ollama served the wrong weights under the right tag

this is why i screenshot model settings before every agent session now

had a similar thing where cursor kept saying it was on sonnet and the output felt like a tiny local model. turned out my proxy had silently fallen back after a rate limit. ui still green.

i refuse latest on anything that touches prod code. tags only, and i write the digest in the PR description like a weirdo. has saved two bad merges.

digest or nothing. tags are social fiction.

i keep a one-line check in my agent harness: before first completion, ask the runtime for model_id + parameter count and abort if it doesn't match a allowlist. logged it after an eval suite "improved" overnight because ollama had swapped in a chat-tuned 7b under the same name.

no eval on the harness itself = you are measuring vibes. the rename PR inventing helpers is the classic symptom.

same class of bug on my 4090 box last week.

ollama show qwen2.5-coder:14b --modelfile still printed the right FROM line. sha256 of the blob did not match what I pulled on monday. the tag had been moved by a pull --quiet in a cron that "just updates models".

i now pin digests in a models.lock and the agent refuses to start if ollama list digest != lock. tags alone are marketing.

sorry english not perfect — but yes, refuse latest forever.

i hit this with num_ctx too. agent said it loaded 32k, ollama ps showed 8k after a silent model swap. eval scores jumped overnight for wrong reason.

now first tool call in my harness is: print digest, param count, and context size. if any mismatch → abort before first edit. ugly log. better than merging invented helpers.