALITY CHECK reflection (conditionally skippable)
• Quality Gate [Fable 5.1, max] — 10-point rubric + web novelty + per-claim grounding verification + impact annotation (informational) + META-VALIDATION reflection
• Session Analyst [Fable 5.1, high] — Post-pipeline meta-learning: strategy performance, kill patterns, bridge type analysis → knowledge/meta-insights.md
• Cross-Model Validator [Fable 5.1, high] — Calls GPT-5.5 Pro (background poll, reasoning xhigh, web search + code interpreter) + Gemini Deep Research Max (Interactions API agent: google_search + url_context + code_execution) for independent validation → consensus report (requires API keys; falls back to export files)
• Convergence Scanner [Fable 5.1, high] — Post-QG: searches ClinicalTrials.gov, NIH Reporter, patents for independent convergence signals + partial mechanism confirmations from non-pipeline sources
• Dataset Evidence Miner [Fable 5.1, high] — Post-QG: queries HPA, GWAS Catalog, ChEMBL, UniProt, PDB via scripts/query-biodata.py to verify specific molecular claims in passing hypotheses. Suggests computational follow-up queries
• Holdout Evaluator [Fable 5.1, max] — Validation framework: compares MAGELLAN output against known post-cutoff discoveries with contamination check + mechanism similarity scoring

Conceptual Foundation

MAGELLAN operationalizes Don Swanson's Undiscovered Public Knowledge (1986): logically connected knowledge that exists across disjoint scientific literatures but hasn't been linked because researchers don't read across fields. Swanson's method was bibliometric (citation graphs, MeSH co-occurrences). MAGELLAN replaces the method with frontier LLMs — models trained on both literatures already have the connections latent in their parameters. The challenge shifts from detecting citation disjointness to eliciting cross-domain connections through structured multi-agent reasoning.

State of the Art (March 2026)

MAGELLAN sits in a sparsely populated niche: fully autonomous target selection. Most comparable systems (Google AI Co-Scientist, FutureHouse Kosmos, SciAgents) require human-specified research objectives.

Key validations from the field:

• MOOSE-Chem (ICLR 2025): LLMs encode "latent scientific knowledge associations not yet recognized by humans" — direct validation of MAGELLAN's UPK thesis
• FrontierScience Benchmark: 52-point gap between structured (77%) and open-ended research (25%) tasks validates multi-agent approach
• Google AI Co-Scientist: 3 experimentally validated discoveries using parallel architecture (Generate/Reflect/Rank/Evolve)

See docs/methodology-v5.md for full comparison with state-of-the-art systems, evidence-based design rationale, and risk analysis. See docs/CHANGELOG.md for the evolution history of the pipeline.

License

Software: Apache License 2.0 — Copyright 2026 Alberto Trivero / Kakashi Venture Accelerator

Discovery outputs: Dual-track licensing based on mode of operation:

• Autonomous (/discover pure) → CC0 1.0 (public domain, voluntary citation requested)
• Guided (/discover A × B, --context, --papers, --interactive) → CC-BY 4.0 (attribution required)

See DISCOVERY_LICENSE.md for full details on output licensing and attribution.

Built where science meets AI. kakashi.ventures