accept-clef-bf16-interim-spark (plan orin-decision-models, t17 — stage A only)
date: 2026-10-14 (UTC)
box: DGX Spark GB10 (128 GB unified)

SCOPE: stage A only. Stage B (routing decider:general through the Spark,
Orin and Thor gateways) was DROPPED by approved deviation d6: Clef is kept
only as the W4A16 build on the Jetson AGX Orin, served through the orin-decider
shape (t9) and its live acceptance (t11); the Spark interim is stopped after
t11 passes. The acceptance criterion "answers through the Spark gateway and,
via the mesh, from the Orin and Thor gateways" was therefore NOT exercised
here and is not claimed.

== what was served ==
service    decider-general (container model-gear-decider-general), added via
           ~/.lobes/docker-compose.override.yml (the Spark compose is hand-kept)
backend    DECIDER_BACKEND=clef (Cloudflare's own transformers code and
           joint_schema_model.py from the checkpoint)
model      Cloudflare/clef @ 2f3de3dd85f379784083b0814d997ab627200f0c (bf16),
           mounted read-only from a verified local snapshot
image      lobes-decider:odm-local
           sha256:bd3c10347acaaeb6cda87ab5f4f6f644ebcce3e7cd729480a62470fec1fc7df2
           (branch wheel on vllm/vllm-openai@sha256:c1c9f6fd… v0.31.0 base)

== backup before the change ==
~/odm/backups/lobes-20261014-180402-pre-clef-interim.tgz
~/.lobes/docker-compose.override.yml.bak-20261014-180402-pre-clef-interim

== memory ==
freed beforehand: host 'available' 9,046 -> 100,108 MiB, by stopping a
  non-fleet container with the operator's go (not named, by rule)
before the sidecar started (15:05 UTC): used 8,792 MiB, available 115,818 MiB
sidecar healthy 15:10:50 UTC (load ~5 min, ~52 GB of bf16 weights)
after load: used 62,443 MiB (+~53.7 GB), available 62,166 MiB;
  after the probe run: used 72,110 MiB
GPU memory held by the sidecar process: 60,445 MiB

== probe (culture-dev/clef-probe-set @ cf9b911a, 320 records / 418 questions),
   direct to the sidecar, one request at a time ==
top-1 accuracy 0.9354, failed records 0
per-question probabilities: max |sum(p) - 1| = 0.0002 (criterion ±0.01)
latency by input length (median / p90):
  <= 1024 tokens  (n=279)   441 / 545 ms
  1025-4096       (n=13)   2974 / 3035 ms
  > 4096 (to 10,606, n=28) 15.4 s median, 16.7 s max

These bf16 results are the reference the quantized builds were scored against
(docs/evidence/2026-10-14-verify-clef-quants.txt).
