cesnimda
2.5 KiB
Bounded real-adapter evaluation authority and limits
AUTOMATION-HANDOFF-01 REQUEST section7 authorises targeted model evaluations when approved immutable artifacts and safe isolated development conditions exist. The earlier owner request approved creation of the local Ollama instance matching the recorded production model name. LOCAL_RUNTIME_ACCEPTANCE.md records the approved official Qwen GGUF/manifest and dedicated local runtime. This permits a bounded synthetic compatibility probe; it does not approve model activation, production access or factual-accuracy acceptance.
Preconditions, checked before inference: Desktop Linux context; exact approved local Ollama image; manifest/blob verification with existing verify_ollama.py; cloud disabled and CPU/memory/concurrency bounds; actual development sidecar keeps CV_IMPORT_AI_ENABLED=false. Inspection found that the existing local service has a read-only root but its dedicated model volume is writable. Do not infer read-only cache protection from that root setting. Instead start a disposable Ollama container using the same immutable image and existing model volume mounted read-only, on a generated internal-only network with no published ports. A separate bounded Python client container on that network exercises the adapter. Neither container receives a daemon socket. Remove only these exact owned containers/network afterward. The existing local Ollama service is not changed or restarted.
Exercise only cv_import_ai.analyse using the Python standard-library transport to the isolated ollama service alias and explicit model qwen2.5:7b. Three independently authored synthetic EN/Bokmål/adversarial block sets have explicit fact whitelists. Each request retains the existing schema, grounding checks and120-second bound; the entire run is finite. No download, model replacement, cache write, external provider or real personal document is allowed. Record actual native exit, timeout/rejection, structured suggestions, exact expected matches and unsupported claims. Quotes alone do not establish that instruction text is a career fact.
These cases are development compatibility probes, separate from the frozen benchmark and active deterministic Docling pipeline. A small success cannot establish general extraction precision/recall, multilingual quality, safe activation or end-to-end PDF/DOCX model accuracy. Any timeout, malformed output, missed field or unsupported claim stays visible. Phase10 does not gain authority from this test. No service configuration is changed.