🗺️ Roadmap

Active development continues. Near-term priorities:

• 🔬 Search pipeline overhaul — major refactor of dare-web and dare-scholar search flows to better integrate with AlphaXiv MCP and upcoming Perplexity MCP, reducing redundancy and improving retrieval quality
• 🧬 Method-evolve full loop (v3.2+) — AlphaEvolve-inspired evolutionary improvement of DARE's own research methods. Core tools (mutate, crossover, evaluate) implemented; full autonomous Elo-tournament loop next
• 📝 Paper-writing implementation (v3.1+) — automated academic paper composition from research outputs. Strategy interface defined, implementation pending
• 🔗 GitHub MCP integration — native GitHub MCP adapter for issue tracking, PR-driven experiment workflows, and automated result reporting
• 🌐 Perplexity MCP adapter — leverage Perplexity's search-augmented generation as an additional web research backend

🎯 Design Philosophy

🤔 Why "De-Anthropocentric"?

The bottleneck in modern research is not data or compute — it's the human in the loop. Every existing "AI research assistant" still requires a human to decide what to search, what to read, which gaps matter, and which ideas are worth pursuing. DARE removes this bottleneck entirely. The human provides only the initial direction; everything after that is autonomous.

This isn't about replacing researchers — it's about creating a parallel research capacity that operates on timescales and breadths impossible for any individual.

🎖️ The Military Metaphor: Four-Layer Command Structure

DARE's architecture follows a military command hierarchy — not because research is war, but because the decomposition pattern is remarkably effective:

General (Meta-Strategy)  →  "Take that hill"         →  WHAT to research
Colonel (Strategy)       →  "Flank from the east"    →  WHEN and WHY
Captain (Tactic)         →  "Squad A cover, B move"  →  HOW to combine
Sergeant (SOP)           →  "Fire, reload, advance"  →  HOW to execute

Each layer has a single concern and calls only the layer directly below it. A Strategy never touches MCP tools directly; a Tactic never decides research direction. This strict layering means every component is independently testable, replaceable, and composable.

🤖 Micro-Agent Paradigm: Every Tool Thinks

Traditional MCP tools are dumb functions — they take input, return output, no reasoning involved. DARE's dare-agents tools are fundamentally different. Each of the 49 tools is a single-responsibility LLM micro-agent with its own system prompt, personality, and reasoning chain.

When DARE runs "root-cause-drilling", it's not calling a template — it's spawning an AI agent whose entire existence is devoted to drilling from surface symptoms to root causes. When "debate-critic" runs, it genuinely tries to destroy the idea it's reviewing. This is what makes DARE's outputs qualitatively different from prompt-chaining systems.

Built on pi-ai — a lightweight framework for building LLM-powered tools as MCP servers.

⚔️ Adversarial Validation: Ideas Must Survive Attack

Every significant output in DARE goes through adversarial debate before being accepted. The Proposer-Critic-Judge architecture isn't decoration — it's the core quality mechanism:

1. A Proposer presents the gap/insight/idea
2. A Critic attacks it from every angle (4 rounds of critique)
3. A Defender responds to each attack (3 rounds of defense)
4. A Judge evaluates the exchange and scores the result

Ideas that survive this gauntlet are genuinely robust. Ideas that don't are discarded or refined. No hand-waving, no "sounds good to me."

🎲 Quality × Diversity: Not Just Good Ideas, Different Ideas

Most AI systems optimize for a single quality metric — they'll give you 10 variations of the same good idea. DARE uses MAP-Elites, a quality-diversity algorithm that maintains a population of ideas spanning multiple dimensions of variation. The result: you get the best idea in each niche, not 10 copies of the same insight.

📏 Research Depth & Breadth Enforcement

AI agents naturally take the path of least resistance — searching a handful of papers and declaring victory. DARE embeds hard enforcement mechanisms directly into every skill to prevent this:

• Research Budget: Every strategy declares quantitative floors with three topic-size tiers (Small / Medium / Large). A literature survey on a Medium topic must fetch 40+ papers and 50+ web pages.
• State Ledger: A progress table printed before every iteration — the AI cannot lose track of where it stands.
• Budget Gate: A <HARD-GATE> that blocks the strategy from exiting its loop until 80% of the budget is met. The AI cannot stop early no matter how "satisfied" it feels.
• Adversarial Completeness Probe: After the budget is met, a qualitative self-check probes for blind spots (missing sub-areas, unchecked citations, unexplored perspectives). Up to 2 extra iterations if gaps are found.
• Yield Reports: Every tactic prints execution metrics (papers fetched, ideas generated, methods used) that feed the calling strategy's ledger.
• Cross-Domain Discovery: Before any ideation method runs, a mandatory phase searches 3+ unrelated domains for analogical inspiration — because the best ideas come from unexpected collisions.

You ask a question
    ↓
┌────────────────────────────────────────────────────────────────┐
│  Phase 0: Brainstorming (structured requirement clarification) │
│  Phase 1: Intake (research brief)                              │
│  Phase 2: Research Loop (Stages 1-3, up to 7 rounds)           │
│    ├── Literature Survey (S:20 / M:40 / L:60+ papers)          │
│    ├── Gap Analysis (S:10 / M:15 / L:25+ papers)               │
│    ├── Insight (7-step pipeline)                               │
│    ├── Ideation (cross-domain discovery → 31 methods × 5)      │
│    └── Review → Selective Redo → Review (score ≥ 8/10)         │
│  Phase 3: Experiment Design                                    │
│  Phase 4: GPU Execution (remote pod, fully autonomous)         │
└────────────────────────────────────────────────────────────────┘
    ↓
Results returned via git

Each stage runs SEARCH → READ → REFLECT → EVALUATE cycles with autonomous gap discovery and dynamic stopping. No human in the loop.

🏗️ Architecture (v3.0)

Four-layer skill hierarchy where each layer calls only the layer below:

┌───────────────────────────────────────────────────────────────┐
│  META-STRATEGY (/dare)                                        │
│  Entry point — orchestrates the full research pipeline        │
├───────────────────────────────────────────────────────────────┤
│  STRATEGY (8)                                                 │
│  intake, lit-survey, gap-analysis, insight, ideation,         │
│  round, paper-writing, method-evolve                          │
├───────────────────────────────────────────────────────────────┤
│  TACTIC