EDGE0 BEGINNER WORKSHOP / LOCAL REVIEW / NOVEMBER 17, 2026

Goal: use a model on your own Mac to build Quote Desk, then change the
purpose of the prompt to build Publication Desk. The model returns text;
edge_workshop.py saves Python source. You inspect it and run it yourself.

PREPARE
This pinned recipe is for Apple Silicon. Our tested machine is a 64GB M1 Max.
It is not an Intel installation recipe; lower-memory Macs have not been tested.
Use macOS 15+ for the current Homebrew preparation below. Our Mac uses 26.5.2.
The model takes about 102GB of disk space. Keep another 25GiB available for
builds and first startup. The advertised 29GB concerns working RAM, not download
size. Close memory-heavy work first. Installation requires internet; generation
uses the local server on 127.0.0.1:19437. No account/API key is used by this kit.
Review the model license before downloading:
https://huggingface.co/Edge0/Edge0-145B-q3k/tree/2d8011637b46d532a65ed4f9d78eb5b466e486d5

Get Python 3.10+ from https://www.python.org/downloads/macos/ if needed.
Use the installer, then reopen Terminal. If HTTPS certificate checks fail,
follow that installer's Install Certificates.command guidance; keep certificate
verification enabled.
Install Apple's compiler tools with:
  xcode-select --install
Finish the native installer. Already installed is a successful preparation.
Install Homebrew using the official instructions at https://brew.sh/ and finish
its printed Next steps so brew is available in a fresh Terminal. Then:
  brew install cmake ninja uv
CMake creates build instructions, Ninja builds the engine, and uv prepares its
private Python environment. These tools download code and packages.

FIND THIS FOLDER
Extract the magazine's Programs + prompt workshop ZIP. Find
lab/edge0-beginner-2026-11-17 inside it. Open Terminal (Applications/Utilities).
Type cd followed by a space, drag that folder from Finder into Terminal, and
press Return. Run:
  pwd
  ls
The file list must contain edge_workshop.py, setup_edge0.py, the two prompt text
files, sample-notes.json and check_programs.py. All commands below start here.
Use the installed Apple Silicon tools in this Terminal:
  export PATH="/opt/homebrew/bin:$HOME/.local/bin:$PATH"
  python3 --version
The version must be 3.10 or newer. The default Apple Python in our SSH session
was older; selecting the installed Homebrew tools resolved the prerequisite check.

INSTALL ONCE
  python3 edge_workshop.py doctor
Resolve each Needs attention line first. Existing model/build lines may be
absent on a new installation; the next three commands create them.
  python3 setup_edge0.py prepare --out my-install
  python3 setup_edge0.py weights --out my-install
  python3 setup_edge0.py build --out my-install
Run each line separately and wait for the Terminal prompt. prepare obtains the
pinned engine source, its pinned math dependency, Apple's Metal C++ headers and
the model file manifest. weights downloads and fingerprints the selected model
files (it omits the unused alternative Q8 projection file). build compiles the
engine, prepares its Python environment, runs two numerical checks and builds
the front end. Each phase writes a JSON record in my-install; look for success:
true. Detailed original commands are readable in setup_edge0.py. The script's
fixed installation folder is ~/Models/Edge0-CAC-20261114. It never creates a Git
remote or uploads source. A resume reuses complete matching downloads; rerun the
same phase after a connection failure. Keep its .part and range records for
resume. A fingerprint mismatch calls for keeping the error and reviewing that
one file rather than bypassing the check. Large transfers can take hours.

The build disables optional ccache because it failed on our earlier Mac setup.
It limits compilation to two threads. If a build fails, read the first compiler
error in that Terminal, verify doctor, and keep my-install's records. This kit
is a pinned November 2026 route, not a promise that future tool versions match.

  python3 edge_workshop.py tune
  python3 edge_workshop.py inspect
tune uses our tested prompt chunk of 64 and saves the original template before
changing it, if a change is needed. inspect must exit 0 with an ok:true report.
An absent file calls for rerunning weights, then inspect. This is a completeness
check; the weights phase performs file fingerprint checks.

FIRST LOCAL REQUEST
In this Terminal:
  python3 edge_workshop.py serve
Wait for Listening:. It loads the model; leave this window open. A second
Terminal must start in the SAME workshop folder using the Finder-drag cd step.
  python3 edge_workshop.py ready
  python3 edge_workshop.py chat
You should see Ready and a short reply about Python and notes. The chat helper
sends one printed example question; it is not an interactive agent or shell.
The client uses the actual model identifier from /v1/models. The server speaks
an OpenAI-compatible HTTP format locally; this request does not use an OpenAI
account, endpoint or API key. The prompts, request JSON and helpers are readable.

BUILD AND RUN
  python3 edge_workshop.py ask notebook --out my-notebook
It saves the exact prompt, request, raw reply, timing and notebook.py. Open that
.py file in a text editor and inspect it before executing it. It should only
read the supplied notes and print a report. Then:
  python3 my-notebook/notebook.py sample-notes.json
A working build shows NOTEBOOK, three source IDs in order and all exact quotations.
Our first fresh build stopped because it added a --help option that argparse
already provides. We preserved that failure and used one focused repair:
  python3 edge_workshop.py repair --source my-notebook/notebook.py --out my-notebook-fixed
The repair command sends prompt-notebook-repair.txt plus your saved source.
Inspect the new file before running it:
  python3 my-notebook-fixed/notebook.py sample-notes.json
If your own first build already passes its checks, keep it instead; use its path
in the final checker. The supplied first-notebook folder contains our failing
original and tested-notebook contains the model's corrected program.

Read prompt-publication.txt. It preserves the common environment and input
requirements while changing the job from inspection to preparing quotations
for publication and showing follow-up work.
  python3 edge_workshop.py ask publication --out my-publication
Inspect the saved source, then:
  python3 my-publication/publication.py sample-notes.json
  echo $?
Success shows READY and NEEDS REVIEW; S2 remains in the latter. The number 1
means useful output with pending follow-up, as the revised prompt requested.
  python3 my-publication/publication.py sample-notes.json --ready-only > ready-draft.txt
Open ready-draft.txt in a text editor: only S1 and S3 belong there. This explicit
shell redirection creates or replaces ready-draft.txt. Pick another filename to
keep an earlier draft. A confirmed flag is the writer's decision, not AI fact
checking. Edit a COPY of sample-notes.json to explore your own invented notes.

CHECK WHAT YOU GOT
  python3 check_programs.py my-notebook-fixed/notebook.py my-publication/publication.py
The checker uses invented temporary inputs, then removes them. It checks sorting,
exact spacing/accents, input preservation, both output modes, pending-work
status, empty inputs, help, errors and duplicate IDs. A pass covers only those
cases. It is not a correctness or security guarantee for arbitrary source.
The supplied tested-* folders preserve the magazine's actual model outputs.
To inspect those instead of generating, use:
  python3 tested-notebook/notebook.py sample-notes.json
  python3 tested-publication/publication.py sample-notes.json
  python3 check_programs.py
Fresh generations can differ. Keep your original prompt/raw reply/failed case;
change one requirement in a copied prompt, predict the effect, generate to a
new folder and check it. This article's two prompts are full independent build
requests, not an in-place edit of the first source file.

RECOVER AND FINISH
- Cannot open edge_workshop.py: repeat the Finder folder-drag cd step and ls.
- Command not found: complete the Python/Homebrew Next steps; reopen Terminal.
- Cannot reach Edge0: start serve in its own Terminal; wait for Listening:.
- Port in use: reuse your own ready server or stop it in its original Terminal.
- HTTP 503 / engine exit: save the server error, check tune, stop the owned
  server with Control-C, then start it again. Do not start a second model copy.
- Too little memory / very slow startup: close memory-heavy work; try again
  when capacity is available. The smaller Ollama workshop is the lighter route.
- Existing output folder: choose a new --out name. The client preserves it.
- Truncated/invalid Python reply: inspect raw.json and reply.txt. Simplify the
  copied prompt, regenerate in a new folder and recheck before running it.
- Failed program check: retain the failure and use it in a focused repair
  prompt. A model can produce plausible code that fails a requirement.
Stop the server with Control-C in its original Terminal when you finish. The
downloaded source, model and your saved programs remain for the next session.

SCOPE
The PDF and session record report exactly which steps ran on our M1. Existing
source/model files may be reused; a replay is distinct from a clean-machine
installation, and first-time installer dialogs are documented preparation.
Private job logs stay outside the reader pack. All example notes are invented.
