Continue after the first build: AI and neural features

No training is required for this continuation. The exported package already has a reviewed, local pure-Perl assistant. First inspect and replay that known baseline. Train a replacement only after you understand its data and quality gates.

1. Keep the four layers separate

Layer	What it does	What it cannot do by itself
Federation Image v2	Stores and resolves byte-exact VFS content and embedded children.	It does not grant an AI permission to alter the workspace.
Packaged Perl NN assistant	Classifies a question into an intent and returns a reviewed knowledge answer.	It does not execute generated code, build a package, or approve itself.
Monk Loom attachment	Runs bounded deterministic neural inference from explicit VFS or Monk signals.	Its output is advisory evidence, not commit authority.
Specialist branch agent	Can return a bounded, proposal-only workspace recipe after separate training and signing.	It cannot trust its signer, accept its plan, apply changes, or start a build.

2. Ask the bundled assistant

Open PowerShell in the generated Perl distribution from the first walkthrough:

Set-Location "$Demo\dist\Example-First-1.0.0"
perl .\bin\vfs-perl-assistant --ask "How do I verify a generated package?" --explain
perl .\bin\vfs-perl-assistant --ask "What is a Monk scope?" --json

The explanation identifies the selected intent and confidence. The answer text comes from reviewed knowledge shipped with the package; the neural network selects among those bounded intents. Low-confidence input stays uncertain rather than acquiring extra authority.

3. Read one neural prediction from input to answer

The text vectorizer turns words and short word sequences into a fixed-size numeric feature vector.
The dense network passes those numbers through its configured hidden layers and activation function.
The final layer produces one score per intent label. A softmax converts the scores into comparable probabilities.
The highest bounded intent selects a reviewed knowledge article. Confidence and the chosen label remain visible as evidence.

Dimensions control the feature-vector width. Hidden layers control model capacity. Epochs are repeated passes over training rows. More capacity or more epochs can memorize data, so Studio evaluates a separate benchmark and does not treat training accuracy as release evidence.

4. Replay the independent benchmark

perl .\bin\plnn-text-evaluate `
  --model .\share\example-first\assistant\model.json `
  --data .\share\example-first\assistant\benchmark.csv `
  --json `
  --output .\assistant-replay-metrics.json `
  --predictions .\assistant-replay-predictions.csv

Compare the reported row count and accuracy with share\example-first\assistant\README.md. Accuracy is the overall correct fraction. Macro-F1 gives each intent equal weight, and minimum per-intent recall exposes a weak label that a strong overall average could hide. A mismatch is a stop condition.

5. Inspect Studio's training controls without starting training

Return to Studio and choose AI Assistant → AI Configuration.
Read the Datasets, Network, Training, Evaluation, Security, Packaging, and Branches tabs. Tooltips explain each field.
Notice the separate workspace-root boundary, worker ceiling, dataset size limits, accuracy floor, macro-F1 floor, and per-intent recall floor.
Close the dialog without saving if you are only learning. Opening configuration never trains or changes the VFS.

6. Understand the six Workbench gates

Choose AI Workbench. The visible stages are intentionally separate:

Configuration saves or loads typed JSON settings.
Preflight validates paths, UTF-8 formats, label coverage, benchmark separation, secrets, links, and limits.
Train and evaluate runs pure-Perl backpropagation outside the UI process and produces measured artifacts.
Review candidate recomputes digests and displays architecture, labels, and quality.
Human approval binds a named decision to the exact candidate fingerprint.
Select for packaging changes configuration only; a later explicit build must still include the candidate.

Do not press Train with improvised data. Continue with Prepare assistant CSV and JSON data, then follow Use the AI Workbench when you want a real replacement candidate.

7. Locate Monk Loom in the architecture

The conversational assistant and Monk Loom are different neural systems. The assistant maps text to reviewed intent answers. Monk Loom is a small deterministic integer network bound into a VFS image. Monk Spindle recipes extract named, ordered features from VFS and policy observations; Loom evaluates those features under explicit budgets. Verified adaptation can produce a proposed attachment, but only a separate review and commit path can replace the bound attachment.

Read these in order: Monk Loom overview, feature recipes, deterministic inference, and direct VFS training.

8. Decide whether you need a main assistant or an agent

Your goal	Use	Next walkthrough
Improve answers for ordinary operator questions	A replacement main assistant candidate	Datasets then AI Workbench
Score bounded VFS or Monk observations deterministically	Monk Spindle plus Monk Loom	Direct VFS training
Suggest project, source, mount, or profile changes	A signed specialist branch agent	Agent reconfiguration

9. Completion check

You have completed the AI continuation when you can explain why the exported baseline works without Python, distinguish training rows from an independent benchmark, identify the assistant and Monk Loom as separate systems, and state which human actions remain required before neural output can affect a Chronicle workspace or a release build.

Permanent boundary: a model score, candidate digest, or trusted signature is evidence. None is an authorization token. Workspace review, application, build, and release remain explicit and separate.