All of these are built + merged this session but default OFF — they need dogfood validation against the real eval instrument (#harness/#data) before flipping, and a cost decision where noted:
Task: for each, measure new-path vs current on real mail, then set in render.yaml. Founder ops, not a code change.
All of these are built + merged this session but default OFF — they need dogfood validation against the real eval instrument (#harness/#data) before flipping, and a cost decision where noted:
JUDGE_INCLUDE_BODY— body into the judge (biggest PUSH-signal gain; eval must have bodies to validate)LEARNED_RULES_IN_JUDGE,SENDER_TRAITS_IN_JUDGE— per-user learned layers (need override volume)JUDGE_ESCALATION_MODEL— cheap→strong dial (cost decision)EMBEDDING_MODEL— semantic correction re-ranking (feat(judge): semantic re-ranking of correction few-shots (free, local-by-default) #642); free via local Ollamanomic-embed-textAUTO_TIER_EXECUTION— AUTO tier executes (feat(firewall): opt-in AUTO-tier execution — wire AUTO to the autonomous agent #643); only after AUTO calibration is provenTask: for each, measure new-path vs current on real mail, then set in render.yaml. Founder ops, not a code change.