Max-Accuracy mode
An optional per-project mode that escalates the hardest claims to a local reasoning judge — higher accuracy on the boundary cases, still no data egress.
A clear picture of when to turn Max-Accuracy on, what it changes about your scores, and how it stays inside your environment.
Every score in variA/Bly is deterministic by default — the same input produces the same score, every run. Max-Accuracy mode adds an optional second pass: on the small share of claims the deterministic engine is genuinely unsure about, it escalates to a local reasoning-model judge for a closer read. You turn it on per project, and none of it sends your data anywhere.
What it changes
Deterministic grounding is fast, reproducible, and right the large majority of the time. But a minority of claims — roughly the hardest 15–20% — sit on the boundary: a paraphrase, a multi-hop inference, a number that is almost supported. Max-Accuracy routes only those uncertain claims to a reasoning judge that weighs the evidence more carefully, then folds the verdict back into the score.
- The deterministic core is unchanged — the other ~80% of claims score exactly as before.
- Only the uncertain minority is escalated, so the judge stays off the hot path.
- Escalated claims are marked in the result (a Local judge pill), so you can always see which verdicts came from the reasoning pass.
Determinism, stated honestly
Deterministic mode is bit-for-bit reproducible. A reasoning judge is not — a local model has hardware-dependent floating point, so a Max-Accuracy score is replayable and audit-logged but not byte-identical across different hardware. That trade-off is disclosed when you enable the mode, and Deterministic mode stays available unchanged whenever bit-for-bit reproducibility matters more than the last few points of accuracy on the hardest claims.
Turning it on
Max-Accuracy is a per-project setting — one project can run it while another stays fully deterministic. Enable it under Settings → Grounding Mode. Turning it on records who enabled it and when, so there are no silent changes; from then on that project's evaluations use the escalation pass automatically, with no SDK change.
No data egress
The reasoning judge runs inside the same environment as the rest of scoring — on the hosted platform for hosted customers, and entirely on-node for self-hosted deployments. It makes no outbound calls, so Max-Accuracy is safe for regulated and air-gapped workloads. (This is separate from the optional external LLM-as-judge overlay, which is off by default and is the only path that would send text to a third party.)