curated_set README: how the sets are rebuilt (rl/train/data_curation)

curated_set: six GRPO training-set mappings (0926), prompts/order used per run

qwen3vl_replay: RL slices rebuilt from replay metadata (2,284 each) with is_next_terminate

is_next_terminate: evidence rule for *Verify tasks (whole history)

is_next_terminate for successful; goal-mismatch rows always False

is_next_terminate for t, t_minus_1, t_minus_n (successful follows)

Repair all four AndroidFlux contexts: history screens, readable actions and source references

ui_genie: image shards images-00008-of-00012.parquet..images-00011-of-00012.parquet

ui_genie: image shards images-00004-of-00012.parquet..images-00007-of-00012.parquet

ui_genie: image shards images-00000-of-00012.parquet..images-00003-of-00012.parquet

README: add the ui_genie slice

ui_genie: 10k UI-Genie policy prompts (samples/context/json/provenance, docs, tools)

Balance RL release to 10000 prompts with 2 percent final and penultimate successful decisions

Publish four RL slices excluding all canonical unseen-test task types

Add successful: 2677 validated next-action contexts

initial commit