temporal_causal/XS-discovery: update model card

models/temporal_causal/XS-discovery/README.md

@@ -6,6 +6,9 @@ cause which, at which lag, from the network's own outputs: a graph-level head sa
 the top `round(sum of pair probabilities)` pairs are marked. Nothing is fitted on the request and there is no
 statistical fallback.
 
+**License: non-commercial use only.** The weights are released under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/);
+commercial use needs a separate license from ProfitOps Inc. See [License](#license) below.
+
 ## What XS is good for
 
 XS is the small, fast entry point: it answers on a CPU in seconds and on a GPU in a fraction of a second, with no
@@ -25,20 +28,19 @@ many steps, before investing in a deeper analysis.
 
 It is strongest on:
 
-* **Process and operations series:** controlled process data on a regular clock (AUROC 94 %, F1 79 % on our
-  development set).
-* **Retail and demand series:** daily sales, prices, promotions and traffic (AUROC 88 %, F1 70 %).
+* **Process and operations series:** controlled process data on a regular clock.
+* **Retail and demand series:** daily sales, prices, promotions and traffic.
 * **Mixed column types:** continuous, count, binary, ordinal and bounded columns, declared with `col_types`.
 * **Lagged effects:** a cause that acts after a delay of several steps.
 
-Use it with care on business-style tables with many weak links (AUROC 79–84 %, F1 41–47 %), and expect weaker answers
+Use it with care on business-style tables with many weak links, and expect weaker answers
 on coarsely sampled or irregular series, long delays (20 to 60 steps), slow hidden trends and same-step links between
 count or binary columns. Effect sizes, identifiability and root cause come with the larger sizes and later XS
 releases.
 
 **Need higher precision or F1?** XS is the fast first look. For higher precision, recall and F1 and for harder
 series, use the S, M and L sizes, coming soon; each release publishes its own measured numbers, so you can choose the
-size by the accuracy your decision needs.
+size by the accuracy your decision needs. To ask about them, contact ProfitOps Inc.
 
 ## Use
 
@@ -117,18 +119,34 @@ For an irregular clock, pass `time_gap`: the gap before each row, in steps of th
   (8 rows). It does not estimate effect sizes, identifiability or root causes.
 * **Known weak spots:** coarsely sampled and irregularly spaced series; long delays (20 to 60 steps), slow hidden
   trends, state memory and delays that vary over time; contemporaneous links between count or binary columns. Control
-  loops, sensor read-through effects and industrial process mechanisms were not part of its training.
-
-## Measured on a development benchmark (files inspected during development, not a holdout)
-
-AUROC / F1 in %, every temporal file sent, unanswered scored as 50 / 0:
-
-| set | files | AUROC | F1 |
-|---|---:|---:|---:|
-| controlled process series | 72 | 94.3 | 79.1 |
-| retail daily series | 26 | 88.4 | 69.6 |
-| business-style files answered by every compared checkpoint | 103 | 83.6 | 47.2 |
-| business-style files, all sent | 118 | 79.3 | 41.2 |
+  loops, sensor read-through effects and industrial process mechanisms were not part of its training. For these, the
+  larger sizes handle more (see below).
+
+## Real-world discovery, against classical baselines
+
+Run on six published real-world datasets used for benchmarking only (never
+seen in training), through the model's own decision layer (pure-neural: the
+graph-level "any edge" head, then the top `round(sum of pair probabilities)`
+pairs). Classical baselines run on the same files with the same settings.
+AUROC / F1 as fractions (0 to 1). In each row, the best AUROC and the best F1 across the three methods are in
+**bold**.
+
+| dataset | XS AUROC | XS F1 | Granger AUROC | Granger F1 | PCMCI AUROC | PCMCI F1 |
+|---|---:|---:|---:|---:|---:|---:|
+| Causal Chamber | **0.72** | **0.34** | 0.67 | 0.19 | 0.66 | 0.00 |
+| Sachs | **0.56** | **0.32** | 0.48 | 0.23 | 0.46 | 0.00 |
+| Tennessee Eastman | 0.70 | **0.16** | 0.60 | 0.03 | **0.71** | 0.08 |
+| Causal Rivers | 0.65 | **0.09** | **0.71** | 0.01 | 0.58 | 0.00 |
+| causRCA | **0.60** | 0.05 | 0.54 | 0.04 | 0.54 | **0.06** |
+| SWaT | 0.36 | 0.00 | 0.64 | 0.07 | **0.65** | **0.10** |
+
+Rows are ordered by how well XS does, best first. XS has the best F1 on four of
+the six datasets and the best AUROC on three, and it wins both on Sachs and
+Causal Chamber. It trails on SWaT, a water-network dataset where it marks no
+edges. These are real industrial and scientific panels; absolute scores are
+modest, as the card's limits note for untrained industrial families. Read them
+against the classical baselines on the same files, not as a fraction of a
+perfect graph.
 
 ## Files