The Research Integrity Council.
The Council reviews every paper we publish, every public benchmark, and every public claim about a result. It can hold a publication, ask for a revision, or refuse one. It can also retract a paper after the fact if a number does not survive scrutiny.
Verification Topology.
A manuscript enters the Council, passes through four checks, and leaves as a publication or a written refusal. None of the checks are optional. None of them can be skipped because of who the author is.
What each step actually tests.
Reproducibility.
A second team rebuilds the result from the released code and weights. The numbers in the paper have to match what comes out of the rebuild. If they do not, the paper does not ship until the gap is understood.
Methodology.
A reviewer with relevant background reads the experimental design before the result. Sample sizes, baselines, ablations, and statistical tests are checked against the claim the paper actually makes.
Hazard review.
Capability research is read for dual-use risk. Where a publication would help an attacker more than a defender, the Council can redact specific details or hold the paper while a fix or warning is put in place.
Citations and prior art.
Every quantitative claim is traced to its source. Numbers we did not generate ourselves carry a citation. Numbers that look like a citation but are not get sent back.
Independent on paper and in practice.
Independent of the research team.
The Council does not report to the author or to the head of research. Members cannot be removed for the substance of their reviews.
External by default.
Most seats are held by researchers outside Blankline. They are paid for their time, and they hold no equity in the company.
Decisions are written.
Every review produces a short written outcome: ship, revise, hold, or refuse, with the reason and the votes. Published papers carry the Council’s decision date.
Retractions are public.
If a paper we published turns out to be wrong, we say so on the paper page and on this site, with the date and the correction.
Integrity is binary.
There is no almost-passing the checks. A paper either survives them and is published, or it does not and is held until it can. When we get something wrong, we say so on the same page where we said it the first time.