Record hardware and model inventory evidence

2026-08-22 11:50:00 +03:00
parent aec63de0c7
commit 93bc4a0fdf
3 changed files with 68 additions and 2 deletions
@@ -0,0 +1,43 @@
+# WP-200 Hardware and model inventory evidence
+
+Task: DEV-000111  
+Evidence date: 2026-08-22  
+Target: `docker-prod` / `docker.cin.su`
+
+## Host evidence
+
+Read-only `ssh docker` inspection returned:
+
+- host: `docker`
+- CPU: `128` logical processors
+- memory: `251 GiB` total, `203 GiB` available at inspection time
+- swap: `8 GiB`, `256 KiB` used
+- root storage: `1.6T` total, `359G` used, `1.2T` available, `24%`
+- `nvidia-smi`: unavailable
+- Docker server: `29.4.0`, `128` CPUs, approximately `252 GiB` Docker memory
+- Docker runtimes: `runc`/`containerd`; no NVIDIA runtime reported
+
+## Runtime evidence
+
+Read-only `docker --context docker-prod` inspection found:
+
+- `assistant-cognitive-worker-1`: production revision `b439a813...`, `/healthz` returned `status=ok`
+- `assistant-cognitive-lifeboat-1`: `ghcr.io/ggml-org/llama.cpp:server`, Docker health `healthy`
+- `llm-local-api`: `ghcr.io/ggml-org/llama.cpp:server`, Docker health `healthy`
+- no model activation, routing change or deployment was performed by this inventory task
+
+## Qualification conclusion
+
+This host has no verified GPU/VRAM capability in the inspected runtime. CPU
+and lifeboat health are evidence of availability, not model qualification.
+WP-201 still requires immutable model digests, holdout quality tests,
+latency/concurrency measurements, license/data-policy evidence and restart
+behavior. Heavy-class activation remains a qualification decision, not an
+inventory inference.
+
+## Evidence limitations
+
+This record does not establish the topology of any separate GPU host, does not
+read secrets, and does not claim that a model is qualified from its image name
+or health endpoint alone.
+