Skip to main content

Bidirectional model fit — escalate and downgrade

· One min read
Anchor maintainers

Anchor already stopped too-hard work with a first-line SUGGEST-ESCALATE. Premium sessions still burned credits on too-easy work with only a soft “say so and ask.”

What shipped

DirectionTokenRule
Too hardSUGGEST-ESCALATE: <target> — <reason>mythos-core 11
Too easySUGGEST-DOWNGRADE: <cheaper> — <reason>mythos-core 10
Good fit(silence)no model pitch

orchestrate.py treats both as fit gates alongside SUGGEST-REROUTE for specialty mismatch (no retry burn; --insist proceeds) — see Dual-axis model fit for the specialty side. Interactive platforms (Claude, Grok, Chat, local) state the same standing check. Downgrade heuristics stay conservative — rename/format/boilerplate, not “any multi-file mid work.”

See Model fitness and Doctrine.