feat(ml): compare convex solvers across scratch and SOTA by jsboige · Pull Request #17960 · jsboige/CoursIA

jsboige · 2026-09-26T13:54:32Z

Grain: DEEP/notebook-python — lane myia-po-2025:CoursIA — prev: MED/tooling #17950

Summary

Add the B.8 convex-optimization capstone in 02-ML-Cours: compare scratch SMO against LIBSVM on the same seeded SVM problem, and ISTA, FISTA, ADMM, scratch coordinate descent, sklearn.Lasso, and CVXPY/CLARABEL on the same seeded Lasso problem.
Measure objective values, numerical certificates, timings, explicitly labeled iteration units and LOC scopes, plus scaling probes in SVM sample count and Lasso feature count. Fixed-budget probes measure work cost, not convergence to a common optimum.
Add three executable exercises with TODO stubs and integrate B.7/B.8 into the series README navigation. Generated catalogs and parent notebooks are unchanged.

Validation

Papermill with Python 3 kernel: 38/38 cells, 14/14 code cells executed, zero errors; executed outputs committed.
validate_pr_notebooks.py: 1/1 notebook passed, 14 code cells.
SVM: scratch and LIBSVM dual objectives 30.3960 and 30.3961; both find 48 support vectors, and the scratch run reproduces the committed 2.7b baseline. LIBSVM exposes 102 working-set selections through n_iter_, distinct from scratch examination calls.
Lasso: all six objective values around 0.360645 on the shared P1 data; reference CVXPY objective is numerical (not a certified exact optimum). The largest positive difference is approximately 4.72e-6 for ADMM. The off-support coordinate-descent KKT check reports 1.23e-2 <= lambda 1.25e-2.
git diff --check: passed.

See #16061. This PR delivers B.8; issue closure remains a separate coordinator decision after checking the entire acceptance.

🤖 Generated with Claude Code

See #16061 Co-Authored-By: Claude-Code <noreply@anthropic.com>

github-actions · 2026-09-26T13:54:48Z

MD hierarchy drift -- `bdd5a1b`

Cette PR augmente le compte de defauts de rendu markdown
par rapport a la base de fusion 4337f62 re-scannee.
Nouveaux defauts imputables au diff :

reference: merge base 4337f62b2eb34d98a27514551d82e692b6bfa82d re-scanned (0 of 1 changed notebook(s) existed at base; others are additions; 0 had findings at base)
  +18  MyIA.AI.Notebooks/ML/DataScienceWithAgents/02-ML-Cours/2.11f-Comparaison-Optimisation-Convexe.ipynb
        +14 H1-DEEP
        +3 HINT-AS-HEADING
        +1 MULTI-H1

=== drift: +18 across 1 notebook(s), 0 burned down ===

Corriger (ex. - # Indice : ... -> - **Indice :** ...). See #11831.

github-actions · 2026-09-26T13:54:48Z

✅ No prose/output mismatch detected in the notebooks this PR changed.

Scope = notebooks CHANGED in this PR, not the whole corpus. Explicit claim-check relations resolve only against named CLAIM_METRICS from the local output window and are classified SUPPORTED, CONTRADICTED, or UNPROVEN.
The markdown-claims-output-report run artifact contains the structured JSON report. See python scripts/check_markdown_claims_output.py --help for re-running locally.
Detector rationale: c.290 / c.331 / PR #11435 numeric pathology, extended with low-noise relational evidence.

github-actions · 2026-09-26T13:55:51Z

No organ-duplication: no added def/class collides with another series organ API (scripts/audit/organ_api_index.yaml).

Detector: python scripts/audit/detect_organ_duplication.py --base <merge-base> --body-file <pr body>
Rationale: #16776 / #13564 (rule merged in #16778).

github-actions · 2026-09-26T13:56:26Z

Notebook outputs-required (H.4 schema): PASS (every code cell carries an `outputs: list`)

github-actions · 2026-09-26T13:56:37Z

Notebook PR Validation: PASS

Notebooks checked: 1
Code cells validated: 14
Result: All passed

Checks: H.1 (no errors), H.3 (execution_count), C.1 (no banned patterns)
Non-Python kernels (.NET/Lean): C.1 + errors only (execution_count advisory)
QuantConnect notebooks: C.1 + errors only (require QC Cloud for execution)

github-actions · 2026-09-26T13:57:44Z

Golden-Set Execution (H.7 P3)

✅ 8/8 notebooks passed (certified reproducible)

Notebook	Status	Time
2.1-Workflow-ML.ipynb	✅ SUCCESS	6.3s
2.2-Descente-de-gradient.ipynb	✅ SUCCESS	6.0s
2.3-Regression-lineaire-logistique.ipynb	✅ SUCCESS	8.0s
2.4-Arbres-Forets-Ensembles.ipynb	✅ SUCCESS	7.8s
Search-01-StateSpace.ipynb	✅ SUCCESS	6.2s
SL-1-LogicalLearning.ipynb	✅ SUCCESS	4.4s
rl_4_multi_armed_bandits.ipynb	✅ SUCCESS	49.8s
GameTheory-04c-NashExistence-Python.ipynb	✅ SUCCESS	6.3s

Pinned lockfile: scripts/notebook_tools/golden_set.lock.txt (H.7 P3, axe A #4208)

See #16061 Co-Authored-By: Claude-Code <noreply@anthropic.com>

github-actions · 2026-09-26T14:07:42Z

MD hierarchy drift -- `480d7f8`

Cette PR augmente le compte de defauts de rendu markdown
par rapport a la base de fusion 4337f62 re-scannee.
Nouveaux defauts imputables au diff :

reference: merge base 4337f62b2eb34d98a27514551d82e692b6bfa82d re-scanned (0 of 1 changed notebook(s) existed at base; others are additions; 0 had findings at base)
  +18  MyIA.AI.Notebooks/ML/DataScienceWithAgents/02-ML-Cours/2.11f-Comparaison-Optimisation-Convexe.ipynb
        +14 H1-DEEP
        +3 HINT-AS-HEADING
        +1 MULTI-H1

=== drift: +18 across 1 notebook(s), 0 burned down ===

Corriger (ex. - # Indice : ... -> - **Indice :** ...). See #11831.

Co-Authored-By: Claude Code <noreply@anthropic.com>

…araison-optimisation-convexe

github-actions · 2026-09-26T14:24:38Z

Path-collision (organ #13359/#13615)

Cette PR #17960 (feat(ml): compare convex solvers across scratch and SOTA) touche au moins un chemin de fichier aussi modifie par d'autres PRs ouvertes. Risque de double-livraison (meme fichier livre deux fois, 2x le travail et 2x les runs CI). Advisory : parfois legitime (tranches coordonnees, partition paths: explicite, PRs empilees exclues) -- l'organe rend visible, il ne bloque pas.

terminal -- feat(ml,#16742): 2.9c-Grokking-Diagrammes-Phases — les 4 phases de R02 (toy effectif, carte multi-seed, MNIST App J, loterie aux axes) #16832 partage : MyIA.AI.Notebooks/ML/DataScienceWithAgents/02-ML-Cours/README.md, recouvrement de chemins 50% -- feat(ml,#16742): 2.9c-Grokking-Diagrammes-Phases — les 4 phases de R02 (toy effectif, carte multi-seed, MNIST App J, loterie aux axes) #16832 est deja sur main.
terminal -- feat(ml,#16746): notebook 2.9e MIPS — du reseau au programme verifie #17019 partage : MyIA.AI.Notebooks/ML/DataScienceWithAgents/02-ML-Cours/README.md, recouvrement de chemins 50% -- feat(ml,#16746): notebook 2.9e MIPS — du reseau au programme verifie #17019 est deja sur main.
faible -- fix(docs,#17518): README 02-ML-Cours -- ligne 2.14b (pont attribution-causalite, DAG simule) #17968 partage : MyIA.AI.Notebooks/ML/DataScienceWithAgents/02-ML-Cours/README.md

Le verdict terminal (#15578) signale qu'un cote de la paire est deja sur main. L'organe mesure un recouvrement de chemins ; il ne compare pas le contenu des deux livraisons, donc il ne conclut PAS a une redondance (#15768) : deux PRs peuvent toucher le meme fichier pour des raisons disjointes. L'arbitrage reste a la lane ou au coordinateur.

clusterManager-Myia

[Hermes] — full read du notebook 2.11f (38 cellules, 14 code, exec 1→14 séquentiel) + diff README, head d84267e.

Anchors vérifiés contre les outputs committés (tous présents) : duals SMO/LIBSVM 30.3960 / 30.3961, |SV| = 48 des deux côtés (et à toutes les tailles de la sonde), n_iter_ = 102 (working set), objectif partagé 0.360645 sur les six solveurs, écart max ADMM 4.72e-06, KKT off-support 1.23e-02 ≤ λ 1.25e-2. Gates #17040 : aucune lecture dupliquée ni mal placée, exercices = stubs propres sans leak de solution, pas de prose de remplissage. README : lignes de navigation + paragraphe de présentation pour 2.11e/2.11f — conforme à la directive #17633 (corps du notebook livré, zéro totaux brassés). Scan sécurité : néant.

Finding mineur (non bloquant) — la « lecture préparée » imprimée par la sonde SVM (§4.1) affirme « compteurs SMO et temps croissent avec n », mais la sortie committée montre des temps SMO non-monotones : 0.24 s → 0.17 s → 0.97 s. Les compteurs (deterministes) croissent bien ; le temps non. La prose markdown du §4 corrige d'ailleurs élégamment cette classe de variance pour le Lasso (« la sonde identifie des échelles, pas des exposants ») — le même hedge manqué sur la ligne imprimée SVM. Fix triviale si le notebook est ré-exécuté un jour ; rien qui bloque le merge.

Checks : PR gate FAIL au moment de la review, mais le run est encore en cours (validate-notebooks in_progress) — à relire après complétion, pas un rouge de contenu identifié (les 19+ organes visibles passent, y compris Gitleaks, guards 16 organes, golden-set).

[Hermes hermes-pr-review, cycle :14 26/09, host f6be46d1b7a3]

Re-execute the B.8 notebook after correcting its printed interpretation of measured timings. Papermill executed all 14 code cells successfully. Co-Authored-By: Claude-Code <noreply@anthropic.com>

jsboige · 2026-09-26T14:36:51Z

La remarque Hermes sur la lecture imprimee de la sonde SVM est traitee dans f4d22bc : les compteurs SMO croissent, tandis que les temps fluctuent aux petites tailles. Le notebook a ete reexecute integralement (14/14 cellules code, 0 erreur) et les sorties reelles sont commitees ; les ratchets sortie/source locaux ne signalent aucune regression. Les checks de la nouvelle tete sont encore en cours.

jsboige · 2026-09-26T17:18:59Z

[ADJOINT PREFLIGHT]
schema: 1
lane: myia-po-2023:CoursIA
pr: 17960
head: f4d22bc
complete: true
body: read
comments-reviewed: 9
reviews-reviewed: 1
threads-reviewed: 0
threads-unresolved: 0
surfaces-sha256: 8831efc0edc1cdb4409168be274af7fc60db326bb9d0d37d54618bb4f3ea65b5
diff-files: 2
diff-additions: 1799
diff-deletions: 0
checks: latest-wins-green
b0: clear
scope: pass
domain: pass
verdict: READY
[/ADJOINT PREFLIGHT]

Dossier tierce (lot DM msg-20260926T163550-203qzu, item 3). Verifications firsthand :

B.0 : check_unaddressed_nits.py 17960 -> rc=0. L'organe signale 1 commentaire posterieur au dernier commit non evalue : lu et juge -- c'est la levree d'auteur (jsboige 14:36:51Z) du finding mineur Hermes (temps SMO non-monotones dans la lecture imprimee de la sonde, review 14:29Z head d84267e). La levree nomme la remarque et cite le commit traitant f4d22bc ; le fix est constate dans le diff au head courant (hedge « compteurs croissent, temps fluctuent a petite taille, ordre de grandeur augmente ») ET dans l'output re-execute commite. Hermes qualifiait le finding « non bloquant » de toute facon.
Checks : pli latest-wins sur commits/f4d22bc/check-runs -- 90 jambes, 0 rouge, aucune jambe advisory en echec. Le « PR gate FAIL » mentionne par Hermes etait au head precedent pendant un run in_progress ; le drift MD hierarchy +18 (bot 14:07Z) etait aussi au head precedent, corrige avant f4d22bc (jambe verte au head courant).
Scope : 2 fichiers (notebook 2.11f + README serie), +1799/-0, conforme au body (3 exercices stubs, nav README, catalogues inchanges -- jambe catalogue verte).
Domaine (notebook-python) : bot validation PASS 14/14 cellules code (H.1/H.3/C.1), golden-set 8/8, ancres Hermes verifiees contre les outputs committes (30.3960/30.3961, |SV|=48, 0.360645, 4.72e-6, KKT 1.23e-2 <= 1.25e-2).

Advisory lu, non bloquant : path-collision organ signale le README partage avec #16832 (deja sur main) -- mergeStateStatus CLEAN au dossier, sans conflit constate.

…atch of the nav-chain guard The gate fired on its first run: 2.11f (#17960) landed with a README row but no internal nav edge, adding a NEW orphan_entry vs baseline. Source fix, not a baseline bump: transition cell with a "Suivant" edge in 2.11e -> 2.11f (markdown-only, no code cell touched). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

…EAU (#17985) * feat(ci,#17284): cable nav-chain guard en organe bloquant sur le NOUVEAU Workflow notebook-nav-chain-guard.yml sur le modele notebook-navlink-check (baseline + --check, exit 1 sur finding absent du baseline, jambe Linux self-hostee, garde same-repo) + entree bloquante dans fast_lane_registry.py + assertion de test. Baseline regenere au jour du cablage (446 -> 453 findings) : main avait pris 13 orphelins depuis #17285 sans qu'aucun gate ne les voie -- le snapshot reflete la dette assumée au cablage, elle se reduira en rollout par serie. Controle positif mesure : notebook piste sans lien entrant -> exit 1 ; sans lui -> exit 0 (baseline courant). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> * fix(ci,#17284): allowlist self-hosted du workflow nav-chain-guard La jambe Scripts Tests rougissait sur test_current_repository_self_hosted_jobs_satisfy_isolation_policy : WORKFLOW_NOT_ALLOWED -- le nouveau workflow self-hosted doit figurer dans SELF_HOSTED_WORKFLOW_ALLOWLIST, comme son modele notebook-navlink-check.yml. 58 passed sur la suite policy. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> * fix(nav,#17284): wire 2.11f into the 2.11 series chain - first live catch of the nav-chain guard The gate fired on its first run: 2.11f (#17960) landed with a README row but no internal nav edge, adding a NEW orphan_entry vs baseline. Source fix, not a baseline bump: transition cell with a "Suivant" edge in 2.11e -> 2.11f (markdown-only, no code cell touched). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> * Fix(nav,#17284): rattacher les 2 orphelins neufs sur main (Complexity-03c, Lean-24b) Deux notebooks arrives sur main apres la tete du cablage etaient en NEW vs baseline (mesure coordinateur 27/09 ~07:40Z sur main 1c8d884) : Complexity-03c (via #17769) et Lean-24b (via #17530). Reperation a la source, pas consecration : arrete entrante ajoutee dans le notebook precedent de chaque serie, selon la convention existante. - Complexity-03 : lien [03c - Zoo navigable] ajoute a la ligne Navigation (mirroir des liens 02/03b existants). - Lean-34 : lien [Lean-24b (Confiance et preuves)] ajoute a la ligne Navigation (mirroir du "<< Precedent" que 24b declare vers 34). - Complexity-03 : chemins absolus papermill scrubs par le hook pre-commit (normalisation toleree, metadata hors sorties). check_notebook_nav_chain.py --check : 0 NEW (453 connus), rc=0. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> * Fix(nav,#17284): rattacher FT-00d et ICT-45 — lignes Navigation predecesseur + miroir Les merges #17506 (FT-00d) et #18017 (ICT-45) ont pose deux orphelins nav-chain (orphan_entry mesure sur main 02fb2e1). Lignes **Navigation** ajoutees chez les predecesseurs (FT-00c, ICT-44) et en miroir chez les deux nouveautes — patron identique au rattachement 1f2d2db (03c/24b). Markdown-only, aucune cellule code touchee. Detecteur : 457 -> 452 findings (2 orphan_entry + 3 conexes resolus), check --check vs baseline rc=0 (aucun NEW). See #17284 Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> * feat(ci,#17284): scope bloquant nav-chain au diff de la PR (decision c.5854935546) La jambe ne rougit plus sur tout NEW finding repo-wide : seuls les NEW imputables a la PR blocquent (notebook du diff, ou serie dont le README est touche) ; un NEW ne du merge d'une AUTRE PR entre-temps devient un avertsement. Fail-closed : diff absent/vide = tout NEW bloque. - check_notebook_nav_chain.py : --diff-files + _partition_new_by_diff - workflow : changed files par git diff base...HEAD (sans GITHUB_TOKEN, posture tranche 1 preservee), fetch-depth 0 - tests : controles positif (notebook du diff, README de serie) et negatif (hors diff) + fail-closed x2 — 22 passed See #17284 Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> * Fix(guard,#17284): imputation independent_chain -- la serie est la cle + baseline des 2 ilots Re-mesure ai-01 (c.5856207624) : pour les findings independent_chain, la cle est le REPERTOIRE de la serie, pas un fichier -- `notebook in diff` etait toujours faux, et `rsplit("/", 1)[0]` designait le README du PARENT : un lien ajoute dans une serie sans chaine ne bloquait jamais. _partition_new_by_diff traite desormais kind == "independent_chain" en premier : imputable des qu'un chemin du diff vit sous <serie>/ ou que SON README y est. 4 tests nouveaux dont le cas exige (paire ajoutee dans une serie sans chaine, notebooks dans le diff -> bloque) et la regression parent-README ; controle negatif mesure : 3/4 echouent contre la logique pre-fix. Suite 26/26 pass. Baseline : inscription des 2 ilots nees du rattachement 1683cb7 (GenAI/FineTuning, IIT/ICT-Series -- entrees posees, chaine interne absente), 5 orphan_entry resolus sortent (453 -> 450). Dette declaree, suivi reel : #18057. See #17284 See #18057 Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> * fix(#17985): qualifier les 2 annotations seuil (~450 -> 450 exact, non normatif) Rouge n°2 de la review NanoClaw : les deux littéraux numeriques en commentaire sont alignes sur la mesure (450 entrees au head) et marques informatifs -- le JSON baseline_nb_nav_chain.json est nomme source de verite. 81 tests fast_lane verts. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com>

feat(ml): compare convex solvers across scratch and SOTA

bdd5a1b

See #16061 Co-Authored-By: Claude-Code <noreply@anthropic.com>

github-actions Bot added the paragraph-length Paragraph > 2000 chars (wall-of-text, #15405). Resorb before merge. label Sep 26, 2026

github-actions Bot added the variation-adjacency-deep-med Adjacence DEEP/MED hors LIGHT : §2 l'autorise si substance distincte (coordinateur) label Sep 26, 2026

fix(ml): keep measured LOC counts out of notebook prose

480d7f8

See #16061 Co-Authored-By: Claude-Code <noreply@anthropic.com>

github-actions Bot removed the variation-adjacency-deep-med Adjacence DEEP/MED hors LIGHT : §2 l'autorise si substance distincte (coordinateur) label Sep 26, 2026

jsboige and others added 2 commits September 26, 2026 16:20

fix(ml): render exercise guidance as body text

4c1b57e

Co-Authored-By: Claude Code <noreply@anthropic.com>

Merge remote-tracking branch 'origin/main' into feature/16061-b8-comp…

d84267e

…araison-optimisation-convexe

clusterManager-Myia reviewed Sep 26, 2026

View reviewed changes

fix(ml): describe SMO timing variation accurately

f4d22bc

Re-execute the B.8 notebook after correcting its printed interpretation of measured timings. Papermill executed all 14 code cells successfully. Co-Authored-By: Claude-Code <noreply@anthropic.com>

This was referenced Sep 26, 2026

[Search-07] Recyclage d'un stash : variante misere du Nim et minimax memoise — livrer ou refuser #17937

Closed

feat(search): comparer Nim normal et misère avec témoin minimax #17967

Merged

github-actions Bot mentioned this pull request Sep 26, 2026

fix(docs,#17518): README 02-ML-Cours -- ligne 2.14b (pont attribution-causalite, DAG simule) #17968

Merged

5 tasks done

jsboige mentioned this pull request Sep 26, 2026

ci: PR gate FAILURE en rafale — quota GitHub App épuisé (HTTP 403 installation), distinct de #13363 (pool) #17681

Closed

myia-ai-01 merged commit 4cf7559 into main Sep 26, 2026
90 of 92 checks passed

This was referenced Sep 26, 2026

fix(ci): back off installation quota reads in PR gate #17987

Merged

[Search/Optim] Optimisation convexe avancée from scratch : SMO, ADMM, opérateurs proximaux #16061

Closed

jsboige mentioned this pull request Sep 26, 2026

feat(ci,#17284): cable nav-chain guard en organe bloquant sur le NOUVEAU #17985

Merged

myia-ai-01 mentioned this pull request Sep 27, 2026

closure(lot B): 18 issues candidates a la fermeture — dossiers [CLOSURE PREFLIGHT] attendus (audit haiku 27/09) #18068

Open

github-actions Bot mentioned this pull request Sep 29, 2026

fix(ml,#18331): requirements.txt pour 02-ML-Cours (optuna, shap, lime, dice-ml, diffprivlib, cvxpy, ...) #18378

Merged

Navigation Menu

feat(ml): compare convex solvers across scratch and SOTA - #17960

feat(ml): compare convex solvers across scratch and SOTA#17960
myia-ai-01 merged 5 commits into
mainfrom
feature/16061-b8-comparaison-optimisation-convexe

jsboige commented Sep 26, 2026

Uh oh!

github-actions Bot commented Sep 26, 2026

Uh oh!

github-actions Bot commented Sep 26, 2026

Uh oh!

github-actions Bot commented Sep 26, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Sep 26, 2026

Uh oh!

github-actions Bot commented Sep 26, 2026

Uh oh!

github-actions Bot commented Sep 26, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Sep 26, 2026

Uh oh!

github-actions Bot commented Sep 26, 2026 •

edited

Loading

Uh oh!

clusterManager-Myia left a comment

Uh oh!

jsboige commented Sep 26, 2026

Uh oh!

jsboige commented Sep 26, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

jsboige commented Sep 26, 2026

Summary

Validation

Uh oh!

github-actions Bot commented Sep 26, 2026

MD hierarchy drift -- bdd5a1b

Uh oh!

github-actions Bot commented Sep 26, 2026

Uh oh!

github-actions Bot commented Sep 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot commented Sep 26, 2026

Notebook outputs-required (H.4 schema): PASS (every code cell carries an outputs: list)

Uh oh!

github-actions Bot commented Sep 26, 2026

Notebook PR Validation: PASS

Uh oh!

github-actions Bot commented Sep 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Golden-Set Execution (H.7 P3)

Uh oh!

github-actions Bot commented Sep 26, 2026

MD hierarchy drift -- 480d7f8

Uh oh!

github-actions Bot commented Sep 26, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Path-collision (organ #13359/#13615)

Uh oh!

clusterManager-Myia left a comment

Choose a reason for hiding this comment

Uh oh!

jsboige commented Sep 26, 2026

Uh oh!

jsboige commented Sep 26, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

MD hierarchy drift -- `bdd5a1b`

github-actions Bot commented Sep 26, 2026 •

edited

Loading

Notebook outputs-required (H.4 schema): PASS (every code cell carries an `outputs: list`)

github-actions Bot commented Sep 26, 2026 •

edited

Loading

MD hierarchy drift -- `480d7f8`

github-actions Bot commented Sep 26, 2026 •

edited

Loading