A research run leaves a trail in context/. Most write-ups bury that trail under narrative prose an agent cannot execute. This package compiles the trail into an artifact an agent can reproduce — no storytelling tax.

📦 ara-from-context

Compile a context/ research record into an Agent-Native Research Artifact

Takes the full trail a DARE research run deposits in context/ — the final experiment-execution report, the iteration history, the produced figures — and compiles it into an ARA (Agent-Native Research Artifact): a machine-executable four-layer knowledge package. No LaTeX, no narrative paper. ARA deliberately rejects storytelling ("an agent needs execution-precision, not persuasion"); the output is a logic arc that closes structurally, not prose that persuades a human reader.

🧭 Part of the De-Anthropocentric Research Engine. This is the terminal "write-up" stage of DARE — it consumes what the research↔experiment-execution loop left in context/ and emits an ARA. Shipped as its own repo first, it is now integrated as the tenth package in the DARE flat body (since v3.2.0). To use it as intended — with the spec-driven orchestrator and the upstream research packages feeding context/ — clone the main repository rather than this repo alone.

---

⚡ What It Does

• 📑 Three-type material sort — reads context/INDEX.md and sorts the directory into report line (→ claims/problem), process line (failures/pivots → exploration tree), and images (→ evidence). The process line is salvaged actively, not dropped.
• 🧭 North-star alignment — deep-reads the original north-star context, distills the artifact's direction, and aligns it with you before compiling, so the reverse-compiled ARA does not drift off-topic.
• 🏗️ One inline compile — runs the external ARA compiler once to produce a globally-consistent four-layer artifact (cross-layer bindings intact), not a stitched-together batch.
• 🔬 Level-2 rigor review — runs the external rigor-reviewer over the result, producing level2_report.json (six semantic dimensions + grade).

---

🏗️ Architecture

campaign: ara-from-context           (no strategy layer — the 4-layer
│                                      architecture is a type system,
│                                      not a mandatory traversal)
├─ tactic: context-review
│    ├─ sop: context-exploring        sort 3 material types + locate north-star
│    └─ sop: north-star-align         distill direction + align with user
│         └─ reuses: present-and-ask, present-candidates  (from DARE body)
└─ tactic: compile-and-review
     ├─ sop: ara-compile              inline external `compiler` → ../ara/
     └─ sop: ara-rigor-review         external `rigor-reviewer` → level2_report.json

All skills are flat under skills/ — no nested tactic/sop/ subdirectories — so the directory can be copied straight into a .claude/skills/ set.

---

🔗 Dependencies

Dependency	Source	What It Provides
compiler	@ara-commons/ara-skills	Reverse-compiles any research input into a complete ARA (4-stage protocol + Seal Level 1)
rigor-reviewer	@ara-commons/ara-skills	ARA Seal Level 2 — six-dimension semantic review → level2_report.json
present-and-ask / present-candidates	DARE north-star-crystallization	Dialogue SOPs reused for user alignment

Install the ARA skills once (into ~/.claude/skills/):

npx @ara-commons/ara-skills

Without them, ara-compile / ara-rigor-review stop with an install prompt rather than failing silently. The north-star dialogue SOPs ship with the DARE body this package runs inside.

---

🚀 Usage

Run the campaign against a workspace that has a context/ directory (with INDEX.md):

/ara-from-context

Output: ara/ (sibling to context/, containing logic/ src/ trace/ evidence/ PAPER.md) + ara/level2_report.json.

Validate the package structure

python scripts/validate.py skills

Exits 0 when every skill's dependencies resolve to a sibling skill or a known external skill.

---

📄 License

Apache-2.0

---

A component of the De-Anthropocentric Research Engine, part of the Yogsoth AI ecosystem. Built by Pthahnix.