Update README (phaseF: Claude Fable 5.2, 855 trajectories; writer GPT-6.1 Sol)

README.md

@@ -865,3 +865,18 @@ reads only the highest `phaseN/` folder of a task (its DAG with its own trajecto
   `provenance` says `engine: claude`, `model: claude-sonnet-5-5` and carries a `writer_note`.
 - Layout: `phaseR/<suite>/<task>/trajectories/<runId>.json`; the DAG stays `phase2/<suite>/<task>/task_dag.json`.
 - Status: 630 of 630 trajectories uploaded (269 tasks) (complete).
+
+## Phase F: Claude Fable 5.2 (2026-11-12)
+
+- Runs: the counted attempt of Claude Code with Claude Fable 5.2 (run `claude_code-fable_5_2-high`, default seed) on
+  each core test (privileged, standard and realtime), successes and failures alike, the same rule as for the other
+  agents. Tests whose task has no phase-2 DAG are not labelled (no DAG is written here): embodiedswe, roboengineer and
+  two newer Embodied Olympics - Match tasks.
+- Writer: GPT-6.1 Sol (`gpt-6.1-sol`, high effort) in Codex; reviewer: Claude Sonnet 5.5 in Claude Code. The same
+  writer and reviewer as phase 2; prompts v3.2, the task's phase-2 DAG as published under `phase2/` (not rewritten;
+  `dag_sha256` is that file), the same task inputs and oracle rule, review rounds and validation. Each file's
+  `provenance` carries a `writer_note` and `writer_access` (ChatGPT subscription or Azure OpenAI API, same model).
+- Videos: the graded `replay.mp4` of the run viewer dataset `MLL-Lab/embodied-exam-runs`; runs without one (privileged
+  runs that submitted nothing) are labelled from the log only.
+- Layout: `phaseF/<suite>/<task>/trajectories/<runId>.json`; the DAG stays `phase2/<suite>/<task>/task_dag.json`.
+- Status: 855 of 855 trajectories uploaded (427 tasks) (complete).