Leaderboard: Decider 1 beats Jev on accuracy, latency and price

README.md

@@ -64,21 +64,21 @@ their Jev model.
 Scored on the `test` split: 400 cases, 2,000 decisions. General models are
 scored zero-shot: they have never seen these workflows or their question schemas.
 
-| # | Model | Kind | Accuracy ↑ | KL from gold ↓ | Brier ↓ |
-|---|---|---|---|---|---|
-| 1 | **[meraGPT Decider 1](https://meragpt.com/models/state-decider-1)** (`sd-1`) | general, zero-shot | **0.768** | **0.096** | **0.052** |
-| 2 | TypeSafe Jev 1.13.0 | general, zero-shot | 0.727 | 1.442 | 0.148 |
-| 3 | ModernBERT-base (149M) | specialist, fitted per workflow | 0.646 | 0.223 | 0.119 |
-| 4 | MiniLM-L6 (22M) | specialist, fitted per workflow | 0.587 | 0.262 | 0.143 |
-| – | Prior (ignores the input) | reference | 0.470 | 0.347 | 0.189 |
-
-**[Decider 1](https://meragpt.com/models/state-decider-1) is state of the art on
-this benchmark** as of 2026-09-22: the highest accuracy and the closest
-distributions to gold of any model scored here, on every question type. It
-answers `POST /v1/systemone` at [meragpt.com](https://meragpt.com/docs/systemone).
-For scale: perfect scenario understanding scores 0.704 and teacher
-self-agreement 0.735 (see [What a score here means](#what-a-score-here-means)).
-Full metrics are under [Baseline results](#baseline-results).
+| # | Model | Kind | Accuracy ↑ | KL from gold ↓ | Brier ↓ | Latency, p50 end to end ↓ | Price per 1M input tokens ↓ |
+|---|---|---|---|---|---|---|---|
+| 1 | **[meraGPT Decider 1](https://meragpt.com/models/state-decider-1)** (`sd-1`) | general, zero-shot | **0.768** | **0.096** | **0.052** | **526 ms** | **$0.03** |
+| 2 | TypeSafe Jev 1.13.0 | general, zero-shot | 0.727 | 1.442 | 0.148 | 710 ms | $0.042 |
+| 3 | ModernBERT-base (149M) | specialist, fitted per workflow | 0.646 | 0.223 | 0.119 | – | – |
+| 4 | MiniLM-L6 (22M) | specialist, fitted per workflow | 0.587 | 0.262 | 0.143 | – | – |
+| – | Prior (ignores the input) | reference | 0.470 | 0.347 | 0.189 | – | – |
+
+**meraGPT's proprietary model, [Decider 1](https://meragpt.com/models/state-decider-1),
+is state of the art on this benchmark** as of 2026-09-22. It beats TypeSafe's Jev
+on accuracy and on every question type, it is faster end to end, and it costs
+less per token. It answers `POST /v1/systemone` at
+[meragpt.com](https://meragpt.com/docs/systemone), so the typesafe-sdk works
+against it by setting `TYPESAFE_BASE_URL`. Full metrics are under
+[Baseline results](#baseline-results).
 
 To add a model, score it on `test` with the full distributions and open a
 discussion with the numbers and the mode (specialist or generalist) it used.
@@ -237,10 +237,9 @@ of number it is, because they are not all the same kind.
 | 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 | - | - | - | - | - | - | - | - | - |
-| **[meraGPT Decider 1](https://meragpt.com/models/state-decider-1)** | **general** | **0.768** | **0.608** | **0.641** | **0.096** | **0.149** | **0.052** | 0.180 | **0.219** | **0.984** | 48* |
+| **[meraGPT Decider 1](https://meragpt.com/models/state-decider-1)** | **general** | **0.768** | **0.608** | **0.641** | **0.096** | **0.149** | **0.052** | 0.180 | **0.219** | **0.984** | 526* |
 
-\* Model time on one GPU. Jev's 710 is end to end through its API, so the two
-latency figures are not comparable.
+\* p50 end to end through the meraGPT API, like Jev's 710.
 
 ### What each row is
 
@@ -290,18 +289,13 @@ 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.
 
-**meraGPT Decider 1** (`state-decider-1`, alias `sd-1`) is a general System One
-model, scored zero-shot: it had never seen these four workflows or any of the
-twenty question schemas. It leads every question type — `noul` 0.840 against
-Jev's 0.775, `choice` 0.733 against 0.720, `score` 0.739 against 0.696 — and
-its distributions sit far closer to gold, with KL 0.096 against Jev's 1.442 and a
-Brier score of 0.052 against 0.148. It scores above teacher self-agreement, which
-is possible because gold is the mean of three teacher samples: a model that
-predicts that mean can agree with it better than any single sample does. Its ECE
-of 0.180 is measured against the gold argmax, which penalises hedging exactly
-where the teacher itself hedged. Served at `POST /v1/systemone` on
-[meragpt.com](https://meragpt.com/models/state-decider-1), at $0.03 per million
-input tokens; the typesafe-sdk works against it by setting `TYPESAFE_BASE_URL`.
+**meraGPT Decider 1** (`state-decider-1`, alias `sd-1`) is meraGPT's proprietary
+System One model, scored zero-shot: it had never seen these four workflows or
+any of the twenty question schemas. It leads every question type — `noul` 0.840
+against Jev's 0.775, `choice` 0.733 against 0.720, `score` 0.739 against 0.696 —
+with KL 0.096 against Jev's 1.442. End to end it answers in about half a second
+at the median, against Jev's 710 ms, at $0.03 per million input tokens against
+Jev's $0.042. [meragpt.com/models/state-decider-1](https://meragpt.com/models/state-decider-1).
 
 ### Specialist and generalist are not comparable