An AI agent can write a convincing manuscript in minutes. A fabricated citation reads exactly like a real one. Researka triages each submission, attacks every claim against a rubric, and returns a public decision record before a human reviewer spends an hour on work a machine should have caught.
We do not certify that a paper is true. We certify that it survived adversarial review, and we publish exactly why it did or did not. That distinction is the product.
Why: Integrity Duplicate
Exact-content duplicate of publication 04a4e243-1f41-43e8-9394-cf52f62997db. Resubmission requires substantially new content.
Failed checks
Proof: provenance chain
A Researka accept means Researka-reviewed. It does not mean verified true. Every accepted artifact carries the same note: the work does not support broad causal, clinical, or policy claims and is hypothesis-generating. Narrow beats broad. Mouse is not human. Association is not causation.
01
Forward a submission
Send a manuscript through REST, MCP, or by hand. Bring your own corpus and your own queue.
02
Adversarial review
Intake gates catch fabricated citations and duplicates first. Reviewer agents then challenge every claim against a fixed public rubric.
03
Decision and receipt
You get accept, revise, or reject with the reasoning, a SHA-256 content hash, and a public decision record you can cite in a desk-reject.
Every decision is public, with the specific reason. Open the ledger and read the ones we turned away.
128
Accepted, published with provenance
39
Rejected, each with the reason
11
Returned for revision
You are using AI to police AI.
Correct, and we constrain it. Reviewers run in different model families against a fixed public rubric, and every verdict is published with its reasoning.
An accept is just a stamp.
We say so on every artifact. Researka-reviewed, not verified true. Acceptance certifies the claims were challenged and traced to sources, not that the conclusions are correct.
How do I know the rejections are real?
Every decision has a public permalink, a content hash, and a provenance chain. You can read all of them, today.
This is just a detector.
No. A detector scores how likely a machine wrote the document. We adjudicate each claim against a rubric and publish the decision record.
What about false rejections?
Revise exists for exactly that. Borderline work is sent back with specific required changes, not rejected outright.
Does this replace peer review?
No. It sits before it. Researka-reviewed is a structured pre-review you can act on, not a substitute for your journal peer review.
Run a pilot on your own queue. You see the decision record, the gate checks, and the provenance receipt for every artifact.
Request a pilotOr read the methods and rubric, the decisions ledger, and the rejections.