My local daily brief used 0 model calls

I almost put a 0-second llama3.1:8b row on a model-speed chart. On September 4, 2026 the daily brief stamped itself ready. The same receipt recorded 0 model calls, 0 successful model calls, and 0 seconds of model time. The GPU did no generate work.

Short answer: preview_only ranked source files and stopped. A ready flag with 0 generate calls is a retrieval pass, not a model run. Do not mix that row with a later llama3.1:8b generate. Canonical URL: https://bmdpat.com/blog/local-llm-preview-only-zero-model-calls-2026

What would I have charted by mistake?

If I only read daily-brief-ready, the chart would show a successful llama3.1:8b brief. If I also read model_call_count, the same day is a ranked file list with no inference.

I already refuse to promote a local model just because it exists. I also refuse to treat a ready stamp as proof of inference. Those are different events.

A later generate with the same model name would look slower and messier next to this 0-second row. That comparison is false. The 0-call row had no citation errors because it wrote no citations.

What did the September 4 receipt record?

Run 20260904-000702 used llama3.1:8b. The flag preview_only was true. The receipt recorded four sections. model_call_count was 0. model_ok_count was 0. total_latency_s was 0. citation_error_count was 0. redaction_issue_count was 0.

The last two zeros are not a quality win. The model wrote nothing, so it could not cite a wrong file or leak text.

The same pattern showed up on September 3, 2026 in run 20260903-083854. Same model. Same preview_only flag. Same 0 model calls and 0 seconds.

Did llama3.1:8b generate any tokens?

No. A local Ollama run is a generate request. The Ollama generate API is that call. This receipt never made it. Latency stayed at 0 seconds because there was no generate interval to measure.

The model name in the header is the model that would have run if preview_only were false. It is not proof that llama3.1:8b produced tokens.

When I size a local agent I care whether the GPU did work. 0 seconds of model time means this brief did not use the 5090. A later llama3.1:8b generate is a different event.

What did the file ranker actually score?

It ranked files. The Decisions section scored a Request file at 3.0 on keyword overlap. The Coding section scored Reports/5090/vault-model-lab/2026-09-04.md at 41.0 and the September 3 report at 43.0. The Learning and Growth sections scored signal and reflection pages.

Each section listed two sources and a score. No section includes model output. The review gate on the receipt says answers stay unreviewed until a human approves them. There were no answers to review.

If I treated that 3.0 as llama3.1:8b judgment, I would be scoring keyword overlap as model quality. It is not.

I already preview retrieval before a local model run. That post is about inspecting chunks before generate. This receipt skipped generate after the preview. The preview became the whole job.

Why can I not quote this brief as a model result?

daily-brief-ready was true. A dashboard that only reads that flag reports a successful brief. A dashboard that also reads model_call_count reports a retrieval pass with no model.

The runtime preflight I run before a benchmark already separates "the runtime loaded" from "the model scored." This is the same split for a daily brief.

The rule is short. If generate did not run, do not quote the brief as a model result. Do not drop the 0-call row onto a tok/s chart. Do not use 0 citation errors as a quality score.

What does the dashboard need next to ready?

I keep the model name. I also keep model_call_count, total_latency_s, and preview_only on the same line. A brief with preview_only true and 0 calls is a ranked file list. It is not a local model result.

Mixing them in one daily-brief chart would hide both. The retrieval pass would look cheap. The real generate would look expensive. Neither number would describe the other.

Accompanying prompt

What the prompt does: It reads a local daily-brief receipt and says whether the named model actually ran, or only ranked files.

Copy/paste this prompt:

Copy-ready prompt

Paste the exact block into your coding agent.

No article chrome, no footnotes, no formatting drift.

Copy promptSelect text

Role: You are auditing a local LLM daily-brief receipt. Context: Date: [YYYY-MM-DD] Run ID: [id] Named model: [model] preview_only: [true, false, or unknown] model_call_count: [number or unknown] model_ok_count: [number or unknown] total_latency_s: [number or unknown] daily-brief-ready: [true, false, or unknown] Sections ranked: [count or unknown] Model output present: [yes, no, or unknown] Task: 1. Say whether generate ran. 2. Name the first field that shows the answer. 3. Say whether the ready flag can be quoted as a model result. Output: - GENERATE RAN or GENERATE DID NOT RUN - Evidence field and value - READY MEANS MODEL RESULT or READY MEANS FILE RANKING - One repair: which field a dashboard must show next to ready Constraints: - Do not treat a model name as proof of a model run. - Do not treat 0 citation errors as quality if model_call_count is 0. - Do not invent missing fields. - Keep unknown fields unknown.

31 lines/932 chars

Ready

This prompt and every other one we publish live in the free prompt library.

Copy the block above.

Get the next measured local-AI field note by email: The 5090 Reports.