card: explain each reference row, add ceilings and a saturation estimate

README.md

@@ -201,39 +201,76 @@ generalist number, unlabelled, misleads the reader.
 
 ## Baseline results
 
-Two specialist baselines, fitted on the `train` split and scored on `test`. Both
-use [Adaptive Classifier](https://github.com/codelion/adaptive-classifier)
-0.2.0, one classifier per question, with the encoder frozen.
-
-| Model | Acc | Soft acc | Macro F1 | KL | TV | Brier | ECE | Score MAE | Within 1 level | ms/case |
-|---|---|---|---|---|---|---|---|---|---|---|
-| MiniLM-L6 (22M) | 0.587 | 0.506 | 0.424 | 0.262 | 0.267 | 0.143 | 0.108 | 0.515 | 0.864 | 22 |
-| ModernBERT-base (149M) | 0.646 | 0.542 | 0.469 | 0.223 | 0.249 | 0.119 | 0.179 | 0.444 | 0.931 | 349 |
-| Prior (input ignored) | 0.470 | 0.430 | 0.207 | 0.347 | 0.317 | 0.189 | 0.088 | - | - | 0 |
-| Uniform | 0.308 | 0.311 | 0.152 | 0.444 | 0.381 | 0.238 | 0.169 | - | - | 0 |
-
-Read these against the floor of 0.520 and the saturation point of 0.735.
-ModernBERT lands 12.6 points above the floor and 5.8 below what perfect
-scenario understanding would give.
+Everything below is scored on the `test` split. The first column says what kind
+of number it is, because they are not all the same kind.
+
+| Model | Kind | Acc | Soft acc | Macro F1 | KL | TV | Brier | ECE | Score MAE | Within 1 level | ms/case |
+|---|---|---|---|---|---|---|---|---|---|---|---|
+| Uniform | reference | 0.308 | 0.311 | 0.152 | 0.444 | 0.381 | 0.238 | 0.169 | - | - | 0 |
+| Prior | reference | 0.470 | 0.430 | 0.207 | 0.347 | 0.317 | 0.189 | 0.088 | - | - | 0 |
+| 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 | - | - | - | - | - | - | - | - | - |
+| Teacher self-agreement | ceiling | 0.735 | - | - | - | - | - | - | - | - | - |
+| A general System One model | estimate | 0.65-0.72 | - | - | - | - | - | - | - | - | - |
+
+### What each row is
+
+**Uniform** puts the same probability on every option. It reads nothing and
+knows nothing. It is here to anchor the KL and Brier scale: 0.444 is what no
+information costs in distribution terms.
+
+**Prior** fits each question's label frequencies on the `train` split, then
+answers those frequencies for every case, ignoring the state entirely. If 67% of
+`needs_human` golds are true, it answers 0.67 true every time.
+
+This is the row to check a learned model against. MiniLM beats it by 9 points
+and ModernBERT by 18. That gap is how much of each score comes from reading the
+input rather than from label frequency. A model that cannot clear it has learned
+nothing about the state, which is the failure that sank the v0.1 prototype.
+
+Prior also has the best ECE on the table, at 0.088, while knowing nothing.
+Guessing the base rate is perfectly calibrated by construction. That is the
+clearest argument for reading KL and Brier here instead of ECE.
+
+**Perfect scenario understanding** is what a model would score if it recovered
+the latent factors that generated each case exactly. Measured by fitting those
+factors to the gold labels with cross-validation. It is optimistic, since the
+factors are more than the text reveals.
+
+**Teacher self-agreement** is a fresh teacher sample scored against gold built
+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.
+
+### Specialist and generalist are not comparable
+
+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
+the better model, since anything it gets right where the teacher is wrong counts
+against it.
+
+### Reproducing the specialist rows
+
+Both use [Adaptive Classifier](https://github.com/codelion/adaptive-classifier)
+0.2.0, one classifier per question, encoder frozen. Configuration was tuned on a
+held-out quarter of `train` and never on `test`: mean pooling, `max_length` 512,
+30 epochs, `prototype_weight` 0.3.
 
 The gap between the two encoders is the trade-off this benchmark exists to
 measure. Six points of accuracy cost 16x the latency.
 
-The prior row is the one to check a learned model against. It fits each
-question's label frequencies on `train` and never reads the state. MiniLM beats
-it by 9 points and ModernBERT by 18, which is how much of each score comes from
-actually reading the input rather than from label frequency.
-
-It also has the best ECE on the table, at 0.088, while knowing nothing. Guessing
-the base rate is perfectly calibrated by construction. That is the clearest
-argument for reading KL and Brier here rather than ECE.
-
-Both learned rows are specialists. Neither can answer a question it was not
-fitted for, so neither can be scored zero-shot.
-
-Configuration, tuned on a held-out quarter of the `train` split and never on
-`test`: mean pooling, `max_length` 512, 30 epochs, `prototype_weight` 0.3.
-
 One harness detail matters for reproducing these. Adaptive Classifier trains on
 hard labels, so the gold distribution is normally thrown away at fit time. Each
 case is instead entered four times, apportioned across labels in proportion to