Write it as a no with two things attached, the eval number and the date you re-test when the next model lands, after one test. In Colin Jarvis's October 22, 2026 account of OpenAI's FDE team, a telecom voice deployment scored around 60% on the key evals at scoping and reached 80% and then 90-plus over six months because Research shipped model snapshots against the customer's requirements. That is a column only a lab has. The handshake every FDE contract should carry sorts each promised outcome into ships, builds, or no, and a 60% belongs in no unless the harness probe moves it: hold the model, change the harness, and rerun the customer's rows. A 60% that survives a harness change is a model gap. A 60% that moves was never one, and goes in builds with a date.
A 60% eval at scoping is a no with an eval number and a re-test date, unless a harness change moves it.
The case, the two scorecard lines, and why the fourth column belongs to the lab are in Two Months, Then One Day: OpenAI Published the FDE Scorecard. The handshake's three columns come from In the Room, Not on the Quota: Who Owns the FDE Promise. The probe to run before calling a gap a model gap is in Which Model Is the Smallest Knob in the Architecture Review.
Last reviewed 2026-10-22 · 1 min read