Every AI tool measures the model's output. Syntonic measures the coupling: the human-and-AI as one unit.
Measure and optimize the quality of the working relationship between your people and your AI: the signal no output-eval can see.
Worse than the best of either alone.
Across 106 studies, human-AI combinations on average performed worse than the best of either alone.
(Nature Human Behaviour, 2024)
19% slower.
Experienced developers were 19% slower with AI in a randomized trial.
(METR, 2025)
+55.8% faster.
Task completion in a randomized trial with well-designed AI assistance.
(GitHub Copilot RCT, 2023)
The spread between −19% and +55.8% isn't the model. It's the pairing, and nothing in your stack measures the pairing. Most human-AI pairings underperform, and no tool on the market tells you whether yours is one of them.
LOCI is the engineering fix to that documented failure: one score for whether the working relationship is actually working. The bar it's built against is the published one, Complementary Team Performance: the pairing has to beat the best of either member alone, not the average. We measure that. We don't assume it.
Code review is where AI errors get expensive: it passed the tests, it passed review, and it still blew up in prod. CEED (coupling-emergent error detection) instruments the code-review process itself. When a result warrants a second look, it gets one, and every flag is tracked against confirmed errors so you can see how well it's working.
"We instrument your code-review loop with a coupling layer and measure, against a pre-registered criterion, whether it catches what your stack misses. You keep the readout either way."
The criterion is registered before we start (flagged outputs wrong at ≥2x your base rate), and it's stated up front because the experiment design is the offer. No catch-rate claims here: your pilot produces your number, on your codebase, not a public leaderboard.
A score for the relationship, not the output.
The only composite we know of that treats the exchange, the human-and-AI as one unit, as the unit of analysis. One normalized number for coupling quality over time: on course, in sync, transferring meaning without loss.
Status: validation in progress.
The evaluator the model can't game.
The score is produced from outside the evaluated system's prompt and state, on a different model family. We built the architecture the field's own findings imply is needed. For researchers: scalable oversight, scored from outside the system prompt.
Status: robustness experiment in progress.
Errors the pairing surfaces that neither party catches alone.
Coupling measurement in practice: our flagship demonstration, deployed first on code review.
Status: design-partner pilots open.
| Output evals (the existing category) | Syntonic | |
|---|---|---|
| What it measures | The AI's output | The coupling: the human-and-AI as one unit |
| The question it answers | "Is the model right?" | "Is the collaboration working?" |
| Who uses it | The AI engineering team | The team doing the work, and its leadership |
To our knowledge, no commercial product treats the human-AI pair as the unit of analysis.
Results ship when experiments land, not before.
Patent pending (nonprovisional submitted for filing, June 2026).
A pre-registered randomized comparison: equal task-success at fewer tokens-to-resolution, against a strong published baseline class. Numbers ship when the study lands.
Ship bar: agreement with human raters at Cohen's kappa ≥ 0.6. Strong public claims reserved for ≥ 0.8. The protocol publishes first.
Preprint: coming.
Research updates. That's all this is.