v3.2.2 — falsification-first-stress-test family

Patch release adding the falsification-first-stress-test truth-seeking family to the stress-test package.

New

• falsification-first-stress-test campaign + 6 strategies (adversarial-debate-truthseeking, red-team-truthseeking, isomorphism-falsification, circular-validation-audit, independent-convergence-audit, elegance-trap-probe) — a truth-seeking sibling family whose win-condition is inverted to refutation: verdicts are BROKEN / CORROBORATED / UNFALSIFIABLE. stress-test package grows 103 → 110 skills. Body graph artifacts refreshed to include them.

Housekeeping

• Bump version to 3.2.2 (root package.json, cli/package.json, README architecture codename).

The v3.2 pure-skill architecture is unchanged.

v3.2.1

Patch release rolling up the mechanism-gap-hunting campaign skill plus a series of small doc and dependency fixes.

New

• mechanism-gap-hunting campaign skill — new campaign-layer skill in the deep-insight package (closure 112 → 113 members). Body graph artifacts refreshed to include it.

Fixes

• Add missing edges key to ara-from-context graph data, resolving a render_combined.py KeyError.
• README: add (DARE) acronym to title, link the ARA repo, document required external dependencies (superpowers / ponytail / ara) in Quick Start.

Housekeeping

• Bump version to 3.2.1 (root package.json, cli/package.json, README architecture codename).

The v3.2 pure-skill architecture is unchanged.

v3.2.0 — ara-from-context (10th package)

Adds ara-from-context: compile a completed context/ research record into an ARA (4-layer agent-native artifact) + Level-2 epistemic review. 10th composable package. writing-specs now emits an optional ARA closing stage with a user-decision gate. Requires external compiler/rigor-reviewer (npx @ara-commons/ara-skills).

v3.1.0 — Self-Contained 9-Package Body

DARE v3.1.0 — Self-Contained 9-Package Body

This release lands the skills-dependency refactory (#25): the skill body is now fully self-contained, its dependency graph encoded inline in frontmatter, the documentation reorganized around 9 freely-composable packages, and the corpus is pure English.

✨ Highlights

🔗 Self-contained dependency graph
Every skill declares its own dependencies inline. Each SKILL.md carries a dependencies block (campaigns / strategies / tactics / sops sub-keys) naming the exact lower-layer skills it may call. No external imports, no reverse used-by sprawl — the entire 905-skill call graph is reconstructable from frontmatter alone, and is machine-verified closed: 2476 / 2476 skill→skill edges resolve.

🧩 9 freely-composable packages
The body is now organized as 9 self-contained research engines — creative-ideation, convergence, deep-insight, stress-test, knowledge-acquisition, experiment-execution, hypothesis-formation, knowledge-structuring, north-star-crystallization — each obeying the same internal 4-layer command hierarchy (Campaign → Strategy → Tactic → SOP). There is no fixed pipeline: CC reads research-catalog after the direction is crystallized and routes across packages as the research demands.

📚 Generated capability tables
786 skills carry deterministic ## Available … tables derived from frontmatter; 9 research-catalog/references/<pkg>.md tables list every skill per package.

🌐 Pure-English corpus
The full skill body is now English-only — 0 non-Latin characters across 905 SKILL.md files, 0 frontmatter parse failures.

📝 README realigned
"Arsenal, not pipeline" pushed all the way; two orthogonal axes documented (9 packages × per-package 4 layers); skill distribution rewritten to 9 packages + infrastructure; counts corrected to real totals.

📊 By the numbers

Layer	Count
Campaign	46
Strategy	209
Tactic	129
SOP	520
Total skills	905

9 packages (≈889 skills) + infrastructure (engine-core 8, literature-engine 3, context-management 3, web-browsing 2, subagent-spawning 1).

✅ Verification

• Closure gate: 2476/2476, missing=0
• Test suite: 60 passed, 1 skipped
• Han/non-Latin scan of skills/: 0 files
• Frontmatter parse of 905 SKILL.md: 0 bad

⬆️ Upgrading

No action required for existing clones beyond git pull. MCP dependencies are unchanged (6 servers: semantic-scholar, wiki-vault, brave-search, tavily-search, apify, alphaxiv).

Full changelog: v3.0.0...v3.1.0

v3.0.0 — Pure-Skill Architecture

What's New

Complete architectural rewrite. DARE is now a pure-markdown skill system — 800+ skills, zero application code, CC as the runtime.

Architecture

• 4-layer military hierarchy: Campaign (8) → Strategy (40+) → Tactic (100+) → SOP (600+)
• 9 orchestrator skills as the control plane above the hierarchy
• Non-linear execution with explicit backtrack conditions and ±10% deviation rules
• Executable Research Specs with checkbox progress tracking and session recovery

Orchestration Flow

1. North Star Crystallization (cold/warm/hot-start)
2. Research Spec Generation (structured questioning → outline → full spec)
3. Spec Execution (autonomous, stage-by-stage, with context checkpoints)

MCP Integrations (5 servers)

• @yogsoth-ai/semantic-scholar-mcp — paper lookup, citations, recommendations
• @yogsoth-ai/wiki-vault — research knowledge graph with BM25 search
• @brave/brave-search-mcp-server — web search, news, LLM context
• @apify/actors-mcp-server — web scraping, Google Scholar
• alphaxiv (HTTP) — arXiv paper search, PDF queries

Design Philosophy

Arsenal, not pipeline. The AI has full cross-stage routing authority. Research Specs define when to backtrack, not just what order to execute. Every existing autonomous research system (AI Scientist v2, AI-Researcher, Dolphin, Agent Laboratory, ARIS) is a fixed pipeline — DARE is the first to implement non-linear agent-decided orchestration with explicit backtrack mechanisms.

Skills by Source

Source	Count	Coverage
north-star-crystallization	~30	Direction finding
knowledge-acquisition	~120	Literature, citation chaining
deep-insight	~80	Gap analysis, abstraction
hypothesis-formation	~60	Abductive/inductive/deductive
creative-ideation	~150	SCAMPER, TRIZ, biomimicry, morphological
convergence	~90	Multi-criteria scoring, Pareto, synthesis
stress-test	~70	Red-teaming, assumption destruction
experiment-execution	~50	Factor design, sensitivity analysis
infrastructure (4 repos)	~40	Web, literature, subagents, context

---

The orchestrator of the Yogsoth AI research ecosystem.

v0.1.0 — First Public Preview

DARE is not a tool that helps you do research — it is the researcher. You set the direction; DARE searches, reads, discovers gaps, generates ideas, designs experiments, and executes them on GPUs. Autonomously. Iteratively. Without asking for permission.

This is the first public preview of the De-Anthropocentric Research Engine.

Highlights

• 49 LLM micro-agent tools — each tool is a single-responsibility AI agent with its own system prompt and reasoning chain. When "debate-critic" runs, it genuinely tries to destroy the idea it's reviewing.
• Autonomous literature survey — searches Google Scholar, downloads full papers (not just abstracts), reads them cover-to-cover with a three-pass Keshav protocol
• 31 ideation methods across 5 categories — SCAMPER, component surgery, cross-domain collision, perspective forcing, structural deconstruction — filtered by MAP-Elites quality-diversity algorithm
• Adversarial debate — Proposer-Critic-Judge architecture validates every gap, insight, and idea through structured multi-round debates
• 7-step INSIGHT pipeline — root-cause drilling, stakeholder mapping, tension mining, question reformulation, abstraction laddering, assumption audit, and validation
• Deep reference exploration — traces citation graphs via Semantic Scholar, enriches metadata from arXiv and Unpaywall

Architecture

DARE follows a four-layer military command hierarchy — each layer calls only the layer directly below it:

General  (Meta-Strategy)  ->  "Take that hill"         ->  WHAT to research
Colonel  (Strategy x8)    ->  "Flank from the east"    ->  WHEN and WHY
Captain  (Tactic x15)     ->  "Squad A cover, B move"  ->  HOW to combine
Sergeant (SOP x60)        ->  "Fire, reload, advance"  ->  HOW to execute
         (Tools x49+)     ->  atomic MCP operations    ->  WHAT to do

84 skills total. A Strategy never touches tools directly; a Tactic never decides research direction. Every component is independently testable, replaceable, and composable.

Under the Hood

Monorepo with 4 packages, 8 MCP servers, 85+ tests:

Server	Tools	Purpose
dare-agents	49	LLM micro-agent tools (ideation, debate, insight, method-evolve)
dare-scholar	5	Academic paper pipeline — search, fetch, read, reference
dare-web	2	Web page fetching and markdown caching
dare-session	—	Git-based context transfer to remote GPU pods
+ apify, brave-search, runpod, alphaxiv	13	External services

Built on pi-ai for LLM completions and the Model Context Protocol for tool orchestration. Designed to run inside Claude Code.

Get Started

git clone https://github.com/Pthahnix/De-Anthropocentric-Research-Engine.git
cd De-Anthropocentric-Research-Engine
npm install

See the README for full setup (API keys, MCP server configuration).

What's Next

• Experiment execution — autonomous experiment design and GPU execution via RunPod
• Method evolution — AlphaEvolve-inspired evolutionary improvement of DARE's own research methods (core tools already implemented)
• Paper writing — end-to-end autonomous paper drafting from research findings