Model ML Pipeline

This meta skill owns selection and ordering, not child methodology. Do not load smoke-test-ml-pipeline as a sibling of evaluate.

Human-facing prose

Details: setup-workspace references/human_facing_prose.md. Design notes, JOURNAL text, and # comments describe this experiment — not the skills framework, the CLI, or the command that produced an output. Questions, replies, and the close narrative use the same data-science language — not skill ids, G-* names, or the wrapper CLI. Trailing locator tokens stay index strings. <!-- results-embed: … --> is a site marker. Authoring hints stay in this skill. style is ruff only.

Design note shell

python -m skore_skills scaffold --journal --stem <NN_short> writes journal/<stem>.md. Do not recreate it from memory. Fill Question, Motivation, Method, and Risks as science (what to learn, why now, what changes, what could invalidate the result). Status lifecycle: planned → approved → running → done | abandoned. Those four content sections freeze after approval; only Status, generated notebook viewers, and the Method <!-- results-embed: pipeline --> embed change afterwards. Abandoned: one-line reason on State; Headline result n/a — abandoned: <reason>. There is no "Success criteria" section. Keep ## Notebooks with Evaluation then Audit.

Entry routing — deterministic

1. Run python -m skore_skills status. If status.setup.pending is non-empty and status.skills.setup-ml-project is true, load setup-ml-project and stop. Do not start this skill. When it returns, continue. Do not load it again on this turn. If that skill is not installed, name the pending pieces in one line and stop. Do not invent git init, scaffold, or env init. If status.setup.env or status.setup.workspace is declined, stop in one line. A declined git or editable is not asked again and does not block modeling.
2. Framing is mandatory. If status.modeling_decisions is not locked, load frame-ml-problem only if status.skills.frame-ml-problem is true and stop. Do not invent the table, offer model choices, write a design note, or declare a pipeline. Missing skill → stop in one line. This includes an approved-stem resume: do not build until the table is locked. When the table is locked, run python -m skore_skills frame show. Anything other than proceed loads frame-ml-problem and stops. A proceed whose translation is null has no splitter translation: say so and stop. Do not write model code. Do not present a choice list: no dummy predictor, standard baseline, EDA-driven proposal, Backlog row, or discussion.
3. Resume beats menu. If the user names an experiment stem, or status.policy.loop.stem / last_history_stem identifies a current design, run python -m skore_skills design consent --stem <stem>. Treat JSON action as authoritative. proceed → resume that stem directly; do not ask how to start again. ask → Design approval below (render JSON context inline per § Gate context); do not write code. stop → missing note: name scaffold --journal --stem (or abandoned: explain and do not implement). Do not infer approval from "build it". Missing shell: no site build, no fill from memory.

4. Otherwise run python -m skore_skills model choices. Treat its JSON as authoritative.

   • action stop / modeling_decisions_unlocked — load frame-ml-problem and stop. Do not offer a model menu.
   • If choices includes baseline and the user did not ask to discuss, propose a pipeline from the EDA, or pick a backlog item, do not AskUserQuestion. Take the locked-baseline contract below. On this turn, run that contract in order — python -m skore_skills status, python -m skore_skills model choices, python -m skore_skills scaffold --journal --stem <NN_short>, then fill the note — before any design-note prose or question. The only gate after the note is the three lines in the locked-baseline contract, not Yes / No.
   • An explicit discuss, EDA, or backlog request takes that branch even when baseline is present. Do not show the menu first.
   • Otherwise present exactly choices[], in returned order, in one single-choice AskUserQuestion. Drop baseline from that question: it is not a menu item.
     • eda_proposal → Propose a pipeline from the EDA
     • backlog → Pick from the Backlog
     • discuss → Discuss the next step

   Carry each choice's JSON reason as its description so the user reads why it is on offer. Choosing backlog asks which row, not whether to build it.

Do not add a disabled choice, infer availability yourself, or reorder the list. In particular: no locked-baseline choice when model_stems is non-empty; no EDA proposal unless data_analysis is present; no Backlog option when backlog is empty. Discussion is present only after the lock.

Gate context

Every approval question carries its own context. Before asking, state in 2–4 lines what the answer authorizes, the facts it rests on — echoed inline — and what each option does, in data-science terms (Approve / Modify / Stop; dummy vs baseline). On the locked-baseline branch, those lines name the baseline and its note only, and the options are the three lines in that contract. Do not name skill ids or G-* tokens in the question. A file link is an addition, never the context: "read journal/<stem>.md and approve" is not an approval request.

On design consent ask, the JSON context holds those facts: question, source, files_touched, change, and up to two risks. Quote them in the message (question and planned change first, then the risks the user may push back on). After the note is populated, and before Approve / Modify / Stop, run python -m skore_skills site build --if-stale when site policy is true. In that same question, say the website has been updated and paste the file:// line from that output verbatim. Do not name the design-note markdown or html/<stem>.html. An empty field means the note does not state that fact: say the note is not ready for approval and offer to populate it; never fill it from memory. Do not site build an empty shell.

Choice contracts

An EDA proposal or a discussion first produces a user-confirmed proposal and an approved design note. The proposal yes agrees that idea. A locked baseline and a backlog row do not: the frame table or the picked row is the decision. The note is approved only by Design approval below. No branch writes model code before that gate is proceed. Use the next available numeric stem; never overwrite an existing note.

• Locked baseline (baseline). The note names the one comparison model the journal already locked. A dummy token is a DummyClassifier or DummyRegressor inside the normal skrub DataOps declaration: it proves loading, fit/predict, and pytest smoke, and it is not expected to add predictive value. Any other token (logistic, seasonal_naive, group_mean, production) is that one comparison model. Do not upgrade it to another estimator. Do not AskUserQuestion before the note. The message lists these three commands, in order, before any design-note text. A workspace summary is not a substitute:

  python -m skore_skills status
  python -m skore_skills model choices
  python -m skore_skills scaffold --journal --stem <NN_short>

  Then fill Question, Motivation, Method, and Risks from the locked baseline, its note, and the comparison metric. Before the question, two to four lines name that baseline and its note. They do not say what Approve allows. The question is only these three lines:

  • Approve — accept this note
  • Modify — change this note
  • Stop — leave this note unchanged

  Nothing follows those lines. Do not mention discuss, an EDA proposal, or the backlog unless the user already asked for one of those. Keep the normal post-smoke Evaluate (Recommended) / Modify / Stop gate for a later turn.

• EDA proposal (eda_proposal). Read data_analysis/data_analysis.md and the project goal. Cite the EDA findings that motivate one pipeline proposal. Do not invent findings or present multiple silent alternatives. Restate the proposal, then one single-choice AskUserQuestion: Yes / No. Do not also ask for a typed yes. "Maybe" is not Yes. No or Stop writes nothing. A later Yes (the tool answer, or an explicit yes on a later turn) is what authorizes the write. On Yes, write the note, then Design approval, before build.
• Backlog (backlog). Load manage-ml-backlog only if status.skills.manage-ml-backlog is true. Else one-line skip; do not invent a Backlog. Pass the backlog rows returned by the CLI. That skill asks for one B<N> when the user has not named one, and asks only for shaping facts the row does not state. It does not ask Yes / No. When it returns the proposal, run scaffold --journal --stem <NN_short> and fill the note from that Item, Source, and the facts just given, then Design approval, before build. Do not add an estimator or a procedure neither the row nor those facts stated. Do not add a new Backlog idea in this branch.
• Discussion (discuss). Have an open conversation about what to learn, why now, and what changes. Restate the agreed idea, then one single-choice AskUserQuestion: Yes / No. Do not also ask for a typed yes. "Maybe" is not Yes. No or Stop writes nothing. A later Yes (the tool answer, or an explicit yes on a later turn) is what authorizes the write. Only after that confirmation create/populate the design note, then Design approval. If no idea is agreed, return to the entry choices.

Before execution

This dispatcher labels the next kind of work but does not duplicate a child's detailed preview.

• For discuss, proposal shaping, or literature-backed design work, emit 1–3 natural sentences saying this is LLM discussion/research over recorded project facts, with no model fit, smoke test, or CV. Name any scratch research note or design note that may be written and stop at confirmation.
• For an approved implementation, say which child comes next and the broad sequence: local pipeline preparation → small real-data smoke fit/predict → optional full-dataset evaluation → review. Then let build-ml-pipeline, smoke-test-ml-pipeline, evaluate-ml-pipeline, and review-ml-experiment each own the single detailed Before execution preview at its actual compute boundary. The dispatcher preview is those four phase names only, one short line each. Do not mention the Method viewer, a site rebuild, a row-count assertion, or skore.evaluate. Do not list their commands (status, frame show, design consent, smoke run) or a DataOps declaration in this preview. Do not name loop notebooks or notebook convert. Notebooks stay off until a report exists. With no shell, that four-phase paragraph is the whole answer: do not name the post-smoke Evaluate / Modify / Stop question, an unfitted site build, or pipeline.svg.

Do not invent minute estimates at dispatcher level. Name a known duration only when explicit measured evidence is available; otherwise leave cost details to the child that knows data scope, folds, and selected views. Pending proposal/design approval may preview the sequence but never starts local work.

If the design-note shell is missing, this turn only names python -m skore_skills scaffold --journal --stem <NN_short> and stops, except on the locked-baseline branch and on a backlog proposal that skill already returned. Those two run the command and populate Question / Motivation / Method / Risks in this same turn, then Design approval. Do not fill those sections from memory when the command did not run this turn.

Design approval

One gate approves a populated note. Run python -m skore_skills design consent --stem <stem>. proceed → already approved; do not ask again. ask → render the JSON context inline (§ Gate context), including the site preview when policy.site, then AskUserQuestion (single choice), in order: Approve / Modify / Stop. Do not also ask in chat whether the note looks right.

• Approve → run python -m skore_skills design approve --stem <stem>. Do not type the date. Re-run design consent; code starts only on proceed.
• Modify → leave State planned, edit the note, and ask this gate again.
• Stop → do not implement.

Approved-design implement loop

1. Load build-ml-pipeline only if status.skills.build-ml-pipeline is true; else one-line skip and stop — do not declare the pipeline from this meta. That skill loads smoke-test-ml-pipeline after the experiment file exists and runs python -m skore_skills smoke run --stem <stem>. JSON stop stays in build (modify the pipeline, re-run smoke run). Do not tell the user to edit the smoke test's expected row count so it matches the sample. proceed: build reports the design, then python -m skore_skills evaluate consent --stem <stem> (Evaluate / Modify / Stop on ask, with that JSON context rendered inline per § Gate context). Build also writes the unfitted Method viewer (scratch/results/<stem>/pipeline.svg and the DataOp report scratch/results/<stem>/pipeline/) and, when policy.site is true and export-ml-site is installed, runs site build --if-stale after that snapshot and before Evaluate so Method shows that report. Missing or skipped EDA does not defer it. The post-loop rebuild refreshes the same unevaluated report; it does not replace this one. Do not fill experiments/<stem>.py at this unfitted snapshot if it already contains skore.evaluate. That ban ends when loop notebooks returns convert.
2. Only if the user chose Evaluate and smoke run is proceed: load evaluate-ml-pipeline only if status.skills.evaluate-ml-pipeline is true. It reuses the DataOp cv and writes skore.evaluate in the experiment script. Re-run status first, then python -m skore_skills evaluate consent --stem <stem>. Consent JSON is authoritative, not the user's wording alone. Missing evaluate-ml-pipeline → one-line skip. Do not invent that skill's steps.
3. After a successful dispatched evaluate (locator returned): run python -m skore_skills review consent --stem <stem>. Treat JSON action as authoritative.
   • stop — no scratch/results/<stem>/report.html. Do not review. Name that file. This is the evaluation snapshot, not the site launcher report.html. Do not record-outcome.
   • audit or proceed — load review-ml-experiment only if status.skills.review-ml-experiment is true. It returns the digest, G-AUDIT-FINDING, locator, and idea paths. Do not load audit-ml-pipeline from this dispatcher. Then § Close, in that order.
   • Missing skill → one-line skip and run § Close with n/a — audit not run.

Do not duplicate child-skill methodology. Before new library symbols are written, children use python -m skore_skills api get <dotted>.... Several symbols go in one call.

Stop conditions

• Pending setup (status.setup.pending non-empty): load setup-ml-project and do not start build. A declined git is not asked again. A declined env or workspace stops.
• Generic model landing: do not hand-author the menu; run python -m skore_skills model choices.
• If journal/NN_<short>.md is missing, this turn only names python -m skore_skills scaffold --journal --stem <NN_short> and stops, except the locked baseline and a returned backlog proposal. Those run that command and populate Question / Motivation / Method / Risks in the same turn, then Design approval. Do not fill the template from memory when the command did not run. Do not site build before the shell exists.
• Require design consent proceed before code. Do not treat "the user asked to build", or a chat yes on the drafted note, as approval.
• Do not open an approval gate whose only context is a file path; render the context facts inline (§ Gate context).
• Preserve identical stems across design, experiment, smoke, audit.
• Do not replace skrub DataOps with bare sklearn Pipeline.
• Do not persist a result as done while smoke run is stop.
• Do not load evaluate (or write skore.evaluate) before the post-smoke HITL answer is Evaluate (evaluate consent ask), except proceed for a stem that already has a persisted report, or while smoke run is stop.
• Do not load smoke-test-ml-pipeline from this dispatcher.
• Do not duplicate child-skill methodology in this dispatcher.
• If status.data_analysis is missing, continue with facts the user stated; do not invent an exploratory data analysis report.
• Literature-backed feature-engineering or learner-family questions: load build-ml-pipeline only if status.skills.build-ml-pipeline is true (it loads research-ml-practice only if status.skills.research-ml-practice is true). Else one-line skip. Do not distill research here; do not invent papers from memory.

Close

Reached only after review returns, a missing review skill, or red smoke. The four-phase preview does not enter here, and neither does review consent stop. Run the steps in this order. Do not reorder them. loop notebooks continues only on skip. record is the loop artifacts action, not this gate. Red smoke skips steps 3 and 4: do not mark the experiment done. This dispatcher owns the user-facing close. Children return locator / digest / finding and do not preview it.

1. python -m skore_skills loop notebooks --stem <stem>. Treat JSON action as authoritative. Do not record-outcome, site build, or git end-turn while action is convert.
2. On convert, run python -m skore_skills notebook fill <source> for every experiments/ or audit/ entry, with --html when html is true. Do not notebook convert those files. Fill does not put. Re-run the matching materialize.py first only when that human call changed since the run that wrote the current artifacts. Filling only audit/<stem>.py does not finish the close. Re-run step 1 until action is skip. Missing jupytext / nbformat / nbconvert → one-line skip naming add-python-package; do not fail the turn. skip / not_evaluated means the report is missing: do not fill an unfitted script. When action is convert, the unfitted-snapshot ban does not apply, including when experiments/<stem>.py already contains skore.evaluate.
3. record-outcome, only once step 1 is skip. Load manage-ml-backlog only if status.skills.manage-ml-backlog is true, in record-outcome mode, handing it the locator evaluate already returned, plus the optional headline and G-AUDIT-FINDING (n/a — audit not run when skipped). Else one-line skip; do not write History from this meta. It writes the journal/JOURNAL.md History row and design-note Status block plus ## Results, then returns. It does not triage idea files. Never mark done while smoke run is stop. Missing headline becomes n/a, never an invented metric.
4. site build only when policy.site is true and export-ml-site is installed: python -m skore_skills site build --if-stale, so the Method DataOp report and Results replace the construct-time snapshot. Skip in one line otherwise. Site build embeds audit/<stem>.nb.html under ## Notebooks; do not add <!-- results-embed: audit -->. If site build errors with mkdocs-material is required, load add-python-package for mkdocs-material (agent) and build once more. Do not pixi add / uv add. Name a build error; do not fail the turn.
5. User-facing close. 2–6 sentences of the result, grounded in the audit digest when present (Checks + Metrics), else the user's headline / report.txt. Do not invent a metric. When policy.site is true and the build ran, say the website has been updated and paste the file:// line from that output verbatim. Do not name html/<stem>.html or journal/<stem>.md. When the site was not built, do not claim it was updated and do not name the markdown. Then G-REPORT-LOCATOR (the locator evaluate passed up, or n/a — backend did not expose a locator) first among tokens, then G-AUDIT-FINDING. Index strings, not the narrative.
6. python -m skore_skills git end-turn --stage implement. If JSON action is invoke, load persist-ml-git only if status.skills.persist-ml-git is true and stop; that skill returns to triage. If persist is missing, name the pending staged paths and stop. Otherwise load triage-ml-task only if status.skills.triage-ml-task is true; else stop. Do not run git commit in this skill.