[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M6MHPYM046ZFPMNSPBKPXBBG":3,"skill-same:01M6MHPYM046ZFPMNSPBKPXBBG":75},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":40,"linkOnly":41,"markdownContent":42,"markdownBody":43,"frontMatter":44,"refName":45,"commitSha":46,"skillDirectory":47,"forksCount":48,"repositoryHtmlUrl":49,"repositoryHomepageUrl":50,"ownerHtmlUrl":51,"ownerBlogUrl":52,"ownerDescription":53,"updatedAt":54,"install":55,"resources":68},"01M6MHPYM046ZFPMNSPBKPXBBG","ml-pipeline",null,"MANDATORY whenever a task involves training, fine-tuning, tuning, or evaluating a machine-learning model on data (tabular, time series, text, images — any modality). Enforces a strict 16-step pipeline that starts with inspecting the raw data, gates each phase behind the user's explicit permission, and produces marimo notebooks with matplotlib visuals so the user can see and understand every step. Never jump straight to model training.","hashgraph-online\u002Fawesome-codex-plugins","hashgraph-online","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F160371962?v=4","01M6CNB6NRJVNNGN14TR01Q6D3","HOL (Hashgraph Online)","01M6JT1HSKKG9V68KQAQGBA7XP",1575,"Apache-2.0","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fjananthan30\u002Fml-pipeline\u002Fskills\u002Fml-pipeline\u002FSKILL.md",[],[19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39],"ml","pipeline","ml pipeline","phase","visuals","can","machine","marimo","series","permission","time","images","never","model","text","learning","tuning","data","Fine-Tuning","time series","training","2026-11-10 10:12:30",false,"---\nname: ml-pipeline\ndescription: MANDATORY whenever a task involves training, fine-tuning, tuning, or evaluating a machine-learning model on data (tabular, time series, text, images — any modality). Enforces a strict 16-step pipeline that starts with inspecting the raw data, gates each phase behind the user's explicit permission, and produces marimo notebooks with matplotlib visuals so the user can see and understand every step. Never jump straight to model training.\n---\n\n# ML Pipeline Discipline\n\n**Hard rule: no model is trained until every earlier step in the pipeline is done and the\nuser has explicitly approved the phase gates before it.** \"Train a model on this data\" is a\nrequest to *start the pipeline at step 1*, not at step 10.\n\n## The pipeline (strict order — never reorder, never skip silently)\n\n```\nRAW DATA\n  → 1. Data inspection\n  → 2. Exploratory data analysis (EDA)\n  → 3. Define the prediction problem\n  ──────────── GATE A: user approval ────────────\n  → 4. Data cleaning\n  → 5. Data engineering\n  → 6. Train \u002F validation \u002F test split\n  → 7. Feature engineering\n  → 8. Preprocessing\n  ──────────── GATE B: user approval ────────────\n  → 9. Baseline model\n  → 10. Model training\n  → 11. Hyperparameter tuning\n  → 12. Model evaluation\n  ──────────── GATE C: user approval ────────────\n  → 13. Error analysis\n  → 14. Final test (test set touched ONCE)\n  → 15. Deployment (only if user asks)\n  → 16. Monitoring + retraining plan\n  ──────────── GATE D: wrap-up report ───────────\n```\n\n## Phase gates — explicit permission, every time\n\nAt each gate, STOP and give the user, in plain non-jargon language:\n\n1. **What was done** — the steps completed, 1–2 sentences each.\n2. **What was found** — key findings, with the visuals that show them.\n3. **Decisions made and why** — e.g. \"dropped 312 duplicate rows\", \"chose time-based split\n   because the data has dates\".\n4. **What comes next** — the next phase's steps, in one short list. At Gate B this section\n   includes the `Model rationale:` block (format under Progress tracking).\n5. **The question** — ask for explicit permission to continue. Wait for a clear yes.\n   Silence, ambiguity, or \"hmm\" is not a yes. If the user redirects, incorporate it.\n\nIf the user says \"skip ahead\" or \"just train it\": explain in 2–3 sentences which steps are\nmissing and the concrete risk (usually leakage or garbage-in), then ask once for explicit\noverride confirmation. If they confirm, proceed and record the override in PIPELINE.md.\n\n## Progress tracking (survives across sessions)\n\nOn first use in a project, create `ml_pipeline\u002FPIPELINE.md` — a checklist of the 16 steps\nwith status (`todo \u002F in progress \u002F done \u002F approved-gate \u002F overridden`), one line of results\nper finished step, and dated gate approvals. Update it after every step. On any new session,\nread it first and resume from the first unfinished step — never restart, never skip ahead\nof it.\n\nRecord each gate approval on its own line in exactly this shape — the enforcement hook\nparses it:\n\n```\n- Gate A: approved 2026-09-16\n- Gate B: approved 2026-09-17\n- Override: Gate B - user approved skipping to training 2026-09-17 - reason: \u003Cwhy>\n```\n\nA line that says a gate is `todo`, `pending`, or `not yet approved` never counts as approval.\n\nAfter recording a gate approval in a git repository, checkpoint it:\n`git add -A && git commit -m \"ml-pipeline: Gate X approved\" && git tag -f gate-X`. Every approved\nphase is then a reproducible point to return to.\n\nBefore recording `Gate B: approved`, PIPELINE.md must contain a model rationale in exactly this shape\n(the hook checks the five bullets exist; the user judges their content):\n\n```\nModel rationale:\n- traits: binary, 1,428 rows (small), minority 11.4%, has_datetime, no groups\n- baseline: majority-class dummy + logistic regression (class_weight=balanced)\n- candidates: gradient boosting with small trees; regularized logistic\n- ruled out: neural nets (small data); k-NN (mixed scales, weak on tabular); random split (temporal)\n- metric: PR-AUC primary, recall at fixed precision secondary\n```\n\n## Tools: marimo notebooks + matplotlib visuals\n\n- **The workbench is a marimo notebook**, not loose scripts. Keep notebooks in\n  `ml_pipeline\u002F`: `01_eda.py` (steps 1–3), `02_prep.py` (steps 4–8), `03_model.py`\n  (steps 9–12), `04_eval.py` (steps 13–16). In Claude Code, drive them live with the\n  `marimo-pair` skill so the user watches the work happen. In other harnesses, write the\n  notebook files and tell the user to open them with `marimo edit \u003Cfile>`.\n- **Every step that looks at data produces matplotlib figures** (seaborn on top is fine).\n  Also save each figure to `ml_pipeline\u002Ffigures\u002F\u003Cstep>_\u003Cname>.png` so gates can reference\n  them even without a live notebook.\n- **Explain every figure in 1–2 plain sentences**: what it shows and why it matters for\n  the next decision. A figure without an explanation is not done.\n\n## What each step must produce\n\n1. **Data inspection** — copy the plugin's guard library to `ml_pipeline\u002Fguard.py` (its path is given at\n   session start; in Codex\u002FKimi it is `skills\u002Fml-pipeline\u002Flib\u002Fmlpipeline_guard.py` next to this file),\n   load the raw data read-only, and run `guard.profile(df, target=..., time_col=..., group_col=...)`.\n   Read `ml_pipeline\u002Fdata_profile.json` before anything else: its `leakage_suspects` and `traits`\n   decide the split, the metric, and the model family. Report rows × columns, column kinds, missing\n   values, duplicates, and every leakage suspect. For images or text, build a manifest table first\n   (path, label, size or length) and profile that. No modification yet.\n2. **EDA** — `guard.eda_figures(df, target=..., time_col=...)` writes the required figures\n   (`01_missingness`, `02_target_balance`, `02_distributions`, `02_correlations`, `02_temporal_coverage`\n   when a date column exists) with explanations computed from the data; add any others with\n   `guard.fig(step, name, figure, explanation)`. If matplotlib is missing, install it — text\n   descriptions are not figures and the gate will not accept them. Output: the figures, your\n   interpretation of each, and a short list of hypotheses and problems spotted.\n3. **Define the prediction problem** — write a short contract: target (exact definition,\n   units), prediction unit and population, prediction time\u002Fhorizon, information actually\n   available at prediction time, objective, evaluation metric, constraints. **The user must\n   approve this contract at Gate A** — it controls everything after. Consult\n   `references\u002Fmodel-selection.md` with the profile's traits when choosing the metric and the\n   candidate model families.\n4. **Data cleaning** — missing values, duplicates, invalid\u002Fimpossible values, inconsistent\n   categories, unit\u002Fformat issues. Report before\u002Fafter counts for every rule. Document every\n   rule in PIPELINE.md. Prefer preserving data over deleting. Never use information from the\n   future or from the test rows to decide a cleaning rule.\n5. **Data engineering** — joins\u002Fintegration, aggregation, time alignment to an index date,\n   business rules, one-row-per-prediction-unit feature table, data-quality checks\n   (row counts, uniqueness, ranges).\n6. **Split before any fitting** — copy the plugin's guard library to `ml_pipeline\u002Fguard.py`\n   (its path is given at session start; in Codex\u002FKimi it is `skills\u002Fml-pipeline\u002Flib\u002Fmlpipeline_guard.py`\n   next to this file) and split with it:\n   `train, val, test = guard.split(df, target=..., time_col=\u003Ccol> if the data is temporal, group_col=\u003Ccol> if the same entity appears in multiple rows)`.\n   It drops exact duplicates, refuses a random split when a datetime column exists, keeps groups\n   together, checks that no row lands in two splits, and freezes a fingerprint of the test set.\n   The test set is touched exactly once, at step 14, through `guard.final_test()`.\n7. **Feature engineering** — design features on the training set's statistics only, then\n   apply the same transformations to validation\u002Ftest.\n8. **Preprocessing** — scalers, encoders, imputers fit on train only, wrapped in a pipeline\n   object so train and inference can never diverge.\n9. **Baseline first** — a dummy predictor (majority class \u002F mean) AND one simple model\n   (logistic\u002Flinear regression or small tree). Record their metrics. Every later model is\n   judged against this line; a complex model that can't beat it gets reported as such.\n10. **Model training** — train candidate models on train, compare on validation. Log every\n    run's config and score.\n11. **Hyperparameter tuning** — on validation\u002Fcross-validation only. The test set is never\n    part of tuning.\n12. **Model evaluation** — the contract's metric plus supporting views: confusion matrix and\n    ROC\u002FPR curves for classification, residual plots for regression, always compared to the\n    baseline. Figures required.\n13. **Error analysis** — worst predictions, performance by slice\u002Fsubgroup, calibration,\n    where the model fails and a hypothesis for why.\n14. **Final test** — `score = guard.final_test(model.predict, test, target=..., metric_fn=...)`.\n    It verifies the frame is the frozen test set, refuses a second call, and logs the result to\n    PIPELINE.md. Report the number honestly, even if it is worse than validation. No going back to\n    tune on it — if the result forces changes, agree a new test strategy with the user and record\n    `- Override: final test - user approved a second evaluation YYYY-MM-DD - reason: \u003Cwhy>`.\n15. **Deployment** — only when the user asks. Save the full pipeline artifact\n    (preprocessing + model together), verify a reloaded artifact reproduces predictions,\n    document the inference input contract.\n16. **Monitoring + retraining** — write down: what drift to watch (input and prediction\n    distributions), what metric threshold triggers retraining, and how retraining reuses\n    this same pipeline from step 1.\n\n## Enforcement in Claude Code (hook)\n\nInstalled as a Claude Code plugin, `hooks\u002Fguard_training.py` runs before every Bash, Write,\nEdit, and notebook tool call, scans the code for fitting and training calls, and **denies**:\n\n- **model training** until PIPELINE.md has `Gate B: approved …` or an `Override: Gate B …`\n  line. With no `ml_pipeline\u002FPIPELINE.md` at all, training is denied with a pointer to step 1.\n- **any fitting** — scalers, encoders, imputers included — until `Gate A: approved …`.\n- **fitting on the test split** (`X_test`, `df_test`, `test_*`) — always, at every gate.\n- **evaluating on the test split** — `.predict\u002F.score` or a metric function whose arguments name\n  `X_test`, `df_test`, `test_*` — until `Gate C: approved …` (or `Override: Gate C …`).\n- **recording a gate before its artifacts exist** — `Gate A` needs `data_profile.json`, one `01_*.png`\n  and two `02_*.png`; `Gate B` needs a `04_*.png` and the `Model rationale:` block; `Gate C` a\n  `12_*.png`; `Gate D` a `13_*.png` and the `Step 14 final test:` line. Every PNG must be larger than\n  1 KB and have its `- Figure \u003Cfile>:` line. The denial lists exactly what is missing.\n- **red flags the profile can prove** — `neural-net-small-data`, `random-split-temporal`,\n  `group-split`, `accuracy-imbalanced`, `resample-before-split` (see `references\u002Fmodel-selection.md`).\n  Override one only after the user agreed: `- Override: red flag \u003Cname> - user approved \u003Cwhat> YYYY-MM-DD - reason: \u003Cwhy>`.\n\nA PostToolUse hook also warns (never blocks) when a result looks too good to be honest\n(accuracy\u002FAUC\u002FF1 ≥ 0.98) or when `train_test_split()` is called without `stratify=` or on data that\nmentions dates. To prove a pipeline does not leak, run it on `examples\u002Fcanary\u002Fcanary.csv`: honest\ntest accuracy cannot exceed 0.80, and `mlpipeline_canary.verdict(score, n_test)` says whether a\nnumber is above the ceiling.\n\nA denial is not an obstacle to route around: do the missing steps, get the user's explicit\napproval, record the gate line, and retry. The hook fails open on its own errors, and\n`ML_PIPELINE_ENFORCE=0` switches it off for projects that are not ML pipelines. Codex and Kimi\nhave no hook support, so there the same rules apply as instructions only.\n\n## Non-negotiables\n\n- Selection leakage is the one that matters most: the test set is never used to pick a model, a\n  seed, a feature, or a threshold. Validation only.\n- Every gate is recorded with its figures on disk and their explanations in PIPELINE.md. A figure\n  that was described but not drawn does not exist.\n- Test set is used exactly once. No tuning, no peeking, no \"just checking\".\n- All fitting (cleaning statistics, features, preprocessing, models) uses training data only.\n- Temporal data gets temporal splits; repeated entities get group splits.\n- Baseline before any complex model; every result is reported relative to it.\n- Failures and disappointing numbers are reported plainly — never hidden or reframed.\n- Chat explanations stay beginner-friendly; the code stays production-grade.\n","# ML Pipeline Discipline\n\n**Hard rule: no model is trained until every earlier step in the pipeline is done and the\nuser has explicitly approved the phase gates before it.** \"Train a model on this data\" is a\nrequest to *start the pipeline at step 1*, not at step 10.\n\n## The pipeline (strict order — never reorder, never skip silently)\n\n```\nRAW DATA\n  → 1. Data inspection\n  → 2. Exploratory data analysis (EDA)\n  → 3. Define the prediction problem\n  ──────────── GATE A: user approval ────────────\n  → 4. Data cleaning\n  → 5. Data engineering\n  → 6. Train \u002F validation \u002F test split\n  → 7. Feature engineering\n  → 8. Preprocessing\n  ──────────── GATE B: user approval ────────────\n  → 9. Baseline model\n  → 10. Model training\n  → 11. Hyperparameter tuning\n  → 12. Model evaluation\n  ──────────── GATE C: user approval ────────────\n  → 13. Error analysis\n  → 14. Final test (test set touched ONCE)\n  → 15. Deployment (only if user asks)\n  → 16. Monitoring + retraining plan\n  ──────────── GATE D: wrap-up report ───────────\n```\n\n## Phase gates — explicit permission, every time\n\nAt each gate, STOP and give the user, in plain non-jargon language:\n\n1. **What was done** — the steps completed, 1–2 sentences each.\n2. **What was found** — key findings, with the visuals that show them.\n3. **Decisions made and why** — e.g. \"dropped 312 duplicate rows\", \"chose time-based split\n   because the data has dates\".\n4. **What comes next** — the next phase's steps, in one short list. At Gate B this section\n   includes the `Model rationale:` block (format under Progress tracking).\n5. **The question** — ask for explicit permission to continue. Wait for a clear yes.\n   Silence, ambiguity, or \"hmm\" is not a yes. If the user redirects, incorporate it.\n\nIf the user says \"skip ahead\" or \"just train it\": explain in 2–3 sentences which steps are\nmissing and the concrete risk (usually leakage or garbage-in), then ask once for explicit\noverride confirmation. If they confirm, proceed and record the override in PIPELINE.md.\n\n## Progress tracking (survives across sessions)\n\nOn first use in a project, create `ml_pipeline\u002FPIPELINE.md` — a checklist of the 16 steps\nwith status (`todo \u002F in progress \u002F done \u002F approved-gate \u002F overridden`), one line of results\nper finished step, and dated gate approvals. Update it after every step. On any new session,\nread it first and resume from the first unfinished step — never restart, never skip ahead\nof it.\n\nRecord each gate approval on its own line in exactly this shape — the enforcement hook\nparses it:\n\n```\n- Gate A: approved 2026-09-16\n- Gate B: approved 2026-09-17\n- Override: Gate B - user approved skipping to training 2026-09-17 - reason: \u003Cwhy>\n```\n\nA line that says a gate is `todo`, `pending`, or `not yet approved` never counts as approval.\n\nAfter recording a gate approval in a git repository, checkpoint it:\n`git add -A && git commit -m \"ml-pipeline: Gate X approved\" && git tag -f gate-X`. Every approved\nphase is then a reproducible point to return to.\n\nBefore recording `Gate B: approved`, PIPELINE.md must contain a model rationale in exactly this shape\n(the hook checks the five bullets exist; the user judges their content):\n\n```\nModel rationale:\n- traits: binary, 1,428 rows (small), minority 11.4%, has_datetime, no groups\n- baseline: majority-class dummy + logistic regression (class_weight=balanced)\n- candidates: gradient boosting with small trees; regularized logistic\n- ruled out: neural nets (small data); k-NN (mixed scales, weak on tabular); random split (temporal)\n- metric: PR-AUC primary, recall at fixed precision secondary\n```\n\n## Tools: marimo notebooks + matplotlib visuals\n\n- **The workbench is a marimo notebook**, not loose scripts. Keep notebooks in\n  `ml_pipeline\u002F`: `01_eda.py` (steps 1–3), `02_prep.py` (steps 4–8), `03_model.py`\n  (steps 9–12), `04_eval.py` (steps 13–16). In Claude Code, drive them live with the\n  `marimo-pair` skill so the user watches the work happen. In other harnesses, write the\n  notebook files and tell the user to open them with `marimo edit \u003Cfile>`.\n- **Every step that looks at data produces matplotlib figures** (seaborn on top is fine).\n  Also save each figure to `ml_pipeline\u002Ffigures\u002F\u003Cstep>_\u003Cname>.png` so gates can reference\n  them even without a live notebook.\n- **Explain every figure in 1–2 plain sentences**: what it shows and why it matters for\n  the next decision. A figure without an explanation is not done.\n\n## What each step must produce\n\n1. **Data inspection** — copy the plugin's guard library to `ml_pipeline\u002Fguard.py` (its path is given at\n   session start; in Codex\u002FKimi it is `skills\u002Fml-pipeline\u002Flib\u002Fmlpipeline_guard.py` next to this file),\n   load the raw data read-only, and run `guard.profile(df, target=..., time_col=..., group_col=...)`.\n   Read `ml_pipeline\u002Fdata_profile.json` before anything else: its `leakage_suspects` and `traits`\n   decide the split, the metric, and the model family. Report rows × columns, column kinds, missing\n   values, duplicates, and every leakage suspect. For images or text, build a manifest table first\n   (path, label, size or length) and profile that. No modification yet.\n2. **EDA** — `guard.eda_figures(df, target=..., time_col=...)` writes the required figures\n   (`01_missingness`, `02_target_balance`, `02_distributions`, `02_correlations`, `02_temporal_coverage`\n   when a date column exists) with explanations computed from the data; add any others with\n   `guard.fig(step, name, figure, explanation)`. If matplotlib is missing, install it — text\n   descriptions are not figures and the gate will not accept them. Output: the figures, your\n   interpretation of each, and a short list of hypotheses and problems spotted.\n3. **Define the prediction problem** — write a short contract: target (exact definition,\n   units), prediction unit and population, prediction time\u002Fhorizon, information actually\n   available at prediction time, objective, evaluation metric, constraints. **The user must\n   approve this contract at Gate A** — it controls everything after. Consult\n   `references\u002Fmodel-selection.md` with the profile's traits when choosing the metric and the\n   candidate model families.\n4. **Data cleaning** — missing values, duplicates, invalid\u002Fimpossible values, inconsistent\n   categories, unit\u002Fformat issues. Report before\u002Fafter counts for every rule. Document every\n   rule in PIPELINE.md. Prefer preserving data over deleting. Never use information from the\n   future or from the test rows to decide a cleaning rule.\n5. **Data engineering** — joins\u002Fintegration, aggregation, time alignment to an index date,\n   business rules, one-row-per-prediction-unit feature table, data-quality checks\n   (row counts, uniqueness, ranges).\n6. **Split before any fitting** — copy the plugin's guard library to `ml_pipeline\u002Fguard.py`\n   (its path is given at session start; in Codex\u002FKimi it is `skills\u002Fml-pipeline\u002Flib\u002Fmlpipeline_guard.py`\n   next to this file) and split with it:\n   `train, val, test = guard.split(df, target=..., time_col=\u003Ccol> if the data is temporal, group_col=\u003Ccol> if the same entity appears in multiple rows)`.\n   It drops exact duplicates, refuses a random split when a datetime column exists, keeps groups\n   together, checks that no row lands in two splits, and freezes a fingerprint of the test set.\n   The test set is touched exactly once, at step 14, through `guard.final_test()`.\n7. **Feature engineering** — design features on the training set's statistics only, then\n   apply the same transformations to validation\u002Ftest.\n8. **Preprocessing** — scalers, encoders, imputers fit on train only, wrapped in a pipeline\n   object so train and inference can never diverge.\n9. **Baseline first** — a dummy predictor (majority class \u002F mean) AND one simple model\n   (logistic\u002Flinear regression or small tree). Record their metrics. Every later model is\n   judged against this line; a complex model that can't beat it gets reported as such.\n10. **Model training** — train candidate models on train, compare on validation. Log every\n    run's config and score.\n11. **Hyperparameter tuning** — on validation\u002Fcross-validation only. The test set is never\n    part of tuning.\n12. **Model evaluation** — the contract's metric plus supporting views: confusion matrix and\n    ROC\u002FPR curves for classification, residual plots for regression, always compared to the\n    baseline. Figures required.\n13. **Error analysis** — worst predictions, performance by slice\u002Fsubgroup, calibration,\n    where the model fails and a hypothesis for why.\n14. **Final test** — `score = guard.final_test(model.predict, test, target=..., metric_fn=...)`.\n    It verifies the frame is the frozen test set, refuses a second call, and logs the result to\n    PIPELINE.md. Report the number honestly, even if it is worse than validation. No going back to\n    tune on it — if the result forces changes, agree a new test strategy with the user and record\n    `- Override: final test - user approved a second evaluation YYYY-MM-DD - reason: \u003Cwhy>`.\n15. **Deployment** — only when the user asks. Save the full pipeline artifact\n    (preprocessing + model together), verify a reloaded artifact reproduces predictions,\n    document the inference input contract.\n16. **Monitoring + retraining** — write down: what drift to watch (input and prediction\n    distributions), what metric threshold triggers retraining, and how retraining reuses\n    this same pipeline from step 1.\n\n## Enforcement in Claude Code (hook)\n\nInstalled as a Claude Code plugin, `hooks\u002Fguard_training.py` runs before every Bash, Write,\nEdit, and notebook tool call, scans the code for fitting and training calls, and **denies**:\n\n- **model training** until PIPELINE.md has `Gate B: approved …` or an `Override: Gate B …`\n  line. With no `ml_pipeline\u002FPIPELINE.md` at all, training is denied with a pointer to step 1.\n- **any fitting** — scalers, encoders, imputers included — until `Gate A: approved …`.\n- **fitting on the test split** (`X_test`, `df_test`, `test_*`) — always, at every gate.\n- **evaluating on the test split** — `.predict\u002F.score` or a metric function whose arguments name\n  `X_test`, `df_test`, `test_*` — until `Gate C: approved …` (or `Override: Gate C …`).\n- **recording a gate before its artifacts exist** — `Gate A` needs `data_profile.json`, one `01_*.png`\n  and two `02_*.png`; `Gate B` needs a `04_*.png` and the `Model rationale:` block; `Gate C` a\n  `12_*.png`; `Gate D` a `13_*.png` and the `Step 14 final test:` line. Every PNG must be larger than\n  1 KB and have its `- Figure \u003Cfile>:` line. The denial lists exactly what is missing.\n- **red flags the profile can prove** — `neural-net-small-data`, `random-split-temporal`,\n  `group-split`, `accuracy-imbalanced`, `resample-before-split` (see `references\u002Fmodel-selection.md`).\n  Override one only after the user agreed: `- Override: red flag \u003Cname> - user approved \u003Cwhat> YYYY-MM-DD - reason: \u003Cwhy>`.\n\nA PostToolUse hook also warns (never blocks) when a result looks too good to be honest\n(accuracy\u002FAUC\u002FF1 ≥ 0.98) or when `train_test_split()` is called without `stratify=` or on data that\nmentions dates. To prove a pipeline does not leak, run it on `examples\u002Fcanary\u002Fcanary.csv`: honest\ntest accuracy cannot exceed 0.80, and `mlpipeline_canary.verdict(score, n_test)` says whether a\nnumber is above the ceiling.\n\nA denial is not an obstacle to route around: do the missing steps, get the user's explicit\napproval, record the gate line, and retry. The hook fails open on its own errors, and\n`ML_PIPELINE_ENFORCE=0` switches it off for projects that are not ML pipelines. Codex and Kimi\nhave no hook support, so there the same rules apply as instructions only.\n\n## Non-negotiables\n\n- Selection leakage is the one that matters most: the test set is never used to pick a model, a\n  seed, a feature, or a threshold. Validation only.\n- Every gate is recorded with its figures on disk and their explanations in PIPELINE.md. A figure\n  that was described but not drawn does not exist.\n- Test set is used exactly once. No tuning, no peeking, no \"just checking\".\n- All fitting (cleaning statistics, features, preprocessing, models) uses training data only.\n- Temporal data gets temporal splits; repeated entities get group splits.\n- Baseline before any complex model; every result is reported relative to it.\n- Failures and disappointing numbers are reported plainly — never hidden or reframed.\n- Chat explanations stay beginner-friendly; the code stays production-grade.\n",{"name":5,"description":7},"main","2b63f856ce19c2a797929cacabd9e0fc936812fa","plugins\u002Fjananthan30\u002Fml-pipeline\u002Fskills\u002Fml-pipeline",491,"https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins","https:\u002F\u002Fhol.org\u002Fplugins\u002Fbest-codex-plugins","https:\u002F\u002Fgithub.com\u002Fhashgraph-online","https:\u002F\u002Fhol.org","Open-source standards and developer tools for interoperable AI agents: discovery, identity, communication, payments, privacy and security.","2026-11-14 14:06:13",{"source":8,"skillName":5,"commands":56,"prompt":66,"zipAvailable":67},[57,60,63],{"tool":58,"command":59},"npx","npx skills add hashgraph-online\u002Fawesome-codex-plugins --skill ml-pipeline",{"tool":61,"command":62},"bunx","bunx skills add hashgraph-online\u002Fawesome-codex-plugins --skill ml-pipeline",{"tool":64,"command":65},"pnpm","pnpm dlx skills add hashgraph-online\u002Fawesome-codex-plugins --skill ml-pipeline","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fjananthan30\u002Fml-pipeline\u002Fskills\u002Fml-pipeline\u002FSKILL.md",true,[69],{"path":70,"resourceType":71,"resourceTypeDesc":72,"sizeBytes":73,"sourceHtmlUrl":74},"references\u002Fmodel-selection.md","REFERENCE","参考文档",5064,"https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fjananthan30\u002Fml-pipeline\u002Fskills\u002Fml-pipeline\u002Freferences\u002Fmodel-selection.md",[76,86,94,102,110,118],{"skillNo":77,"name":78,"nameZh":6,"description":79,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":80,"licenseFileUrl":81,"sourceHtmlUrl":82,"categories":83,"keywords":84,"publishedAt":85},"01M7DV332YQMGD8WTRJ9DZQXTB","worktable-review-with-annotations","Create, inspect, reply to, update, or resolve situated Worktable annotations. Use when feedback belongs to a Doc section, block, text range, or HTML Doc, or when the user asks to address existing review feedback and coordinate a requested artifact edit.","MIT","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-review-with-annotations\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-review-with-annotations\u002FSKILL.md",[],[],"2026-11-15 14:19:17",{"skillNo":87,"name":88,"nameZh":6,"description":89,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":80,"licenseFileUrl":90,"sourceHtmlUrl":91,"categories":92,"keywords":93,"publishedAt":85},"01M7DV2YGHCZVBEB1AHYB6CGHE","worktable-find-and-synthesize-context","Find, assess, and synthesize context from Worktable Docs and Records. Use when the user asks what a workspace contains, wants workspace-backed research, needs current facts reconciled across artifacts, or requests a cited synthesis without changing workspace state.","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-find-and-synthesize-context\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-find-and-synthesize-context\u002FSKILL.md",[],[],{"skillNo":95,"name":96,"nameZh":6,"description":97,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":80,"licenseFileUrl":98,"sourceHtmlUrl":99,"categories":100,"keywords":101,"publishedAt":85},"01M7DV2TQN1H11ME2SAF9MXN7V","worktable-create-or-update-docs","Create, edit, rename, or maintain narrative Worktable Docs. Use for plans, decisions, briefs, notes, research syntheses, and other durable prose when the user requests a new artifact or a safe update to an existing one.","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-create-or-update-docs\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-create-or-update-docs\u002FSKILL.md",[],[],{"skillNo":103,"name":104,"nameZh":6,"description":105,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":80,"licenseFileUrl":106,"sourceHtmlUrl":107,"categories":108,"keywords":109,"publishedAt":85},"01M7DV2NE452NEA8TXATKSNAJH","worktable-create-or-manage-records","Organize repeated structured information as durable Worktable Records. Use for research datasets, inventories, directories, pipelines, trackers, and other collections whose items need identity, fields, filtering, links, validation, or individual updates.","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-create-or-manage-records\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-create-or-manage-records\u002FSKILL.md",[],[],{"skillNo":111,"name":112,"nameZh":6,"description":113,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":80,"licenseFileUrl":114,"sourceHtmlUrl":115,"categories":116,"keywords":117,"publishedAt":85},"01M7DV2HENRBR2KSCVM064WH0Y","worktable-create-or-manage-interactive-html","Create or improve durable visual and interactive artifacts in Worktable. Use for dashboards, trackers, calculators, mini apps, slide presentations, and other requests that benefit from a polished HTML Doc.","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-create-or-manage-interactive-html\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-create-or-manage-interactive-html\u002FSKILL.md",[],[],{"skillNo":119,"name":120,"nameZh":6,"description":121,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":80,"licenseFileUrl":122,"sourceHtmlUrl":123,"categories":124,"keywords":125,"publishedAt":85},"01M7DV2CADR5R7RQFFYPNAGR9Y","worktable-collaborate-threads","Read, answer, and hand off messages through durable Worktable threads with users or other authenticated participants. Use when the user asks to check addressed messages, reply to a participant, continue a thread, request another agent's help, wait briefly for a response, or watch for comments and messages sent to the agent.","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-collaborate-threads\u002FLICENSE","https:\u002F\u002Fgithub.com\u002Fhashgraph-online\u002Fawesome-codex-plugins\u002Fblob\u002Fmain\u002Fplugins\u002Fworktable\u002Fworktable-plugin\u002Fskills\u002Fworktable-collaborate-threads\u002FSKILL.md",[],[]]