The severity ladder
Our "Forensic Metascience AI toolkit" tags every finding with a severity score between 1 to 9. This ranking system was developed from the 1-11 severity ranking system developed at the Black Spatula Project.
Two consequences follow. A finding can be certain and still sit low: a typo is a typo however sure you are of it. And a finding can be alarming and still be capped, because what an individual check is able to establish is bounded by the kind of evidence it produces — ceilings that are enforced in code rather than requested in a prompt.
The nine rungs
Trivial error
Cosmetic slip; nothing to amend.
A cosmetic slip that is self-evidently a slip: a typo, a formatting glitch, a malformed reference. No reader could carry away a wrong fact, because the error is visibly an error. Confirmed and alone, the journal would do nothing formal: the record needs no amending.
Minor error
A wrong fact that nothing depends on.
A stated fact is wrong, but nothing depends on it: a wrong unit obvious from context, a mislabeled figure panel the caption disambiguates, a reference-numbering slip, a value that feeds no calculation. Levels 1-2 are one reporting band -- use 2 when unsure. Confirmed and alone, the journal would at most publish a brief erratum; the core findings stand unchanged.
Notable error
A reporting error that changes one result.
A reporting or assembly error that changes one reported result: a miscalculated supplementary value, a misreported subset of the sample, an assembly-mistake duplicate panel that does not misrepresent the experiment. The underlying work is sound -- correcting it means fixing the report, not redoing the analysis. Confirmed and alone, the journal would issue a localized correction; the paper's main claims still hold.
Localized methodological flaw or omission
One part of the work is flawed and must be redone.
The work itself is flawed in one part: a calibration error behind one figure, a data-processing mistake in one analysis, an omitted method detail that blocks reproduction. That section must be redone or re-derived, not merely relabeled, yet the paper's overall interpretation survives the fix. Confirmed and alone, the journal would require a public corrigendum for the affected portion.
Major error affecting conclusions
A main claim of the paper does not survive.
A named claim of the paper does not survive: a miscalculation that changes an important quantitative result, a compromised dataset, a statistical error behind one main conclusion. You can still say WHICH claim falls and which stand. Confirmed and alone, the journal would publish a major corrigendum with revised conclusions, or retract the affected portion while the rest stands.
Severe error -- results unreliable
Too much is wrong to say what survives.
Problems pervasive enough that you cannot say which results survive: errors across many analyses, a fundamental design flaw biasing the study, a contaminated or misidentified key reagent or cell line. This rung records UNBOUNDED SCOPE, not a journal's procedural state -- whether an investigation has opened is not something this pipeline can know. Confirmed and alone, the journal would issue an expression of concern and investigate.
Fatal error invalidating main findings
An honest mistake that destroys the central finding.
An honest mistake that is determinedly fatal: the wrong cell line or strain, a coding or calculation error that erases the key result, data misread so the primary conclusion is wrong. Nothing can be isolated or repaired. Confirmed and alone, the journal would retract the paper as an honest error.
Serious misconduct
Deliberate manipulation, evidenced, possibly confined.
A breach of research integrity rather than a mistake, possibly confined in scope: plagiarism, deliberate data manipulation or falsification in part of the work, research conducted without ethical approval. This rung requires POSITIVE evidence of manipulation, never mere improbability. Confirmed and alone, the journal would retract, and the notice would name the misconduct category.
Malicious or extreme misconduct
Systematic fraud, or fraud that harmed the public.
Willful, systematic fraud broad in scope or with direct consequences for public welfare or policy: fabricated datasets or images across figures or papers, a manipulated clinical trial, deception that misled a field. Confirmed, the journal would retract with condemnation and institutions would pursue sanctions. This rung is never machine-assigned: it is a human's call, recorded through review, not through this pipeline.