LOCAL AI WORKSHOP / ISSUE TWO / NOVEMBER 14, 2026

Start with benchmark.py. It sends the two printed briefs to a loopback Ollama
service, saves each exact response and tests its build_report function against
four invented-note fixtures. A single outer Markdown fence is stripped for
extraction; the raw reply remains in results.json. Fixture counts measure
function behavior, not compliance with the requested reply formatting. Read
generated code before extending it. The AST
checks and restricted builtins narrow this teaching exercise; they are not a
security sandbox for arbitrary untrusted programs.

Example (from the extracted program pack):
python3 lab/local-workshop-2026-11-14/benchmark.py \
  --endpoint http://127.0.0.1:11434 \
  --models qwen3.5:4b llama3.2:latest --out my-local-results

The primary-results and primary-llama-results folders contain actual model
outputs from the 128GB M5 Max. Calls used Ollama 0.32.0 with cloud disabled and
one cold generation per brief. The first validator disallowed sorting lambdas;
we corrected the validator and rechecked the exact saved Qwen responses. We did
not regenerate or repair model code. Notebook retains all notes; triage creates
ready and needs_review lists of source_id/quote records. Duplicate IDs, exact
text, sorting, empty input and input preservation remain requirements.

Model tags can move. Compare the recorded digest, options, runner and hardware,
and inspect errors as well as fixture counts. Generation timing and correctness
answer different questions. This small experiment does not rank models in
general or reproduce a vendor's complete benchmark suite.

The editorial installation and owned-server helpers stay in the magazine's
production project under tools/local_ai, outside this reader pack. Installation
records in edge0-installation retain source/model revisions, commands and file
hashes. The pinned public guides are in the magazine's sourcebook. Downloaded
weights are not included. Keep source and weights licenses with your own copy.

An extra clarification experiment used the original M5/Llama setup with one
sentence added to the notebook purpose. It kept all notes and introduced
brackets into unconfirmed quote text: 1/4 fixtures passed. Exact prompt, reply
and cases are in primary-clarification-results. To reproduce that brief:
python3 lab/local-workshop-2026-11-14/benchmark.py \
  --endpoint http://127.0.0.1:11434 --models llama3.2:latest \
  --briefs notebook-clarified --out my-clarification-results

Edge0 on the M1: the original Metal template returned HTTP503/engine exit5
before generating code for both briefs. A reversible local change reduced
STRATA_METAL_PROMPT_CHUNK from8192 to64 and retained a native log. Both fresh
follow-up functions passed4/4 cases. These are different configurations; the
failed defaults remain recorded separately. The two decode rates were26.9 and
34.4 tokens/s; wall times31.98s and6.39s include prompt processing on the loaded
server. Startup is separate. The top-level success field records procedure
completion; inspect each request's error/cases fields for its actual outcome.
