| Status | Design lesson — NOT a proposed feature |
| Parent epic | #255 — Kong et al. 2026 auto-research survey implications for ARS |
| Paper anchor | Kong et al. (2026), AI for Auto-Research: Roadmap & User Guide, arXiv:2605.18661 |
| Verified | 2026-06-08 — verified against the tracked repo (see Verification note) |
This is a recorded boundary, not a roadmap item. It exists so that future work and reviews apply the same line consistently.
- Paper: Kong et al. survey AI auto-research systems organised around eight stages × four phases of the research lifecycle — idea generation, autonomous experimentation, end-to-end paper pipelines, autonomous peer review, Paper2X dissemination. The paper's own cross-cutting conclusion (§7.3.2, §7.4.8) is that human-governed collaboration is the most reliable deployment mode and that fully autonomous science is not a credible near-term objective.
- ARS:
POSITIONING.md— "the human decides at every gate" (Design philosophy); ARS is a human-led copilot, not an autonomous paper-writing system.
ARS and an auto-research system can look superficially alike. Both use multi-agent orchestration, a layered architecture, and simulated critique (ARS runs a five- reviewer panel plus a Devil's Advocate). A naive cross-walk could read these surface features as "ARS is most of the way to auto-research" — or, in the other direction, a reviewer could wave through a new agent on the grounds that "we already have many agents, one more is fine."
Both reads miss the actual line. The line is not how many agents there are or how layered the architecture is. It is who controls the next research-state transition — who gets to create, select, execute, or advance a research object of record, and whether the scholar saw and confirmed the relevant state first.
Any PR that adds or extends an agent, mode, or automation is reviewed against this question before merge:
Does this change let ARS create, select, execute, or advance a research object of record — hypothesis, RQ, evidence set, experiment result, claim, manuscript section, or dissemination artifact — without an explicit scholar-authored seed or scholar confirmation after inspecting the relevant state?
- If yes → it crosses into auto-research and is rejected, regardless of how useful or well-tested the mechanism is.
- If no → it stays copilot/advisory and is evaluated on its own merits.
These are the specific surfaces where ARS could drift across the line if a change were not tested against the question above:
- Multi-agent orchestration. Adding agents is fine; letting an agent advance the pipeline state (select the next stage, accept its own output as final) without a scholar gate is not. Mandatory checkpoints are the gate.
- Layered architecture. Layers may transform and annotate; they may not become a closed loop that produces a submission-ready object the scholar never confirmed.
- Simulated critique. The reviewer panel and Devil's Advocate critique the scholar's draft; they do not author, accept, or act on their own critique. AI review is advisory input to the scholar, never a substitute for the scholar's decision (and never a substitute for actual peer review).
Three first-class architectural commitments, all already present:
- Mandatory checkpoints. FULL and MANDATORY checkpoints cannot be skipped; "full mode" is full-pipeline execution, not full autonomy.
- Scholar-as-author. The researcher remains the author of every claim, citation, experimental design, and interpretation. AI assists; it does not author.
- Disclosure, not detection. ARS makes AI assistance visible and discloseable; it does not position itself as an autonomous producer whose output passes as unassisted.
Surface-level rejection ("multi-agent systems are auto-research, avoid them") would throw away orchestration, layering, and simulated critique — all of which are useful when the scholar owns every state transition. State-authority rejection ("no AI- owned transition of a research object of record") keeps those primitives while holding the line.
The reverse is equally true. Permitting a new agent on the grounds that "we already have many" silently re-introduces the risk: the question is never the count, it is whether the new agent can advance state the scholar has not seen and confirmed.
The Co-Scientist analysis recorded a related boundary from a different angle — who may rank and whether the user sees the full candidate set. Those docs and this one share the same underlying principle (the scholar must see and own the state the system acts on): Co-Scientist L1 (hidden ranking), L2 (feedback propagation), L3 (mechanism transfer), L4 (who may write / rank / route). The Kong L2 sharpens the specific advisory-vs-generation seam for research questions.
- Verifiable (2026-06-08, against the tracked repo): first-party ARS does not
implement an end-to-end autonomous research pipeline, an idea-generation agent, a
Paper2X generator, an autonomous experiment executor, or a wet-lab automation API.
"The human decides at every gate" is present in
POSITIONING.md; the mandatory checkpoints are real (academic-pipeline/SKILL.md). The Kong-derived features that did ship — wording-pattern advisory (#257) and experiment provenance intake (#260) — are advisory / provenance gates, not autonomous generation or execution layers. A targeted review of the first-party implementation, agent prompts, and schemas found no autonomous mechanism; the onlywet_lab/ slides / experiment references are domain-evidence-profile labels or separate, non-ARS companion workflows, not automation surfaces. - Design commitment (not a guarantee): this doc records a first-party scope boundary and a review criterion for future changes. It is not a proof that the software cannot be modified, a guarantee about forks or downstream deployments, or a runtime enforcement mechanism.