A protocol that recasts the primary research object from narrative document to machine-executable knowledge package — so AI agents can navigate, reproduce, and extend published research without re-discovering every dead end.
Publishing compiles a rich research object into a lossy narrative (left); ARA preserves the original as a high-fidelity, machine-executable knowledge package (right).
Research produces a branching knowledge object — months of hypotheses tested and rejected, implementation tricks discovered through trial and error, design alternatives weighed. Publishing compiles this into a linear narrative, discarding everything that doesn't fit the final story.
This was tolerable when every consumer was human. It is not when AI agents routinely read papers to reproduce experiments and extend published methods.
The numbers:
- Only 45.4% of 8,921 reproduction requirements from 23 ICML 2024 papers are fully specified in their PDFs (PaperBench)
- Failed agent runs account for 90.2% of total dollar cost across 24,008 runs on RE-Bench — agents without prior failure records rediscover every dead end independently
ARA organizes research into four interlocking layers:
artifact/
PAPER.md # Root manifest + layer index (~200 tokens)
logic/ # Cognitive layer — What & Why
problem.md # Observations → gaps → key insight
claims.md # Falsifiable assertions with proof refs
concepts.md # Formal definitions
experiments.md # Declarative experiment plans
solution/
architecture.md # System design + component graph
algorithm.md # Math + pseudocode
constraints.md # Boundary conditions
heuristics.md # Implementation tricks + rationale
related_work.md # Typed dependency graph
src/ # Physical layer — How
configs/ # Hyperparameters with rationale
environment.md # Dependencies, hardware, seeds
trace/ # Exploration graph — Journey
exploration_tree.yaml # Research DAG with typed nodes + dead ends
evidence/ # Raw proof
tables/ # Exact result tables
figures/ # Extracted data points
Cross-layer forensic bindings thread claims in /logic to code in /src and evidence in /evidence. Dead-end nodes (×) in the exploration graph preserve failure modes.
- Progressive disclosure —
PAPER.md(~200 tokens) tells agents whether the artifact is relevant. Deeper files load on demand. - Cross-layer binding — Claims reference experiments, experiments reference evidence, heuristics reference code. Everything is linked.
- Dead ends preserved — Failed approaches and rejected alternatives are first-class nodes in the exploration graph, preventing agents from rediscovering known failures.
- Provenance tracking — Every entry carries a tag (
user,ai-suggested,ai-executed,user-revised) distinguishing human-confirmed facts from AI inferences.
This repository ships three open-source agent skills that work with ARA:
| Skill | Description | Invoke |
|---|---|---|
| compiler | Compiles papers, repos, notes, or any research input into a structured ARA artifact | /compiler <path> |
| research-manager | End-of-turn recorder that captures decisions, experiments, and dead ends with provenance tags | /research-manager |
| rigor-reviewer | ARA Seal Level 2 semantic review — scores six dimensions of epistemic rigor | /rigor-reviewer <artifact_dir> |
Converts ANY research input into a complete ARA artifact. Accepts PDFs, GitHub repos, experiment logs, code directories, raw notes, or combinations. Follows a 4-stage epistemic protocol:
- Semantic Deconstruction — extract raw knowledge atoms
- Cognitive Mapping — map to claims, concepts, experiments
- Physical Stubbing — generate configs and code stubs
- Exploration Graph Extraction — reconstruct the research DAG
/compiler path/to/paper.pdf
/compiler https://github.com/org/repo
/compiler path/to/paper.pdf path/to/code/ --output ./my-artifact/
See skills/compiler/SKILL.md for the full specification.
An end-of-turn recorder that runs after every turn and writes research-significant events into the ara/ artifact via a three-stage pipeline (Context Harvester → Event Router → Maturity Tracker). Trace events (decisions, experiments, dead ends, pivots) are recorded immediately; knowledge events (claims, heuristics, concepts, constraints) are staged and crystallize only on closure signals — so research knowledge accrues as a side-effect of ordinary development.
/research-manager
See skills/research-manager/SKILL.md for the full specification.
A semantic epistemic review that assumes Level 1 structural validation has passed, then scores six dimensions — evidence relevance, falsifiability, scope calibration, and more — producing a level2_report.json with severity-ranked findings and an overall recommendation.
/rigor-reviewer path/to/artifact/
See skills/rigor-reviewer/SKILL.md for the full specification.
npx @orchestra-research/ara-skillsAuto-detects Claude Code, Cursor, Gemini CLI, OpenCode, Codex, and Hermes, then prompts for skills, agents, and install scope (global vs. local).
# All three skills, every detected agent, user-level
npx @orchestra-research/ara-skills install --all
# One skill, one agent
npx @orchestra-research/ara-skills install --skill compiler --agent claude-code
# Into the current project (.claude/skills, .cursor/skills, …) instead of $HOME
npx @orchestra-research/ara-skills install --all --local
# List / update / remove
npx @orchestra-research/ara-skills list
npx @orchestra-research/ara-skills update
npx @orchestra-research/ara-skills uninstall --skill rigor-reviewerFull CLI reference: packages/ara-skills/.
These skills follow the Agent Skills open standard and work with:
- Claude Code (Anthropic)
- Codex CLI (OpenAI)
- GitHub Copilot
- Cursor
- Any agent supporting the Agent Skills specification
If you use ARA in your research, please cite:
@article{ara2026,
title = {The Last Human-Written Paper: Agent-Native Research Artifacts},
author = {Liu, Jiachen and Pei, Jiaxin and Huang, Jintao and Si, Chenglei and Qu, Ao and Tang, Xiangru and Lu, Runyu and Chen, Lichang and Bai, Xiaoyan and Zheng, Haizhong and Chen, Carl and Chen, Zhiyang and Ye, Haojie and Fu, Yujuan and He, Zexue and Jin, Zijian and Zhang, Zhenyu and Sun, Shangquan and Harmon, Maestro and Wang, John Dianzhuo and Zeng, Jianqiao and Sun, Jiachen and Wu, Mingyuan and Zhou, Baoyu and You, Chenyu and Lu, Shijian and Qiu, Yiming and Lai, Fan and Yuan, Yuan and Li, Yao and Hong, Junyuan and Zhu, Ruihao and Chen, Beidi and Pentland, Alex and Chen, Ang and Chowdhury, Mosharaf and Zhang, Zechen},
year = {2026},
eprint = {2604.24658},
archivePrefix= {arXiv},
url = {https://arxiv.org/abs/2604.24658}
}See CONTRIBUTING.md for how to add or improve skills, or contribute ARA artifacts to ara-output/.