mixle.task.design_prior module¶
Persistent what-works prior over structural design families.
DesignModel already ledgers every evaluated design point with an arbitrary
tag dict, so a caller can tag each accepted structural recipe with which FAMILY it belongs to (a
quotient leaf vs a plain head, a richer factorization vs a simpler one, or
capability-conditioned recipes). This module is the thin
query layered on that existing ledger: rank the families seen so far by their mean recorded quality,
so the NEXT round’s structural-search proposal starts from a sharper prior instead of from scratch –
“what has actually worked” persisted across rounds and design searches, not re-derived each time.
record_accepted_recipe(design, point, quality, violations, family=”quotient_leaf”) rank_design_families(design) # [(“quotient_leaf”, 0.91), (“plain_head”, 0.62)] best_family(design) # “quotient_leaf”
An untried candidate family (never recorded) has no evidence and ranks below every family that has
been recorded, via an explicit, named default_score – never silently tied with a proven winner.
- best_family(design, *, tag_key='family')[source]
The single top-ranked recorded family, or
Noneif nothing has been recorded yet.
- rank_design_families(design, *, tag_key='family', candidates=None, default_score=float('-inf'))[source]
Rank every family tag recorded in
designby its mean quality, best first.candidates, if given, are ALSO included in the ranking even if never recorded – an untried family getsdefault_score(-infby default: no evidence ranks strictly below any recorded family, however weak, rather than being silently omitted or tied with a proven winner).
- record_accepted_recipe(design, point, quality, violations, *, family, fingerprint=None, **tag)[source]
Record an accepted structural recipe under its
familytag – the training signal forrank_design_families(). A thin, named wrapper overDesignModel.addso callers do not have to remember which tag key the prior reads.