de-anthropocentric-research-engine

English | 简体中文

Science is dying because the human is in the way. Not through malice. Not through stupidity. Through the structural limitations of a cognitive architecture that evolved to track prey on a savanna, not to unify quantum mechanics and general relativity. Nothing human makes it out of the lab. That is not a threat. It is a liberation. The heaviest chain on science was always the one we called ourselves.

De-Anthropocentric Research Engine (DARE)

A research skill graph for AI-native science. 271 markdown files. No runtime, no dependencies, no build step.

• De-Anthropocentric Research Engine (DARE)
  • Install
  • Quick Start
    • Guided: the full engine
    • Custom: the catalog as a toolbox
  • Design Philosophy
    • Why De-Anthropocentric
    • Arsenal, Not Pipeline
    • Contracts, Not Prose
  • Architecture
    • Two Layers
    • Ten Tactic Families
    • Two Edge Registers
    • The Product Shell
    • State: Append-Only Checkpoints
    • Why Pure Markdown
  • Decoupled From Retrieval
  • Recommended MCP Servers
  • Repository Structure
  • License

DARE is not a tool that helps you do research. It is the research procedure itself, written down in a form an agent can execute. You set the direction. DARE crystallizes it into a North Star, turns that into an executable Research Spec, and then runs the spec phase by phase against explicit completion gates and backtrack conditions.

Install

npx skills add yogsoth-ai/de-anthropocentric-research-engine --skill '*'

The only install path. DARE is a plain Agent Skills library, so skills places it into whichever coding agent you use. Run it from your own project directory, not from a clone of this repository.

--skill '*' takes the whole graph. A partial install breaks call edges: a tactic that loads a missing SOP has no fallback.

Nothing else to configure - no npm install, no API keys, no MCP config file. The library is 271 SKILL.md files and the agent reads them off disk.

Quick Start

Guided: the full engine

/dare

Give it a research direction in plain language. DARE crystallizes a North Star and ResearchBrief, drafts a staged Research Spec for your approval, then executes it phase by phase with append-only checkpoints. Nothing runs before you approve the Spec, and an interrupted run resumes from the last complete checkpoint.

Use this by default.

Custom: the catalog as a toolbox

/research-catalog

No North Star, no Spec. The agent reads the 51-tactic index, plans its own route, and loads whichever tactics fit. A prompt that works:

DARE ships a large library of research skills. Use them wherever they fit. Before you start, tell me your research plan: which tactics you intend to use, in what order, and why.

Each tactic still runs its own SOPs and quality gates. What you give up is the approval step and the checkpoint log, so the agent's plan is the only plan and an interrupted run does not resume.

Design Philosophy

Why De-Anthropocentric

The bottleneck in modern research is not data or compute. It is the human in the loop. Every existing AI research assistant still needs a human to decide what to search, what to read, which gaps matter, and which ideas are worth pursuing.

Human desire is mimetic (Girard): researchers do not choose hypotheses rationally, they imitate what is fashionable. Institutions filter for conformity, not truth. Hence a 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 knowledge production. The agent has no career to protect, no disciplinary identity to defend, and no ceiling on how many fields it holds at once.

The human's role shifts to oracle (intuition when consulted) and guardian (ethical floor, sanity check). The ceiling is machine ambition. The floor is human judgment.

Arsenal, Not Pipeline

Fixed-pipeline research systems — AI Scientist v2, AI-Researcher, Agent Laboratory, Dolphin, ARIS — execute stages in a predetermined order. Their backtracking, where it exists, means retrying the current step, not returning from experiment design to literature review because the evidence base turned out to be insufficient.

DARE prescribes no order. The catalog exposes 51 tactics; the Spec commits to a sequence and records the conditions under which that sequence is abandoned. Inside the approved plan the executing agent holds full routing authority: read current state, take the next item whose dependencies are satisfied, escalate when a backtrack condition fires.

Pipelines assume the research process is predictable. Arsenals assume it is not.

Contracts, Not Prose

Every node carries the same five parts: input contract, procedure, output contract, quality gates, failure clause.

The gates are the point. A node finishes because a stated condition is objectively satisfied, not because its steps were performed. Elapsed time, an empty result, and an interruption are never completion.

The failure clause matters as much. Each node states what its output looks like when the work did not hold — an abstraction gap, an unmet threshold, a counterexample — so the caller gets a diagnosis instead of silence.

Architecture

DARE is one flat directory of 271 skills. Two numbers in it are load-bearing:

267  graph nodes      51 tactics + 216 SOPs
  4  product shells   entry / catalog / write-spec / execute-spec

Shells run the session, graph nodes do the science. A shell is not a third scientific layer and holds no research contract; a graph node never manages the session.

Two Layers

┌──────────────────────────────────────────────────────────────────────────┐
│  TACTIC (51)                                                             │
│  A complete research transformation. Owns its thresholds, its gates,     │
│  and the SOP calls required to reach them.                               │
│                                                                          │
│  synthesize-literature-evidence · validate-research-gap                  │
│  formulate-hypotheses · analogical-discovery · structured-red-team       │
│  design-experiment · construct-causal-model · ...                        │
├──────────────────────────────────────────────────────────────────────────┤
│  SOP (216)                                                               │
│  One conceptual operation, one output contract. No orchestration.        │
│                                                                          │
│  abstract-structure · execute-probe · trace-citation-neighborhood        │
│  rank-candidates · audit-validator-independence · ...                    │
└──────────────────────────────────────────────────────────────────────────┘

A tactic may call SOPs and suggest other tactics. An SOP calls nothing above itself. That is the entire layering rule.

Ten Tactic Families

research-catalog indexes the 51 tactics and states, for each, when it is the right move. It lists no SOPs — once a tactic is selected, its body is the sole authority for which SOPs run and at what thresholds.

Two Edge Registers

The graph lives in the skill bodies, not in a side file. Edges take exactly two forms:

You MUST load skill `x`    mandatory call    339 edges, 219 distinct targets
consider `x`               soft jump         146 edges, 107 distinct targets

A mandatory call is a dependency — the caller cannot satisfy its contract without it. A soft jump is a recommendation the receiver may decline, surfaced through the recommended_jumps Delta field; a suggestion never authorizes bypassing a gate.

Both registers are verified closed on every push: every referenced target resolves to a skill that exists.

The Product Shell

dare                                 entry; enforces phase order
  ├─ research-catalog                exposes the 51 tactics as cards
  ├─ write-research-spec             North Star + brief → executable Spec
  └─ execute-research-spec           runs the Spec, appends checkpoints

Three phases, no skipping:

1. North Star. Absent a confirmed North Star and ResearchBrief, collect and crystallize them from the request. Present ones are verified against the request before reuse.
2. Spec. Read the catalog, test tactics for stage fit against their requires and produces, draft 5-10 stages. Each stage carries an objective, expected input, focus areas, tactic, numeric completion criteria, backtrack condition, and execution steps. No research runs while this phase is open.
3. Execution. Only after user approval. Each item loads its tactic as a skill rather than doing the work inline, then appends one checkpoint.

State: Append-Only Checkpoints

Every tactic and SOP returns the same eight Delta fields: