zeekay 8447bd1851 decision: kai-1's harness answers, and Kai asked over HTTP ...

results/kai-1 holds the answers the harness rows of the Decision Suite
were scored from: kai-1 (a7-512, weights 0834a74f) through bench preds on
CUDA in BF16, 11,099 of 11,099 questions, its model field a name in place
of a local path. merge.py takes --kai, --jev and --out; with --kai
results/kai-1/preds.json.gz it writes the scores every published Kai and
Jev harness figure equals, and with no flags it writes the same bytes it
did. --jev recorded, the default, is Jev's committed preds.

run_kai.py sends results/states.json.gz to POST /v1/decisions,
api.hanzo.ai or a local kai serve, and writes preds in Jev's shape.
digest.py prints each suite's digest of a preds file as a research run's
notes record it; all 24 harness runs of kai-1 and Jev match.

The cache directories are ignored and the two committed .pyc files go.

2026-10-15 10:43:17 -07:00

1.5 MiB

Raw Permalink History

View Raw