Replace the estimated System One row with a measured Jev 1.13.0 result

README.md

@@ -211,8 +211,8 @@ of number it is, because they are not all the same kind.
 | MiniLM-L6 (22M) | specialist | 0.587 | 0.506 | 0.424 | 0.262 | 0.267 | 0.143 | 0.108 | 0.515 | 0.864 | 22 |
 | ModernBERT-base (149M) | specialist | 0.646 | 0.542 | 0.469 | 0.223 | 0.249 | 0.119 | 0.179 | 0.444 | 0.931 | 349 |
 | Perfect scenario understanding | ceiling | 0.704 | - | - | - | - | - | - | - | - | - |
+| **TypeSafe Jev 1.13.0** | **general** | **0.727** | 0.580 | 0.613 | 1.442 | 0.251 | 0.148 | 0.144 | 0.391 | 0.952 | 710 |
 | Teacher self-agreement | ceiling | 0.735 | - | - | - | - | - | - | - | - | - |
-| A general System One model | estimate | 0.65-0.72 | - | - | - | - | - | - | - | - | - |
 
 ### What each row is
 
@@ -242,12 +242,25 @@ factors are more than the text reveals.
 from the other samples. It is the noise floor of the labelling process. Scoring
 far above it means predicting the teacher's quirks rather than the task.
 
-**A general System One model** is an estimate, not a measurement. We have no
-access to one and have run nothing. The range comes from the two ceilings above:
-such a model should read these scenarios about as well as anything can, which
-puts it near 0.704, and it cannot reliably pass 0.735 because the remaining gap
-is teacher noise it has no way to predict. Treat it as a reading of where the
-benchmark saturates, not as a result.
+**TypeSafe Jev 1.13.0** is a measurement, taken on 2026-09-18 through the
+TypeSafe API (`POST /v1/systemone`, `model: jev-latest`, which reported itself
+as `jev-1.13.0`). All 400 cases, all 2,000 decisions, zero errors, p50 710ms per
+case, $0.016 total at the published $0.042/1M input rate. Earlier revisions of
+this card carried an estimated range here instead; that estimate is gone.
+
+Jev scores **0.727 against a 0.735 ceiling**, so it has effectively saturated
+this benchmark. It also clears the 0.704 factor ceiling, meaning it reads these
+scenarios better than a model that recovers the generating factors exactly.
+
+**Its distributions are a different story.** Jev's KL from gold is 1.442 against
+ModernBERT's 0.223 -- six times worse -- while scoring 8 points higher on
+accuracy. Jev picks the right label and commits to it; the specialist is right
+less often but its uncertainty tracks the teacher's spread much more closely.
+Jev is not badly calibrated in absolute terms (ECE 0.144, overconfidence
++0.023); it is confident because it is usually correct. The KL gap is mostly
+that this gold is a three-sample teacher spread and Jev does not reproduce that
+spread. Which number matters depends on whether you consume the argmax or the
+distribution.
 
 ### Specialist and generalist are not comparable
 
@@ -255,9 +268,11 @@ Both learned rows are specialists, fitted on the `train` split of the same four
 workflows they are scored on. Neither can answer a question it was not fitted
 for, so neither can be run zero-shot.
 
-A general model scoring 0.70 against a specialist's 0.646 has not beaten it by
-six points. It would be answering question schemas it has never seen, which the
-specialists cannot do at all. A general model can also score lower while being
+Jev at 0.727 against the specialist's 0.646 has not beaten it by eight points.
+Jev answered all twenty question schemas cold, having never seen this benchmark;
+the specialists were fitted on the `train` split of the very workflows they are
+scored on and cannot answer anything else at all. Read the gap as the price of
+generality, not as a quality ranking. A general model can also score lower while being
 the better model, since anything it gets right where the teacher is wrong counts
 against it.