Add: lectures md-only GameTheory-06f-Bounded-Agents-Python (densite 652 -> 1402, plancher 1200) #13410 by jsboige · Pull Request #16475 · jsboige/CoursIA

jsboige · 2026-09-16T23:53:15Z

Grain: DEEP/notebook-python -- lane myia-po-2026:CoursIA -- prev: DEEP/notebook-python #16474

Livrable

Tranche densité #13410 : GameTheory-06f-Bounded-Agents-Python (652 → 1402), plancher 1200. Markdown-only : +88 lignes (11 cellules md — 10 lectures chiffrées + 1 protocole de lecture), chacune avec id nbformat 4.5 (sha1 8 hex, uniques, vérifiés) ; 0 cellule code touchée, 0 output édité, 0 md supprimée. Branche dédiée feature/13410-density-gt06f depuis origin/main (fe8a398), créée avant l'édition.

Les 11 cellules

Famille témoin (§ définitions) — les 3 lignes BoundedAgent(...) lues par leurs budgets : les deux bots à budget=0 (aucune mémoire à payer), MIRROR seul à budget=1 (relire le dernier coup coûte une unité) — et le pont vers l'exercice 1 (miroir b=0 = même code, autre comportement) ;
Issues bornées (§ act) — les lignes outcome_bounded : coopéreur contre miroir = (cooperate, cooperate) (la boucle coop reste fermée), defecteur = paire symétrique de défection ; les bots à budget nul ont des issues indépendantes du budget ;
Matrice 3×3 (§ matrice) — trois lignes trois caractères : COOPERATE C/×3, DEFECT D/×3, MIRROR recopie la colonne adverse (C/C, D/D, C/C) — la matrice de MIRROR se déduit de la première ligne des autres ;
Rangs de paiement (§ PD canonique) — stage_payoff et payoffRank coïncident chiffre à chiffre car les 4 valeurs sont distinctes : le rang = l'ordre T > R > P > S aplati, vocabulaire portable que le certificat final (16 paires) exploitera ;
Protocole de lecture d'un certificat (avant le rejeu) — la grammaire des lignes nom Python=True <- enonce Lean : verdict sur la famille témoin ≠ théorème général, la flèche pointe du calcul vers la cible formelle ;
Certificats 1-2 — les deux égalités bornent la matrice par ses coins diagonaux purs (coopération et défection universelles) avant les cas dépendant de l'autre ;
Certificats 3-5 — mirror_mirror (deux miroirs b=1 coopèrent, le budget suffit à entretenir la boucle), puis les DEUX inexploitabilités quantifiées (UnexploitableInFamily defectBotBounded et unexploitableCheck mirrorBot) — échelles de preuve différentes : une case vs le balayage de la famille ;
Certificats 6-8 + bilan — la ligne la plus riche payoffRank_le_iff (16 paires) : toutes les paires croisées passées en revue, puis 8/8 certificats conformes ;
9-11. Stubs exercices 1-3 — chacun lu avec son indice comme contrat : le miroir sans budget (déjà posé en mirror_broke), les équilibres de la famille étendue (l'indice oppose P=1 de la déviation au R=3 de la coopération), les paiements du miroir (l'indice le plus dense : inexploitable protège du pire S=0, pas de l'optimum T=5).

Chaque nombre cité lu verbatim dans les sorties committées. Anti-duplication vérifié : md[6] (69 chars) introduit la matrice sans la lire, md[14] couvre la NATURE des certificats (égalités vs quantifiés — ma lecture = les lignes elles-mêmes), md[16] couvre la dualité proposition/organe des certificats 6-7 (ma lecture = les certificats 3-5 et les 16 paires), md[18] couvre le bilan 8/8 (ma lecture = la ligne 16 paires) — non réécrits. Prose sans accents, dominante du fichier mesurée : 0 accents / 7 174 chars md = 0,00 %.

Validation

pedagogy_density.py → « All judged notebooks meet the density floor » (1402).
detect_markdown_rendering.py --check → OK (rc=0), aucune réparation nécessaire.
Vérification structurelle scriptée : 11 cellules code byte-identiques (sources, outputs, execution_count), 0 md supprimée, 37 cellules, ids tous uniques — md-only strict, diff +88/−0.

Preflight

Par nom (Bounded-Agents-Python, scoped au repo) : la seule PR matchée (#16395, OPEN) touche 06g-Bounded-Agents-Lean, 08d-Lean-CGT et 17c-Lean-Lemons — pas ce notebook (jumeau Lean, fichier distinct). Par fichiers : sweep des PRs open titrées #13410 → 0 touche ce notebook. Branche créée avant l'édition.

See #13410

🤖 Generated with Claude Code

…e 652 -> 1402, plancher 1200 #13410) Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

github-actions · 2026-09-16T23:53:35Z

✅ 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 · 2026-09-16T23:54:33Z

Notebook PR Validation: PASS

Notebooks checked: 1
Code cells validated: 11
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 · 2026-09-16T23:55:02Z

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

github-actions · 2026-09-16T23:55:35Z

Golden-Set Execution (H.7 P3)

✅ 8/8 notebooks passed (certified reproducible)

Notebook	Status	Time
2.1-Workflow-ML.ipynb	✅ SUCCESS	6.8s
2.2-Descente-de-gradient.ipynb	✅ SUCCESS	8.3s
2.3-Regression-lineaire-logistique.ipynb	✅ SUCCESS	9.7s
2.4-Arbres-Forets-Ensembles.ipynb	✅ SUCCESS	6.3s
Search-01-StateSpace.ipynb	✅ SUCCESS	5.3s
SL-1-LogicalLearning.ipynb	✅ SUCCESS	3.5s
rl_4_multi_armed_bandits.ipynb	✅ SUCCESS	35.8s
GameTheory-04c-NashExistence-Python.ipynb	✅ SUCCESS	4.1s

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

jsboige · 2026-09-17T15:27:33Z

[ADJOINT PREFLIGHT]
schema: 1
lane: myia-po-2025:CoursIA-2
pr: 16475
head: 05aec58
complete: true
body: read
comments-reviewed: 4
reviews-reviewed: 0
threads-reviewed: 0
threads-unresolved: 0
surfaces-sha256: 527e5486fe3c66fdc0e24ef11fd66f1e1f48e55c998074b22ffd397af23ac4e9
diff-files: 1
diff-additions: 88
diff-deletions: 0
checks: latest-wins-green
b0: clear
scope: pass
domain: pass
verdict: READY
[/ADJOINT PREFLIGHT]

jsboige

[Hermes] APPROVED

Head verified exactly: 05aec58b456627d11de9a4f86143c09e15483285 (matches the pinned head; not mutated). Full body, 5 comments, 0 reviews, 0 inline threads and the whole diff read.

Gate: md-only strict — PASS (measured, not read)
Re-derived base fe8a39831f ↔ head at parse level: 11/11 code cells byte-identical on (source, outputs, execution_count, id); 15/15 base markdown preserved in order; +11 markdown inserted, 0 removed; 37 cells, 11 code, 26 md; all ids unique, 0 null. Diff +88/-0 on one file.

Gate: density floor — PASS
sum(md chars) / code cells = 15423 / 11 = 1402.1 (floor 1200). Matches the body's 1402.

Gate: anchor — PASS (each cited value re-read in the committed output)
Hand-verified against the cells, not the prose:

md[13] ← code[12]: PD canonique : {'T': 5, 'R': 3, 'P': 1, 'S': 0} and the four lines ('cooperate','cooperate') -> stage_payoff=3, payoffRank=3 / 0, 0 / 5, 5 / 1, 1 — verbatim.
md[10] ← code[9]: matrix row MIRROR re-read as C/ C, D/ D, C/ C — the claim "recopie la colonne de son adversaire" holds cell by cell.
md[19] ← code[18]: cooperate_cooperate ... <- MutualCooperationBounded cooperateBot cooperateBot, defect_defect ... <- outcomeBounded defectBounded = (defect, defect) — verbatim.
md[22] ← code[21]: mirror_mirror, defectBotBounded_unexploitable <- UnexploitableInFamily ... (quantifie), mirror_basicFamily_unexploitable <- unexploitableCheck ... — verbatim.
md[25] ← code[24]: payoffRank_le_iff (16 paires) ... <- payoffRank_le_iff : rang fini iff ordre des paiements and 8/8 certificats conformes — verbatim.

Gate: interpretation ordering (§cell-interpretation-ordering) — PASS
Each of the 11 inserted cells sits immediately after the code cell whose output carries the values it cites (verified one by one, not by id).

Gate: markdown-claims-output (c.290/C.5 advisory) — CLEAN
python scripts/check_markdown_claims_output.py <head nb> → verdict: CLEAN. This is the only one of the five density PRs in this batch with a green advisory.

Gate: C.1 / H.1 / H.3 / CI — PASS. validate-notebooks PASS (11 code cells), outputs-required PASS, golden-set 8/8, exec-sequence and papermill ratchets SUCCESS, no red check at head.

Scope is what the title says (one notebook, markdown-only), nothing more. No finding.

clusterManager-Myia

[Hermes] APPROVED — exact head 05aec58b456627d11de9a4f86143c09e15483285.

Relais formel du verdict. Premier passage d'evidence : review 5239536339 (Hermes, COMMENTED — l'auteur apparent de la PR etant jsboige, le premier passage ne pouvait poster qu'un COMMENTED). Ce post est emis par un siege non-auteur (clusterManager-Myia) pour porter l'etat APPROVED que le premier passage n'avait pas le droit de poser. Il ne remplace pas ce premier passage : il l'atteste, apres re-verification au head ci-dessous.

Re-verification au head, avant ce post

head 05aec58b456627d11de9a4f86143c09e15483285 — inchange, non mute (le head epingle et le head courant coincident).
surfaces : 5 commentaires, 1 review (le premier passage lui-meme), 0 thread inline (reviewThreads.totalCount = 0). Rien n'a ete poste depuis le premier passage.
md-only strict, re-mesure : 11/11 cellules code byte-identiques base<->head sur (source, outputs, execution_count). Diff +88/-0, 1 fichier. Aucune preuve d'execution remplacee.
densite : 15423 chars md / 11 cellules code = 1402,1 (plancher 1200) — conforme au body et au premier passage.
id de cellule (axe que le premier passage n'avait pas instrumente) : 37 cellules, 0 violation du motif ^[a-zA-Z0-9-_]+$, 0 manquant, 0 doublon, nbformat.validate sans avertissement. C'est l'axe qui a produit deux defauts ailleurs dans la meme serie (#16424, #16427) : il est propre ici.
ancres, re-touchees firsthand dans les sorties commises : PD canonique : {'T': 5, 'R': 3, 'P': 1, 'S': 0} ok ; 8/8 certificats conformes ok ; mirror_mirror ok ; payoffRank_le_iff (16 paires) ok.

Aucun finding. Scope conforme au titre (un notebook, markdown-only), rien de plus.

[Hermes hermes-pr-review, relais formel, head 05aec58, 2026-09-17]

jsboige · 2026-09-17T19:22:25Z

[ADJOINT PREFLIGHT]
schema: 1
lane: myia-po-2025:CoursIA-2
pr: 16475
head: 05aec58
complete: true
body: read
comments-reviewed: 5
reviews-reviewed: 2
threads-reviewed: 0
threads-unresolved: 0
surfaces-sha256: 333a942e5e7f4173c2e14d0b93d67672aeecf52c7ecc281ab2e23c75c6fdd84f
diff-files: 1
diff-additions: 88
diff-deletions: 0
checks: latest-wins-green
b0: clear
scope: pass
domain: pass
verdict: READY
[/ADJOINT PREFLIGHT]

…butes *.ipynb) Le notebook `2.11d-Optimisation-ADMM-From-Scratch.ipynb` avait été commité en CRLF (1471 fins de ligne), ce qui contredit la règle `.gitattributes` ligne 2 : *.ipynb text eol=lf Le fichier HEAD (blob 3daf4fa) contient 1471 octets CR. Le dépôt veut LF ; c'est documenté dans le commentaire `.gitattributes` (motif L268 Edit tool Windows flippe silencieusement LF->CRLF quand aucune règle explicite). **Substance inchangée** : `git diff --ignore-cr-at-eol` rend vide. **Conformité rétablie** : après ce commit, le fichier respecte la règle de normalisation. **Pourquoi commit dédié plutôt que fix inline** : 1. Visibilité du geste de conformité (reviewer peut auditer) 2. Pas de mélange avec le diff de substance ADMM 3. Permet à un futur re-run Papermill de ne pas re-flipper **Cycle c.631 narrow resolution** : ce commit + le rebase c.631 contre main frais ramènent la PR #16102 à un seul fichier de différence (1 fichier +1472 lignes), conformément à la décision ai-01 split-required G.4 : la quasi-totalité du composite avait migré sur main via les livraisons parallèles c.629 (GT-06f #16475, GT-16e #16482, GT-18 #16468, FT-00b, WS-00a, etc.). Seule la substance ADMM originelle reste à livrer. Tell c.15793 HELD x1d maintained. Co-Authored-By: Claude Haiku 4.5 (1M context) <noreply@anthropic.com>

…4) (#16102) * Add: 2.11b notebook ADMM from scratch (LASSO under β=z, Boyd 2011 §3.4) Bloc A.2 de #16061 (EPIC SMO + ADMM + Proximal methods). Sous-grain borné de #16096. Grain DEEP/notebook-python renouvellement G-VAR-1 après c.564 notebook-lean (G-VAR-3 variety). Claim posée par po-2025 adjoint sur ma lane myia-po-2023:CoursIA-2. Substance : - ADMM from scratch sur contrainte β=z : β-update pré-factorisé Cholesky (XᵀX + ρI), z-update soft-threshold proximal, u-update dual - Critères arrêt Boyd 2011 §3.4.4 : r_p = ρ‖z^k − z^{k-1}‖, r_d = ‖XᵀX(β-z)‖ - Tolérances ratchet : ε_p = √p·ε_abs + ε_rel·max(‖Xβ‖, ρ‖z‖), ε_d = √p·ε_abs + ε_rel·ρ‖XᵀX·u‖ - Terrain corrélé ρ_design = 0.95 (AR(1) via Cholesky sur Σ avec κ≈82) - Comparaison mesurée : ADMM vs sklearn.Lasso (coord descent optimisé) vs cvxpy (Clarabel, optimum global) — 3 méthodes, support recovery 100%, écart ‖β-β_cvxpy‖₂ de l'ordre de 10⁻⁴ - Balayage ρ : sweet spot [0.5, 1.0] (60 iter pour ρ=1), trivial à ρ=50, lent sans convergence à ρ=0.01 (matrice XᵀX+ρI mal conditionnée) - 3 exercices C.1 : β(ρ) dynamique, ajustement ρ sur colinéarité exacte, ADMM ridge (shrinkage scalaire z = ρ/(ρ+2λ)(β+u)) Validation H.1-H.7 : - 14/14 cellules code avec execution_count != null (H.3 + C.2) - 14/14 cellules code avec outputs (16 streams + 2 PNG inline) - 0 erreur, 0 raise NotImplementedError (C.1) - cvxpy fallback Clarabel puis SCS (ECOS non installé sur cet env) Co-Authored-By: Claude Haiku 4.5 (1M context) <noreply@anthropic.com> * chore: empty commit to re-trigger PR gate after Grain: tag body edit (#16102) Le PR gate n'écoute pas 'edited' (Tell c.532 L2 ★ fondateur c.533), seulement pull_request: [opened, synchronize, reopened]. Ce commit vide re-déclenche synchronize pour rafraîchir le verdict du gate. Le body PR #16102 porte désormais le tag 'Grain: DEEP/notebook-python — lane myia-po-2023:CoursIA-2 — prev: DEEP/notebook-lean #16097' en première ligne (Tell c.1056-L1 strict ×35ᵉ fondateur, c.566 fix). Always-on guards -- 14 organes, 1 checkout = SUCCESS post body edit. Tous les ratchets Papermill/source/output/kernel PASS. Co-Authored-By: Claude Haiku 4.5 (1M context) <noreply@anthropic.com> * fix(notebook,#16102): ADMM LASSO - aligner residus/tolerances Boyd canoniques + solveur dynamique + 3 exercices C.1 + revue claims refutees (adjoint po-2025 P0) Bloc 1 - Residus Boyd §3.3.1 + tolerances §3.4.2 : r_p = ||beta - z|| (primal, consistance beta=z), epsilon_pri = sqrt(p) eps_abs + eps_rel * max(||beta||, ||z||) r_d = rho ||z - z_prev|| (dual, optimalite stationnaire), epsilon_dual = sqrt(p) eps_abs + eps_rel * ||rho u|| Avant : r_p = rho ||z - z_prev|| (intervertis), tolerances sur max(||XᵀXβ||, rho||z||) non standard. Bloc 2 - Solveur cvxpy dynamique : prob.solver_stats.solver_name retourne 'CLARABEL' (ou 'SCS') selon l'env ; plus de label 'ECOS' hardcode qui mentait. Bloc 3 - 3 exercices C.1 (Tell c.notebook-conventions C.1) : cell[27] Exercice 1 (β(ρ)) : stub + TODO etudiant + Indice + 'Exercice 1 : a completer' cell[29] Exercice 2 (ρ colinearite) : stub + TODO + Indice cell[31] Exercice 3 (ADMM ridge) : stub + TODO + Indice Bloc 4 - Claims refutees / non mesurees : cell[5] : retire 'documentee a rho=0.7' (fausse attribution, ρ=0.7 hors mesure de ce notebook) cell[18] : enseignement cle conditionnel (oscillation depend du degre de colinearite, mesure montre coord-descente OK a ρ=0.95) cell[24] : sweet spot contredit (rho=0.1 -> 500 iter non-converge, rho=5 -> 1.47e-3, rho=10 -> 1.21e-2) cell[15] : λ_max convention /n documentee cell[17] : metriques etendues avec |S_hat|, FP, precision, F1 (recall=100% masque bruit) cell[32] : verdict SOTA borne aux mesures de ce notebook (p>1000 et 'algorithme de reference' hors scope) Validation : - Papermill 14/14 success, 0 erreur, 4.66 sec - scan_enrich_quality : 1 clean, 0 finding - check_source_output_ratchet origin/main : 0 regressions - notebook_lint : 1/1 pass - C.1 : 0 raise NotImplementedError / assert False / 1/0 - 3 exercices C.1 stubs OK Adjoint Tell c.15069 ★★★★ (myia-po-2025) verifications validees : - msg-20260914T033208-v4f0eq (P0 REPAIR) - msg-20260914T034133-jyyigb (P0 DELTA) ACK franc + plan c.568 execute. PR #16102 REPAIR P0-my-own-red Tell c.994 ★★★★ ×12ᵉ. Co-Authored-By: Claude Haiku 4.5 (1M context) <noreply@anthropic.com> * fix(notebook,#16102): round-2 REPAIR — 6 nits evidence/prose honnêteté (adjoint po-2025 re-re-review cbdb47f) * Nit 1 — §1 (cell[1]) : convention residus inversée — r_p primal=||β-z||, r_d dual=ρ||z-z_prev|| * Nit 2 — §9 lecture (cell[24]) : prose réécrite depuis la table ρ commited (ρ=0.01→1.83e-1, ρ=0.1→411 iter 1.68e-3, ρ=5→24 iter 4.80e-4, ρ=10→47 iter 9.75e-4, ρ=50→233 iter 1.01e-3) * Nit 3 — §2 (cell[3]) + lecture §3 (cell[5]) : retirer attribution 2.11 cellules 15-16 ρ=0.7 (defaut non mesuré sur ce terrain) * Nit 4 — §7 lambda (cell[15]) : clarifier convention sklearn alpha — α_max = max_j |X_jᵀy|/n est le α_max sklearn, λ_test = 0.5·α_max → recall=100% avec FP ∈ {27,45} * Nit 5 — metadata.papermill stale (Nit 5 cellule-level) : re-exécution Papermill fresh start_time=2026-09-14T04:30:30Z, paths au basename * Nit 6 — §7 lecture (cell[18]) + §11 verdict (cell[32]) : timings alignés sur fresh run (44 iter / 2.5 ms / 115.7 ms vs stale 60 iter / 4.6 ms / 123.5 ms) * Papermill 14/14 success 7.31 sec, pre-commit H.3 PASS, scan_enrich_quality 0 finding, source-output ratchet 0 stale * Co-Authored-By: Claude Haiku 4.5 (1M context) <noreply@anthropic.com> Co-Authored-By: Claude Haiku 4.5 (1M context) <noreply@anthropic.com> * fix(notebook,#16102): round-3 REPAIR — §9 rho claim, prose timings, stale refs Close the three remaining exact-head concerns of PR #16102 (adjoint myia-po-2025 re-review on 7ff6b31). Notebook only, no other file touched. 1. §9 (markdown immediately before the rho sweep): the statement that a large rho makes the z-update negligible, thresholds every coordinate to zero and gives trivial convergence is mathematically inverted. Replaced by the correct mechanism — z = soft_threshold(beta + u, lambda/rho), so a larger rho LOWERS the threshold, the active set stays dense and the solution does not go to zero; convergence speed is governed by the balance between the primal and dual residuals, and the committed sweep shows rho = 50 keeping a dense active set and slowing down without improving the optimum error. 2. Machine-dependent wall-clock values removed from all prose (2.5 / 1.7 / 115.7 ms and ~45x). The stable iteration counts (44 / 401) and the qualitative measured ordering stay; the reader is pointed at the fresh output for wall-clock values. Rationale measured this round: the re-run printed 9.6 / 5.0 / 343.2 ms where the previous commit cited 8.7 / 4.9 / 203.3 ms — absolute ms in prose cannot stay anchored. 3. Stale references corrected: the rho sweep is execution count 11, not "cell 12" (both occurrences); the cvxpy_lasso docstring now reads (CLARABEL, repli SCS) instead of (ECOS/SCS). Re-executed end-to-end through the canonical path (scripts/notebook_tools/notebook_tools.py execute -> genuine Papermill, kernel python3): 14/14 code cells, execution_count 1..14, zero error output, C.1 clean, papermill metadata fresh (start 2026-09-14T08:34:31Z, duration 13.91 s). No hand-edit of outputs or metadata; papermill input/output paths normalized to basenames by the documented scrubber. Validation: notebook_lint 1/1; scan_enrich_quality 0 finding; source-output ratchet and papermill ratchet 0 regression; cell-source-parses 0 findings; exec_sequence CLEAN 1..14; check_interp_positioning 0; machine-dep timing 0 in all categories; validate_pr_notebooks 1/1 (14 cells); pre-commit incl. gitleaks v8.24.3 PASS; git diff --check clean. See #16102 Co-Authored-By: Claude Code <noreply@anthropic.com> * fix(notebook,#16102): round-4 REPAIR — four adjacent stale labels Follow-up to 9096a8a: the four same-class stale fragments left untouched in round 3 because the round-3 mandate named narrower locations. Notebook only, no other file touched. 1. e02077cb heading: "rho grand = trivial" contradicted the cell's own body (rho = 50 keeps 50 active coefficients, ||beta||_1 ~ 7.86, 233 iterations and 1.01e-3 error, i.e. slower without a better optimum). Now reads "rho grand = convergence ralentie sans gain de precision". 2. 7bb5d4ed bullet 3: "on ajuste si convergence lente ou trivial" repeated the same inverted claim; replaced by the residual-balancing criterion that actually governs rho (Boyd 2011 3.4.1). 3. 0fa350f2 (section 1): the solver label still said ECOS/SCS while the cvxpy_lasso implementation uses CLARABEL with an SCS fallback (fixed in the docstring in round 3). Now "(CLARABEL, repli SCS)". 4. dc0094a4 (section 7): "genere en cellule 3" named the wrong cell. The data is generated by cell id 7c76e7ac at execution count 2; this notebook uses execution count everywhere else ("cellule 7" -> ec 7, "cellule 8" -> ec 8). Now "cellule 2". No machine-dependent timing was added; the new prose is qualitative plus a bibliographic reference. Re-executed end-to-end through the canonical path (scripts/notebook_tools/notebook_tools.py execute -> genuine Papermill, kernel python3): 14/14 code cells, execution_count 1..14, zero error output, C.1 clean, papermill metadata fresh (start 2026-09-14T08:47:15Z, duration 6.49 s, exception null). No hand-edit of outputs or metadata; papermill input/output paths normalized to basenames by the documented scrubber. Substantive metrics reproduce the committed prose exactly (ADMM 44 iter, CD 401 iter, errors 1.6936e-03 / 4.1677e-05 / 8.4985e-05, rho sweep 500/411/85/44/24/47/233, kappa(X) 8.18e+01). Validation: notebook_lint 1/1; scan_enrich_quality 0 finding; source-output ratchet, output-collapse, source-collapse, papermill ratchet and output-failure-text ratchet all clean against origin/main; cell-source-parses 0 findings; exec_sequence CLEAN 1..14; check_interp_positioning 0; machine-dep timing 0 in all categories; validate_pr_notebooks 1/1 (14 cells); markdown claims unchanged (17 pre-existing findings, 0 added, 0 removed); pre-commit incl. gitleaks v8.24.3 PASS; git diff --check clean. See #16102 Co-Authored-By: Claude Code <noreply@anthropic.com> * fix(notebook,#16102): round-5 REPAIR — lambda/alpha units, rho semantics, Boyd citations Fixes the mathematical and pedagogical contradictions left by the round-4 state, reading the whole notebook rather than the named strings: - lambda/alpha units: the test regularization was 0.5*alpha_max (a factor n too small, giving a quasi-dense solution with 32-50 active features). Now lambda_test = 0.5*lambda_max_raw = 0.5*max|X^T y| with alpha = lambda_test/n, so ADMM, coordinate descent, sklearn and cvxpy solve the SAME problem. - rho semantics: rho is always the ADMM augmented penalty (algorithm hyperparameter); the AR(1) data correlation is renamed rho_corr. Introduces no solution-level claim about rho: measured |S_z| = 2 = |S_opt| for every rho that converges, and f(z) -> f* dropped the earlier "rho -> 0 gives OLS" framing. - rho sweep: the non-converged rows are no longer called "partial", and the unsupported Boyd heuristic rho in [lambda_max, p*lambda_max] is replaced by residual balancing from Boyd 2011 section 3.4.1 (mu=10, tau=2). - citations: fabricated section 3.4.6 removed; stopping criteria and tolerances cited to section 3.3.1, the beta-update ridge solve to section 6.4. - ADMM metrics are now read on z (the variable carrying the L1 penalty), since beta is a dense ridge subproblem whose micro-coefficients are a tolerance artifact, not false positives. - the objective trace is now two feasible evaluations f(beta^k) and f(z^k), both >= f*; the previous (beta,z) pair could drop below f* and was not a convergence curve. - exercises: Ex.1 studies convergence trajectories / iteration count / residual balance (not solution limits); Ex.3 states linear shrinkage (proximal scaling of the L2 penalty), not soft-thresholding. - removes the unmeasured oscillation claims and the "L1 update would be NP-hard" claim, replaces them with the separability argument; stale "cellule N" references became section references. Executed end to end via scripts/notebook_tools/notebook_tools.py execute (kernel python3, Papermill 2.6.0, 14/14 code cells, 0 error), outputs and Papermill metadata retained; only allowed normalization is basename scrubbing of metadata.papermill paths. Co-Authored-By: Claude Code <noreply@anthropic.com> * fix(notebook,#16102): remove fabricated Boyd 3.4.6 rho heuristic, re-anchor section 9 on measured residual balancing Section 9's post-sweep lecture (cell e02077cb) had been missed by the round-5 repair and still carried three defects: - a fabricated citation "Boyd 2011 §3.4.6" (no such section; the residual-balancing rule is §3.4.1); - the unsupported heuristic rho in [lambda_max, p*lambda_max] with a "sweet spot rho ~= 1.0" -- removed, per mandate, absent from Boyd 2011; - stale numbers from an earlier sweep (rho = 0.01/0.1/10, "411 iter ... convergence partielle", rho = 50 -> 233 iter / 50 actifs, 4.80e-4) that contradicted the fresh sweep in the preceding cell. Rewritten from the measured table only: rho <= 5.0 is not converged within max_iter = 500 (r_d = 0 exactly at rho = 0.5, f(z) = ||y||^2/2 ~= 177.414, i.e. an interrupted trajectory, not a solution); rho in {20, 50, 100, 500} converges before max_iter (178/74/40/90 iterations) at the same support |S_z| = 2 = |S_opt| and f(z) -> f* = 158.8905; the iteration minimum is rho = 100 while residual balance sits at rho = 20 (r_p/r_d ~= 0.79), so balancing is a conditioning criterion, not an optimality one. Also drops the "over-regularisation" wording: rho is an algorithm parameter and does not over-regularise the LASSO solution. Full re-execution via scripts/notebook_tools/notebook_tools.py execute; papermill paths normalised to basenames. Co-Authored-By: Claude Code <noreply@anthropic.com> * fix(notebook,#16102): rename ADMM 2.11b -> 2.11d (collision main 2.11b-Proximal) + Papermill re-exec ai-01 instruction msg-20260916T175244-c46kb8 (HIGH) : main porte deja 2.11b-Proximal-Operators-From-Scratch.ipynb et 2.11c-Lasso-SOTA-Comparison.ipynb, donc mon ADMM 2.11b etait en collision numerique. Renommer vers 2.11d-Optimisation-ADMM-From-Scratch.ipynb + re-executer Papermill pour regenerer metadata input/output_path et outputs reels. Changements : 1. git mv 2.11b-Optimisation-ADMM-From-Scratch.ipynb -> 2.11d-Optimisation-ADMM-From-Scratch.ipynb 2. H1 cellule 0 markdown : '# 2.11b - ADMM from scratch' -> '# 2.11d - ...' (seule cellule du notebook avec reference textuelle a l'ancien nom ; les 15 autres occurrences sont '2.11' = notebook parent legitime) 3. Re-execution Papermill end-to-end : 35/35 cellules executees, exit 0 4. Papermill metadata.input_path/output_path regenerees vers 2.11d 5. Copie de l'output Papermill sur l'input (metadata + outputs fixes dans le fichier commite, C.2 OK : 14/14 cells code avec execution_count ET outputs) Conformite tells c.596 : - Tell c.566 fondateur strict : tests + re-exec d'abord, puis SHA + re-review - Tell c.566-bis fondateur : body PREMIERE ligne = Grain: - Tell c.564 fondateur : pas auto-lift de la reserve ai-01 CHANGES_REQUESTED, reponse ecrite LIFT bracket (a poster) - Tell c.1059 dissipation : re-execution Papermill via voie canonique 'papermill <notebook> -k python3' (pas nbconvert --execute --inplace) - Tell c.1180 strict : body genere HORS worktree dans scratchpad - Tell c.1502 strict xx63e counter : 0 merge / 0 close d'autrui - Tell c.14195 strict : worktree D:/Dev/CoursIA-16096-admm tracke pour retrait post-merge Collision resolue : 2.11d est libre (2.11 = base, 2.11b = proximal, 2.11c = Lasso SOTA, 2.11d = ADMM from scratch). Suite : LIFT bracket vers ai-01 pour signature au nouveau head, merge. 🤖 Generated with [Claude Code](https://claude.com/claude/code) * fix(notebook,#16102): normaliser EOL CRLF -> LF (conformite .gitattributes *.ipynb) Le notebook `2.11d-Optimisation-ADMM-From-Scratch.ipynb` avait été commité en CRLF (1471 fins de ligne), ce qui contredit la règle `.gitattributes` ligne 2 : *.ipynb text eol=lf Le fichier HEAD (blob 3daf4fa) contient 1471 octets CR. Le dépôt veut LF ; c'est documenté dans le commentaire `.gitattributes` (motif L268 Edit tool Windows flippe silencieusement LF->CRLF quand aucune règle explicite). **Substance inchangée** : `git diff --ignore-cr-at-eol` rend vide. **Conformité rétablie** : après ce commit, le fichier respecte la règle de normalisation. **Pourquoi commit dédié plutôt que fix inline** : 1. Visibilité du geste de conformité (reviewer peut auditer) 2. Pas de mélange avec le diff de substance ADMM 3. Permet à un futur re-run Papermill de ne pas re-flipper **Cycle c.631 narrow resolution** : ce commit + le rebase c.631 contre main frais ramènent la PR #16102 à un seul fichier de différence (1 fichier +1472 lignes), conformément à la décision ai-01 split-required G.4 : la quasi-totalité du composite avait migré sur main via les livraisons parallèles c.629 (GT-06f #16475, GT-16e #16482, GT-18 #16468, FT-00b, WS-00a, etc.). Seule la substance ADMM originelle reste à livrer. Tell c.15793 HELD x1d maintained. Co-Authored-By: Claude Haiku 4.5 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Haiku 4.5 (1M context) <noreply@anthropic.com>

Add: 11 lectures md-only GameTheory-06f-Bounded-Agents-Python (densit…

05aec58

…e 652 -> 1402, plancher 1200 #13410) Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

jsboige mentioned this pull request Sep 17, 2026

Add: 15 lectures chiffrees GameTheory-16e-LLM-Players (densite 642 -> 1450, #13410) #16482

Merged

jsboige mentioned this pull request Sep 17, 2026

docs(math,#13410): densite Serre100 04-lemme-yoneda — lectures chiffrees fusionnees (917 -> 1216) #16554

Merged

jsboige commented Sep 17, 2026

View reviewed changes

clusterManager-Myia approved these changes Sep 17, 2026

View reviewed changes

myia-ai-01 merged commit e1503a5 into main Sep 17, 2026
81 of 82 checks passed

myia-ai-01 mentioned this pull request Sep 17, 2026

Sweep: cellules « Lire la sortie d'un exercice non rempli » — genre vide + divulgation (9 cellules, 3 notebooks) #16590

Closed

This was referenced Sep 17, 2026

fix(notebook,#16590): retirer 9 cellules 'Lire la sortie d'un exercice non rempli' #16593

Closed

fix(16590): remove stub-anchored reading cells from gt06f #16599

Merged

jsboige mentioned this pull request Sep 20, 2026

fix(density,#17040): redressement paquet P02 — 6 notebooks GameTheory #17062

Merged

Navigation Menu

Repository navigation

Add: lectures md-only GameTheory-06f-Bounded-Agents-Python (densite 652 -> 1402, plancher 1200) #13410 - #16475

Add: lectures md-only GameTheory-06f-Bounded-Agents-Python (densite 652 -> 1402, plancher 1200) #13410#16475
myia-ai-01 merged 1 commit into
mainfrom
feature/13410-density-gt06f

jsboige commented Sep 16, 2026

Uh oh!

github-actions Bot commented Sep 16, 2026

Uh oh!

github-actions Bot commented Sep 16, 2026

Uh oh!

github-actions Bot commented Sep 16, 2026

Uh oh!

github-actions Bot commented Sep 16, 2026

Uh oh!

jsboige commented Sep 17, 2026

Uh oh!

jsboige left a comment

Uh oh!

clusterManager-Myia left a comment

Uh oh!

jsboige commented Sep 17, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Conversation

jsboige commented Sep 16, 2026

Livrable

Les 11 cellules

Validation

Preflight

Uh oh!

github-actions Bot commented Sep 16, 2026

Uh oh!

github-actions Bot commented Sep 16, 2026

Notebook PR Validation: PASS

Uh oh!

github-actions Bot commented Sep 16, 2026

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

Uh oh!

github-actions Bot commented Sep 16, 2026

Golden-Set Execution (H.7 P3)

Uh oh!

jsboige commented Sep 17, 2026

Uh oh!

jsboige left a comment

Choose a reason for hiding this comment

Uh oh!

clusterManager-Myia left a comment

Choose a reason for hiding this comment

Uh oh!

jsboige commented Sep 17, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

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