Built for the expensive era
Model time is subsidized today, the way early ride-hailing was: prices are set to win adoption, not to cover cost. The pattern that follows is known, not predicted. Once adoption is won, prices rise and each token buys less; enterprise contracts already show it. magus is built for that era, not this one.
The answer is to move what a model would otherwise have to work out (the graph, declared outputs, what a change can break, which conventions hold) into deterministic, checked answers. A smaller model can then do work that today takes a frontier one. A feature that only works with a frontier model is not finished. This prevents a tool that looks capable only while someone else pays for the reasoning.
Enforced by
- The model-tier lens: the economy tier's pass rate should rise over time.
- The static tier table in ADR 0009: a person picks the model for a tier, and nothing routes per call.
Good
One deterministic answer replaces a model reading files to rediscover it:
magus describe file docs/MAGUS.md
magus refs decodeFile --occurrences
Bad
Read every magusfile under libs/ and tell me which targets write docs/MAGUS.md.
The model spends tokens deriving a fact the graph already holds, and the answer is a guess a cheaper model gets wrong.