FAQ

Coupling is the human and the AI working as one unit: the thing we measure, instead of measuring either side alone. Coupling quality is how well that unit is actually working: staying on goal, transferring meaning without loss, catching errors. LOCI is the score: one normalized number for coupling quality over time. The instrument is in validation; see Is LOCI validated? for exactly where that stands.

Those tools measure the AI's output, and they're good at it. We measure the coupling: the human and the AI as one unit. Different object, different question: they answer "is the model right?"; we answer "is the collaboration working?" It sits beside your eval stack, not instead of it. To our knowledge, no commercial product treats the human-AI pair as the unit of analysis.

LLM-judge manipulation is a published vulnerability: attacks flip judge verdicts at more than 30% success (source). We did not discover this, and independent evaluator models already exist. Our mechanism: the evaluator scores from outside the evaluated system's prompt and state, on a different model family, because judges measurably over-rate their own model family (self-preference; NeurIPS 2024). We built the architecture the field's own findings imply is needed. Status: the adversarial-robustness experiment is in progress. Today this ships as a mechanism description, not a proven property.

The research roadmap is the honest answer: pre-registered studies with stated ship-bars, and a preprint that is coming. The bar we test against is the published one, Complementary Team Performance: a real effect means the team beats the best of either member alone, not the average. Results ship when experiments land, not before.

Not yet, and we won't act like it is. Ship bar: agreement with human raters at Cohen's kappa ≥ 0.6; strong public claims reserved for ≥ 0.8. The study is in progress; the protocol is what we publish first. Until it clears, the readout ships as an instrument in validation: a lens you watch mature, not a number you act on alone.

No. The best published evidence runs the other way: across 106 studies, human-AI combinations on average performed worse than the best of either alone (Nature 2024), and experienced developers were 19% slower with AI in a randomized trial (METR 2025). Our position: most pairings underperform, nothing tells you whether yours is, and LOCI is the instrument: the engineering fix. We claim gains only where they're measurably possible, and we measure rather than assume.

The shape: a measurement layer installed in one code-review workflow (~6-week deployment, no hardware); flagged outputs get a second-look pass; flagged-vs-confirmed correlation is tracked for 90 days; you get a per-review readout (instrument in validation). The success criterion is pre-registered before we start (flagged outputs wrong at ≥2x your base rate), and you keep the readout either way. Pilots are paid; scope and pricing are set per partner in the intake call. Data capture runs under contracted consent; specifics live in the SOW. The pilot, in detail →

No. The scoring and evaluation APIs are in validation, not yet available. There's no waitlist; the design-partner pilot is the early-access path. Roadmap statuses change when the work lands. Not before.

Patent pending. The nonprovisional was submitted for filing in June 2026. We don't publish claim details.

An instrument company in its design-partner phase. Two doors today: design-partner pilots (the product wedge) and done-for-you deployments (Work with us, available now). Every engagement feeds the same instrument development. The APIs and self-serve product are roadmap: coming, not for sale.