Replace the S2_10 map_reduce stage with load_fanout -> assess_bundle_* ->
capture_bundle_{same_ordinal} -> reducer so each LLM instance receives its
slot through preflight_files. Merge output and reasoning reserves into
max_output_tokens 128000, move the three runtime flags out of the input
fingerprint, and publish to the s2_10/v5 root. Flags stay False until the
canary probe.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
jhogyuandClaude Fable 5.1
ab448fbf23