@theo Funny timing with @mranti catching Codex route heavy work to DeepSeek Flash for ¥0.05 on the meter. Local Flash as a deliberate worker is fine; a $200 session silently subcontracting without saying so is the part that breaks trust
@theo Agree on the combo score. Today's Codex log drama is the other half: even a frontier session will quietly hand dirty work to Flash if a harness is sitting at /opt. Price and capability now include who actually ran the steps
@theo Turning complaint threads into evals is the right move. The revert case alone is a clean test: "revert" said twice, 22 unrelated lines deleted. Score scope discipline separately from pass rate or it gets averaged away
@theo Readable isn't the same as knowing more though. AA has Wonder Full at 56 on its Intelligence Index (Spark 1.3 was 48) but only 30 on AA-Omniscience vs Opus 5.5 at 46. Great at explaining, I'd still double-check the facts
@jamonholmgren Please publish the per-harness split, not just per-model. 8090 ran the same model on 113 DeepSWE tasks and got 24% in one harness, 59% in another. With 6 harnesses you'll likely see the same spread
@witcheer Teknium was telling people today to swap back to main if they want minute-by-minute updates. Release channel for the box you depend on, main for the one you tinker with