Clarify which runs use which key

KEYS.md

@@ -4,10 +4,10 @@ These are the keys under which the texts in this dataset were generated, release
 reproduce the detection results. They are research keys used only for this evaluation: anyone holding them can produce
 text that these detectors flag under these keys, so they must not be reused to watermark real content.
 
-| splits | key `k` |
+| runs | key `k` |
 |---|---|
-| `thicket`, `gumbelmax`, `synthid`, `kgw`, `unigram`, `sweet`, `ewd`, `kgw_delta*` (and their C4 / multilingual / other-model runs) | `965853231` |
-| `textseal`, `none` (reference runs; `none` carries no watermark) | `1777` |
+| every scheme on every dataset and model (ELI5, C4-RealNews, multilingual-1000, Llama / Gemma / Mistral), including TextSeal's C4, multilingual-1000 and other-model runs | `965853231` |
+| the original TextSeal and no-watermark reference runs on the main ELI5 protocol (splits `textseal` and `none` of the configs `main`, `dilution_answers`, `multilingual`, `benchmarks`; `none` carries no watermark) | `1777` |
 
 How the integer key is consumed (the evaluator first reduces it to `k mod (2**31 - 1)`):
 - **Thicket** — generator `build_generator(model, tokenizer, k)` and detector `build(None, tokenizer, proxy_model, k)`;

keys.json

@@ -13,5 +13,6 @@
   "textseal": 1777,
   "none": 1777
  },
- "note": "integer key k given to the evaluator for each scheme (all models and datasets of that scheme); see KEYS.md"
+ "note": "integer key k given to the evaluator for each scheme (all models and datasets of that scheme); see KEYS.md",
+ "rule": "Every run of every scheme uses key 965853231, on every dataset and model (ELI5, C4-RealNews, multilingual-1000, Llama/Gemma/Mistral), except the original TextSeal and no-watermark reference runs on the main ELI5 protocol (splits `textseal` and `none` of the configs main, dilution_answers, multilingual, benchmarks), which used key 1777."
 }