Frame ML Problem

Write ## Modeling decisions in journal/JOURNAL.md. The table is the contract. This skill does not declare a learner and does not evaluate one.

Human-facing prose

Details: setup-workspace references/human_facing_prose.md. Journal rows describe this dataset. Do not name the skills framework, the CLI, or a splitter class in the table. Questions use data-science language — not skill ids, G-* names, or the wrapper CLI.

Procedure

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; continue. If data_analysis is missing and explore-ml-data is installed, AskUserQuestion: explore first (default) or continue from facts the user stated. Explore loads explore-ml-data and stops. Do not invent dataset facts. Do not ask this again once data_analysis is present or skipped.

2. Run python -m skore_skills frame show. When the user is changing a locked constraint and named one decision, do not add --revise and do not treat proceed as the end of the turn. Show this command once, in a fence. The only sentence for it is: That command drops that one recorded decision. Then run frame show with no --revise.

   python -m skore_skills frame clear --cell <key>

   Ask the replacement in this message, the way step 5 asks a missing decision. Do not say that frame show will ask it. When the named decision is the comparison metric and the message does not state the new value, say that metric role, prediction goal, and folds stay filled. Then name the experiment file from the workspace state and say it still uses the previous metric and is not run, and that the next build or evaluate rewrites it after the table is complete again. The last line is: The comparison metric is MAE. Which metric replaces it? Nothing follows that line. When the named decision is the prediction goal and the message already states the new value, write that value. Metric role and the comparison metric are unanswered. Status is draft. Ask those two in this message. Do not write Status locked. Otherwise, if the message already states the new value, write it and ask only the decisions that are still empty. Do not ask Modify / Keep / Stop. Do not say the new value waits until next turn. Do not load model-ml-pipeline or git-close on this turn. When they are changing a constraint and did not name a decision, AskUserQuestion one pick among the filled decisions (skip n/a) and stop. Do not also ask for a typed answer. Do not --revise and do not edit the journal on that turn. JSON action is authoritative. Do not invent a menu. If the command is missing or exits without JSON, follow Fallback below. Do not open another reference. Do not guess candidates.

3. stop — say the JSON reason and stop.
4. ask / uncovered — follow Fallback below, including the table written in this message and the Status rule there. Say the reuse and change lines, and say there is no splitter translation. Do not load build-ml-pipeline. Stop this turn.
5. ask / missing_keys — read each distinct reference in questions once before asking. Do not open any other file under references/. Ask every key in missing in one message. For a question that has candidates, those are the options. When candidates is absent, ask for the value the reference describes. Draw on three sources, and only what they actually say: the EDA report, free-form text that came with the data if any is present (notes, a dictionary, or a README beside the raw files), and facts the user stated. If none of that text is present, do not invent it. When one of them already states the fact, quote it in the question. Write every Value the user answered in this turn. Do not rename the Variable column. Do not stop after the first decision. When the deployment makes other rows inapplicable, set those rows to n/a in the same edit. Horizon, gap, and time role are n/a unless deployment is time. Generalize-to is n/a unless deployment is groups. A fold count of 1 is one train/test split drawn from a single table. When the EDA report, the text shipped with the data, or the user already names a separate training table and test table, offer using that split in the folds question and write predefined if they choose it. Do not offer it otherwise. Do not write prefit in the table. Do not type Revised on; only that command writes that date. If any key in missing is still unanswered, set Status to draft once any decision is filled, ask those keys, and stop. Do not invent their values. Do not set Status to locked. Do not ask to confirm the table. If the write fills every required decision, set Status to locked in that same edit, not draft. Run python -m skore_skills frame show again in this turn. On proceed, say the reuse and change lines, then follow step 8. On ask / set, write Status locked only, say those lines, run frame show again, and follow step 8. If that frame show still returns missing_keys, the table was not complete: ask those keys and stop, and leave Status draft. Reuse and change lines, quoting JSON context in 2–4 lines: these choices are reused for the rest of the experiment so models stay comparable, and any one of them can be changed by naming it (for example the comparison metric). Do not say "lock" in those lines. Do not AskUserQuestion.
6. When the user named one decision and Status is draft, do not treat set as accepting the table. Show the same fenced command for that decision and stop. Do not write the new value. Do not say any other decision was dropped. The command's JSON list of dropped keys is the record. Status stays draft. That command stamps Revised on; do not type that date. The next frame show asks only keys that are still empty or invalid.
7. ask / set — the table was already complete and Status is still draft. Write Status locked only. Say the reuse and change lines from step 5. Run frame show again and follow step 8. Do not AskUserQuestion. The user sentence that opened this screen is not a choice. ask / revise is not a user question. Do not present Modify / Keep / Stop. Ignore those choices. Drop the named decision and ask the replacement, as in step 2.
8. proceed — the table is locked, and the user is not changing a named decision. If they are, step 2 already handled it. If translation is null, say that this lock has no splitter translation. Do not load build-ml-pipeline and do not return to model-ml-pipeline. Stop. If model-ml-pipeline dispatched this turn, return to that coordinator and stop. Do not start build, write a design note, or run the git close from here. If no experiment script exists and History has no running, done, or abandoned model row, and status.skills.model-ml-pipeline is true, load that skill and stop. Do not write a design note here. Do not git end-turn. Otherwise run 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. Otherwise load triage-ml-task only if that skill is installed.

Live preview

After any turn that changes journal/JOURNAL.md (filled decisions, Status, that command, or lock), coalesce all edits, then, when policy.site is true and export-ml-site is installed, run python -m skore_skills site build --if-stale before the next question or stop. In that message, say the website has been updated and paste the file:// line from that output verbatim. Do not name journal/JOURNAL.md. policy.site null or false does not enable or build the site. A read-only frame show with no Markdown change does not build.

Stop conditions

• Do not write Python, a pipeline, a test, or a design note.
• Do not put a class name or a constructor argument in the journal. TimeSeriesSplit, KFold, GroupKFold, and gap= stay out of the table.
• Do not re-ask a key that is absent from missing.
• Do not add an option that is absent from candidates.
• Do not open a reference the JSON did not name. When the command is missing, follow Fallback below instead of opening a file.
• A recorded decision changes when that command drops it, then filling it like any missing one. A complete fill sets Status to locked again. Do not ask Modify / Keep / Stop.
• That command is the only journal edit that drops a recorded decision, and only for the decision the user named. It stamps Revised on; do not type that date. Do not rewrite experiments/, audit/, or a report in this skill.
• After that decision is dropped, do not run an existing experiment script. Say that it still uses the previous splitter and metric. Show the command once, in a fence. The next build or evaluate rewrites it after the table is complete again.

Fallback

Use this when frame show is missing, or when the problem is not classification, regression, or multi-output regression. Do not open any other file under references/. Do not invent the closed candidate menu.

Tell the user, in a few lines, that the closed menu does not cover this case, so the comparison has to be written in words. Ask one question: what would count as a better result, and what is an honest baseline. Draw on the EDA report, free-form text that came with the data if any is present, and facts they stated. Quote a fact those sources already give. If that text is absent, do not invent it.

Write the table in this same message:

• Prediction goal: uncovered
• Metric: the comparison, in their words, or empty if they did not state it
• Baseline note: the baseline, in their words, or empty if they did not state it
• Every other decision row: n/a, unless they already stated a value

Do not use placeholders. Do not leave the table for a later message. Status is locked only when both the comparison and the baseline are already stated. Otherwise Status is draft. The words lock, modify, and stop stay out of the message.

Do not name a splitter class or write code. Say these choices are reused for the rest of the experiment so models stay comparable, and that any one of them can be changed by naming it. There is no splitter translation.

If they say it is actually classification or regression, write that goal instead of uncovered and stop. Do not set Status to locked. The next frame show uses the closed menu.

If the command is missing, there is no JSON. The same write uses the same lines. There is no splitter translation.