Things I build and try out.
A methods write-up of the CRAP-score reduction loop — why the coverage term dominates the metric, why the threshold is strict, and the three verification layers that caught what the tests missed.
Benchmarking a 26M single-shot router against a 1B dispatch model for switchboard's semantic routing tier, and settling on confidence-gated escalation instead.
Rebuilding the humanness critic from anti-correlated to useful, and discovering it has detection ability but not ranking ability — a valid gate and an invalid reward.
Why the shipped humanness critic ran below chance, how it punished real writers for using em-dashes and triads, and which detectors actually reproduced the blind judge.
Building a modern-prose reference register to test whether the judge's tell was AI-ness or epoch, and finding the register swap alone did not close the null.