Active voice, present tense, plain English, and consistent brand
capitalization across the README and the notebook's prose cells.
Code identifiers wrapped in backticks where missing; a sentence
fragment in the verify step rewritten as a complete clause.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Adds a masked mo.ui.text(kind="password") field so users can authenticate
by pasting a key instead of running `wandb login` in a shell. The resolved
value is kept out of the run config (never logged) and feeds
wandb.login(key=...) in the gated training cell; blank falls back to ambient
auth (shell login, WANDB_API_KEY, or netrc). Downstream cells read the key
via the returned wandb_api_key.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

GitHub resolves multi-segment (slashed) branch names in blob/<ref>/<path>
and molab fetches from GitHub, so a branch-based link points at the latest
revision and needs no per-commit comment maintenance. Content detection
still pins to the head SHA. Skip the comment write when the body is
unchanged so no-op pushes don't churn it.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The view-only link failure now names the required role (Member) and the
three ways an admin can grant it: the Registry Members UI, the Python SDK
(wandb.Api().registry(...) then add_member()/update_member()), and SCIM
(PATCH /scim/Users/{id} with registryRoles), with a link to the configure
registry access docs. README prerequisite updated to match.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The bare molab GitHub URL renders a static preview. Appending /server opens
the notebook in a hosted runtime so reviewers can actually run it — needed
here since the notebook depends on torch, which the in-browser /wasm mode
cannot run.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Adds a push-triggered workflow that runs `marimo export md` on each notebook
under examples/marimo/, writing a peer <notebook>.md, and commits the result
back to the branch so the rendered Markdown always tracks the notebook. marimo
is pinned for byte-deterministic output (no version-only churn), and three
guards prevent commit loops (GITHUB_TOKEN pushes don't re-trigger, *.py-only
trigger vs *.md-only commits, and [skip ci]).

Includes the initial export of mnist_registry.md.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Collapses 19 cells into 5, breaking only where marimo requires it or where
the user actually provides input:

- intro (markdown)
- setup + form: imports, device detection, and all input widgets in one cell
- train button: its own cell (a widget must be defined separately from the
  cell that reads its value)
- run training: one gated cell that configs, authenticates, inits, builds and
  trains the model, logs, saves, and links the artifact — streaming each
  milestone with mo.output.append so it isn't a black box
- verify + next steps (markdown), rendered once a run exists

run_button.value resets to False after its cascade, so editing the form after
a run re-runs the training cell but mo.stop halts it immediately — no silent
retrain. Regenerates the Markdown peer.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Reorders the notebook to: intro -> setup+form -> training pipeline -> Train
button -> verify. The button still gates the pipeline (so it doesn't run on
open), but now reads as the explicit "run the code above" trigger. The
training cell's pre-run message explains what it does and points to the button
below.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The entity form field read just "W&B entity", which reads as jargon. Label it
"username or team" to match the prerequisites wording. Also tighten the
registry-link remediation: an entity is a team or username (not an org), so it
now reads "set W&B entity to a team in an org where you have Registry write
access."

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Personal-username entities were removed for W&B accounts created after
2024-05-21, so a run's entity must be a team. Relabel the entity field
accordingly, fix the prerequisite, and wrap wandb.init so an
"entity ... not found" failure renders actionable guidance (set the entity to
a team) instead of a raw CommError traceback.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Pulls the CNN definition into its own cell so it can be shared, then adds a
cell that downloads the model from W&B (preferring the registered version,
falling back to the run's artifact so it works even when the registry link is
blocked), loads the weights into a fresh network, and classifies 10 held-out
MNIST test digits — rendering each as an image with predicted vs. true label
and an N/10 correct tally. No new dependency (numpy ships with torch; mo.image
renders the arrays).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The consume cell reused `correct` and `pred`, which the training cell already
defines — marimo forbids defining a name in more than one cell, so the cell
refused to run. Rename them to `n_correct` / `prediction` (both cell-local to
the consume step). marimo export does not enforce the single-definition rule,
only the kernel does, so this slipped through earlier validation.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

)

* Refactor the mnist notebook for readability and usability

- Move imports and constants into a setup cell
- Wrap the UI elements in a submittable form, removing the
  ui.run_button()
- Refactor the monolithic training and logging cell into multiple helper
  functions, each declared in their own cell as a reusable function
- Separate logic cells from view cells when possible
- Clean up the evaluation presentation, using a table instead of
  markdown+emoji
- Add a mo.outline() so users can see an outline of the notebook at a
  glance

  Core principles:

  * Gate execution once on mo.stop(), then let the graph handle the
    rest. All other cells reference run/config, which don't exist
    until the form is submitted. This lets us remove lots of
    confusing conditions.
  * Separate logic from presentation
  * Push temporary variables into functions to reduce notebook globals.
    marimo notebooks work best when there are as few global variables
    as possible.

* docs: export marimo notebook(s) to Markdown [skip ci]

---------

Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>

Bring back example predictions the right way: a wandb.Table of un-normalized
digit images with true/predicted/confidence, logged once (no per-step slider).
Add a confusion matrix so "accuracy" is concrete — which digits the model
nails and which it confuses. Both log under the Training prefix after training.
Update the intro and verify narration to define accuracy and describe the new
panels.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Expand log_eval_report to also log a per-digit PR curve and penultimate-layer
embeddings (Net.features) for W&B's 2D projector — one cluster per digit,
colored by label. Set run.summary with final/best accuracy and parameter count
so they surface on the run overview. Document how to add the 2D Projection
panel in the verify steps.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>