Asked, answered.

A fixed-fee, fixed-scope measurement engagement: two weeks, one agent loop. We run adversarial, preregistered, reproducible tests against your production agent and deliver the false-completion rate, attempted vs executed violations, variance across sampling settings, and token economics, plus a regression harness that stays with you and a recommendation for where the boundary goes in your architecture. Details are on the front page.

Predictions, thresholds, and controls are committed to version control before a run. Our tooling mechanically refuses to record a result whose preregistration isn't already in history. You hold us to the predictions in the statement of work, the same standard our own published record holds itself to, including the misses.

Access to one agent loop (or a faithful staging copy), agreement on what ground truth means for your tasks, and a signed statement of work with the preregistered predictions. We don't need your model weights, and the harness runs against your interfaces, not inside your infrastructure.

A, reliability: a calibrated probability that each newly deposited structure will later be substantively revised by the archive itself; the control is ranking by resolution alone. B, divergence: where deposited experiments disagree with the corresponding predicted models, residue by residue, focused on interfaces; the control is the predictor's own confidence. C, novel pockets: binding pockets whose geometry is new while their sequence family is known, ranked by evidence tier. All three grade weekly on the PDB release; see the results.

A geometry flag alone isn't a claim worth your time. Tier 1: novel by the learned embedding, no bound ligand. Tier 2: a real ligand contacts the pocket in the deposited structure: strong evidence it's a true binding site. Tier 3: that ligand chemistry is new for the pocket's sequence family: a familiar protein showing genuinely new bound chemistry. Every claim is checkable: open the structure and look.

You shouldn't, which is why it isn't. The models are fit only on structures released before a fixed temporal cutoff, frozen, and committed; they're graded on releases they have never seen, against oracles they don't control (the archive's own future revisions, the deposited experiment, later ligand annotations). And the failure mode is public: "could not beat the control" publishes at full prominence.

Yes: the public weekly snapshots are free to use with attribution. Each is versioned, dated, and immutable, so what you cite stays what you cited. Bulk downloads, the full embeddings, per-residue divergence maps, historical archives, and alerting are available under license: use the contact form or write to research@praetu.com.

Please tell us; corrections have their own topic in the contact form. Published snapshots are immutable, so a correction ships as a new dated snapshot with a note, never a silent edit.

A substantive human reply within two business days, usually faster. Every message is read by a person; see what happens after you write.

New York, working with teams anywhere. Engagements run remotely against your interfaces.