An agentic pipeline for small-molecule and binder design in which refusal is a first-class output. A calibration gate sits before ranking and decides whether a method may rank at all. On our own first programme, it fired.
A scoped computational review, delivered with the evidence that produced it and the limits that bound it. Assays are agreed separately.
Structures or sequences with the design rationale, and the comparator each one was built against.
Model outputs, the controls that ran, and the conditions under which each number is interpretable.
The comparator, the test and the decision the result resolves — written before the experiment, not after.
A calibration gate sits between modelling and candidate design. It asks of a method on a target: has it been shown to track compounds whose potency is already known? A method that cannot be shown to work does not get to produce a ranking. Inadequate evidence and demonstrated sub-floor performance are different findings, and the vocabulary keeps them apart.
Discrimination above the row's own permutation null, with a lower bound above the declared floor. Both conditions must hold.
The upper bound sits below the floor. Sub-floor performance was demonstrated, not assumed.
Everything else. The evidence does not settle the question, which is not the same as the method failing.
The instrument broke or a control did not return. A tooling failure never becomes a scientific finding.
We publish the audit trail, including the parts that went against us.
Three mechanisms convert unasked questions into blocking nulls: instrument positive controls, so an absence from a corpus that failed its own control cannot be read as evidence; an authority split, so no module both proposes and eliminates; and the calibration gate, whose unit of abstention is a method on a target rather than a prediction on a molecule.
The paper reports its own withdrawal log, the defects we found by auditing our benchmark, and an adversarial review that found our proposed remedy overstated. We argue that these artifacts — not the shortlist — are how agentic systems should be compared. Results and methods are in the paper; this page does not reproduce them.