calybris1.3.0One order, four suppliers
The kernel rejects at hard limits, then ranks what survives on one utility number. Move the order or a price and it decides again.
What would it take?
Pick a supplier that lost. The kernel is run again on a copy of the policy with one of that supplier's numbers moved, and the boundary is searched for: the smallest single change after which it wins.
Is this decision really in the log?
Every decision is written to a hash-chained, HMAC-keyed WAL. 1.2.0 adds RFC 9162 Merkle proofs over it: one record is checked without the rest of the log, and yesterday's published head is proven to be a prefix of today's.
Would another supplier have done better?
The kernel always picks the same winner, so on its own it never learns what the others would have done. A keyed, replayable draw takes a small share of near-best alternatives and records exactly how likely each was; an off-policy estimate then scores each supplier as if it had been chosen every time.
These are preview features: released and tested, outside the 1.x stability promise until reviewed. Every 1.0 decision and digest is unchanged.