refactor: update about tab with experimentation details

ui.py

@@ -50,12 +50,13 @@ there is no single score across InferBench.
 
 ### OneIG Alignment
 Prompt-image **alignment** on anime / stylization, human / portrait, and
-general object prompts. This is the alignment slice of OneIG, not the full
-suite. Default rank is the mean of the alignment categories that exist for
-that row.
+general object prompts (100 prompts each). This is the alignment slice of
+OneIG, not the full suite. Default rank is the mean of the alignment
+categories that exist for that row.
 
 ### Qwen Image Bench
-A shared prompt set used to compare endpoints on the same prompts.
+100 prompts from the 1,000-prompt Qwen Image Bench set, sampled for coverage
+across its fine-grained (L3) categories.
 """
 
 ABOUT_DETAILS_CONTENT = """
@@ -69,8 +70,9 @@ ABOUT_DETAILS_CONTENT = """
   not a pairwise Elo. It is not a text-to-image model on the leaderboard. Use
   it alongside Datapoint and Rapidata Elo, not instead of them.
 - **Datapoint Elo**: human-preference Elo from Datapoint pairwise comparisons.
-- **Rapidata Elo**: human-preference Elo from the Rapidata evaluation on the
-  same suite.
+- **Rapidata Elo**: human-preference Elo from Rapidata pairwise comparisons.
+  Rapidata rejects prompts over 400 characters, so this Elo is on a subset
+  of each suite (see Setup).
 - **Generation time**: median and minimum generation time in seconds, as
   reported in the evaluation table. This is not a p95, and we do not state
   warm vs cold or concurrent load.
@@ -83,17 +85,30 @@ models *within* a column.
 
 ## Setup
 
-- **Evaluation window:** —
+- **Evaluation window:** July–August 2026.
 - **Update policy:** numbers come from evaluation snapshots in the tables,
   not a live API poll.
-- **Prompt counts:** OneIG —, Qwen —.
+- **Prompt counts:** OneIG Alignment uses the first 100 prompts from each of
+  the three categories (300 total). Qwen Image Bench uses 100 prompts sampled
+  from the 1,000-prompt pool for roughly even coverage of its fine-grained
+  (L3) categories.
 - **Generation:** one image per prompt per endpoint when the run exists.
-  Resolution, steps, CFG, seed policy, prompt rewrite, and safety filters: —.
-- **Pairwise comparisons (Datapoint / Rapidata):** —
+  Default resolution is 1024×1024. Exceptions: FLUX 1.1 Pro Ultra at 2K,
+  FLUX 2 Flex at 1008×1008, and any endpoint labeled 2K. The seed is derived
+  from the prompt, so every model gets the same seed for the same prompt.
+  Steps, CFG, prompt rewrite, and safety filters follow each endpoint's
+  default.
+- **Datapoint:** every model pair is compared on every prompt, with 10 votes
+  per battle.
+- **Rapidata:** prompts longer than 400 characters are dropped, leaving 212
+  OneIG prompts and 85 Qwen Image Bench prompts. 4 votes per pair; about
+  26,000 votes on OneIG and 35,000 on Qwen Image Bench.
 
 ## Limits
 
 - Empty cells mean that track was not run or not reported for that model.
+- Rapidata Elo is not on the full prompt suite, so it is not directly
+  comparable to Datapoint Elo even on the same benchmark.
 - Elo ratings can shift when the comparison pool changes: treat them as
   relative rankings for the snapshot, not absolute constants.
 - Close scores can be a tie in practice; the table does not show confidence