⚡ What It Does

• 🧭 Autonomous direction crystallization — cold-start from zero, warm-start from vague interest, or hot-start from specific question. Produces a structured North Star without human hand-holding
• 📚 Deep literature acquisition — multi-pass academic paper discovery via Semantic Scholar, citation chaining, snowball sampling, cross-database verification. Not keyword search — systematic coverage
• 🔍 Gap discovery at scale — 15+ gap detection methods (coverage analysis, white-space identification, niche mapping, boundary unfolding) that find what the field is missing, not what you tell it to find
• 💡 Structured hypothesis formation — abductive, inductive, and deductive generation pipelines with falsifiability audits and competing hypothesis matrices
• 🎨 31+ ideation methods — SCAMPER, component surgery, cross-domain collision, biomimicry, TRIZ contradiction resolution, morphological analysis, concept blending, lateral thinking, and more
• ⚔️ Adversarial stress testing — multi-perspective attack, sacred cow hunting, assumption destruction, worst-case design, winner stress testing. Ideas must survive attack before acceptance
• 🔬 Convergence & synthesis — multi-criteria scoring, Pareto frontier construction, pairwise ranking, structured consensus, dialectical synthesis across competing threads
• 📏 Executable Research Specs — machine-readable documents with checkbox progress tracking, quantified completion criteria, backtrack conditions, and session recovery. Another CC instance picks up where you left off
• 🧪 Experiment design — full experimental methodology generation (factor-level design, parameter screening, sensitivity analysis) ready for execution
• 🌐 5 MCP integrations — Semantic Scholar, Brave Search, AlphaXiv, Apify web scraping, and Wiki Vault for persistent knowledge graphs

🎯 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.

Human desire is mimetic (Girard): researchers don't choose hypotheses rationally — they imitate what's fashionable. Human institutions filter for conformity, not truth. The result: 90% decline in scientific disruptiveness since 1945 (Park et al., 2023), while researcher headcount exploded. DARE's response is architectural: remove the mimetic agent from the center of the knowledge-production process. The AI has no career to protect, no disciplinary identity to defend, no cognitive ceiling on how many fields it can hold in working memory at once.

The human's role shifts to oracle (providing intuition sparks when consulted) and guardian (maintaining ethical floors and sanity checks). The ceiling is AI ambition. The floor is human wisdom.

For the full philosophical argument, see assets/DE-ANTHROPOCENTRIC.md.

🎖️ Four-Layer Command Structure: Campaign → Strategy → Tactic → SOP

DARE's architecture follows a military command hierarchy — not because research is war, but because the decomposition pattern is remarkably effective for autonomous multi-stage operations:

Campaign (8)    →  "Take that hill"         →  WHAT to research (full research stage)
Strategy (40+)  →  "Flank from the east"    →  WHEN and WHY (iteration loops, stopping conditions)
Tactic (100+)   →  "Squad A cover, B move"  →  HOW to combine (orchestrates multiple SOPs)
SOP (600+)      →  "Fire, reload, advance"  →  HOW to execute (single-responsibility operations)

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.

Campaigns are the 8 stages of the research pipeline — from north-star-crystallization through experiment-execution. Each campaign owns a complete research phase and defines its own completion criteria, backtrack conditions, and context protocol.

Strategies are the iteration engines within campaigns. A literature survey strategy manages the search-read-reflect loop; a gap analysis strategy manages coverage scoring and saturation detection. Strategies hold state (ledgers, budgets) and decide when to stop.

Tactics combine multiple SOPs into coherent workflows. A "cross-domain collision" tactic orchestrates domain scanning, analogy extraction, forced bridge construction, and blend evaluation into a single creative operation.

SOPs are atomic, single-responsibility operations. Each SOP wraps one conceptual action: run one search, score one hypothesis, extract one analogy. 600+ SOPs provide the granular building blocks that higher layers compose.

⚔️ Arsenal, Not Pipeline

Every existing autonomous research system — AI Scientist v2 (Sakana), AI-Researcher (HKUDS), Agent Laboratory, Dolphin, ARIS — implements a fixed pipeline: stages execute in a predetermined order, and the agent's autonomy is confined to local decisions within a single stage. Backtracking, when it exists at all, means retrying the current step — not returning from experiment design to literature review because the knowledge base turned out to be insufficient.

DARE is not a pipeline. It is an arsenal — a strategy book that the AI reads, then decides how to act.

What this means concretely:

In a pipeline system, the workflow is hardcoded: literature → gap → hypothesis → experiment. The agent has no say in the order, cannot skip stages, and cannot go back. If the experiment phase reveals that the literature review missed a critical subfield, the system has no mechanism to return and fix it.

In DARE, the Research Spec defines backtrack conditions for every stage — explicit rules like "if stress-test invalidates >50% of hypotheses, return to hypothesis-formation." The executing agent has full cross-stage routing authority: it reads the spec, assesses the current research state, and decides which campaign to invoke next, which strategies within that campaign to combine, and when the current path has failed hard enough to warrant retreat.

Within each campaign, the agent faces not one method but many. A gap-analysis campaign offers 15+ detection methods (coverage mapping, white-space identification, boundary unfolding, niche analysis...). A creative-ideation campaign offers 31+ generation techniques (SCAMPER, TRIZ, biomimicry, morphological analysis, concept blending...). The agent selects and combines methods based on the research context — not because "more is better," but because different research problems demand different tools, and a system locked to one approach per phase cannot adapt.

The human's role: approve the spec (including its backtrack conditions and recommended campaign combinations) before execution begins. After that, the agent navigates the research space autonomously within the ±10% deviation bounds defined in the spec. If it needs to deviate further — backtrack to an earlier stage, skip a stage entirely, or add one — it asks.

This is the fundamental architectural difference. Pipelines assume the research process is predictable. Arsenals assume it is not.

📏 Executable Research Specs

Traditional research plans are prose documents that humans interpret. DARE produces Research Specs — documents that are simultaneously human-readable and machine-executable:

• Checkbox syntax (- [ ]) tracks progress across sessions
• Quantified completion criteria (no vague "sufficient" — always numbers)
• Explicit backtrack conditions with target stages
• Context protocol (init/checkpoint) baked into every stage
• ±10% deviation rules: CC can adjust within bounds, must document deviations

A spec is a contract between the human who approved it and the CC instance that executes it. Session recovery is automatic: read the spec, find the first unchecked box, read the latest context checkpoint, resume.

🧠 Con