OMNIA Cuts False Accepts, Adds Reviews

The claim is directionally interesting, but it is only defensible if you keep it tightly framed as a bounded damage-proxy result, not a general safety or deployment claim. The layered-policy framing is reasonable; the weak point is that this is still a tiny, hand-curated eval with no evidence yet that the added review load is worth it outside the sandbox.

• –The baseline-vs-OMNIA split is the right framing only if the baseline is explicitly frozen and well-defined; otherwise the comparison is too easy to game.
• –`8 -> 1` on `n=15` is a strong signal, but it is statistically fragile without a held-out set, confidence intervals, and ablations against simple heuristics.
• –False-accept reduction is a valid external proxy if the downstream cost of a bad accept is high, but you also need review precision, reviewer burden, and latency/cost to judge net value.
• –The fastest serious next step is a preregistered, frozen eval with blind labels, stronger baselines, and a cost curve that shows when OMNIA beats simpler structural gates.
• –To make this harder to dismiss as sandbox-only, publish the exact dataset, scoring script, and failure cases, then invite independent reruns on unseen outputs.