fix(gametheory,#14926): le carnet GameTheory-03d passe a la passerelle claudish by jsboige · Pull Request #19634 · jsboige/CoursIA

jsboige · 2026-10-07T00:51:33Z

Grain: MED/notebook-python — lane myia-po-2023:CoursIA — prev: MED/docs #19626

Objet

Tranche de la migration des consommateurs runtime vers la passerelle claudish
(acceptance 3-4 de #14926) : le carnet
MyIA.AI.Notebooks/GameTheory/GameTheory-03d-Le-Joueur-LLM-Python.ipynb cesse de porter
la cle du service et s'authentifie aupres de la passerelle.

C'est le premier carnet du perimetre dont l'appartenance n'est pas discutee : il ne fait
pas d'image generation vLLM directe, contrairement a e2e_quant_validation.ipynb, sur
lequel porte l'arbitrage user demande par la lane precedente.

Ce qui change

avant	apres
`VLLM_ENDPOINT` (defaut `http://192.168.0.47:5002`)	`CLAUDISH_BASE_URL` (defaut `http://localhost:3000`)
`VLLM_API_KEY`	`CLAUDISH_PROXY_KEY`
`VLLM_MODEL`	`CLAUDISH_MODEL` (`qwen3.6-35b-a3b`, servi par la passerelle)
en-tete `Authorization: Bearer <cle de service>`	en-tete `x-api-key: <cle cliente claudish>`
`max_tokens: 1`	`CLAUDISH_MAX_TOKENS` (defaut `2048`)

Le stub C.1 est conserve : sans CLAUDISH_PROXY_KEY, le carnet reste executable et
n'appelle rien.

Le budget n'est pas un choix esthetique, il est mesure

La passerelle ne propage pas chat_template_kwargs : enable_thinking=False n'a plus
d'effet et le raisonnement consomme le budget avant que la reponse n'atteigne content.
Mesure firsthand du 2026-10-07 sur localhost:3000, modele qwen3.6-35b-a3b, temperature 0 :

`max_tokens`	`content`	`reasoning_content`	`finish_reason`
1	`None`	3	`length`
64	`None`	262	`length`
512	reponse atteinte, mais 4 rounds sur 5 (cassette 8, rejeu non identique)	263	`stop`
2048	reponse atteinte, 5 rounds sur 5 (cassette 10, rejeu identique)	263	`stop`

Le meme appel servi par glm-5.2 se comporte comme qwen3.6-35b-a3b : c'est la passerelle,
pas le modele. Temoin negatif : l'endpoint direct rend 401 avec la cle locale — la
rotation est passee, la voie directe est morte sur cette machine.

Avec le budget d'origine, le carnet basculerait sur son stub a chaque appel : une sortie
degradee consacree. Le defaut est donc dimensionne pour laisser le raisonnement finir.

Verifications

Re-execution reelle : Papermill 2.7.0, kernel python3, 26 cellules, 0 erreur,
execution_count renseigne sur les 11 cellules code.
Propriete du carnet restauree : Cassette apres appel : 10 entree(s) et
Rejeu cassette : trajectoire identique ? True — identique a la sortie d'origine sur main.
Sortie non degradee : Trajectoire : CD CD CD CD CD (5 rounds joues), et non le
message de stub.
Ce qui change, et qui est assume : la trajectoire passe de CC CC CC CC CC (100 % de
cooperation) a CD CD CD CD CD. La cause est identifiee — le raisonnement n'est plus
desactivable a travers la passerelle. Aucune prose du carnet n'asserte ce resultat : la
cellule amont ne fixe que l'objectif (remplacer le stub par un appel reel).
Aucun secret dans le diff : la valeur n'apparait nulle part.

Portee mesuree du reste

Mesure par parse des sources de cellules (le grep de chaine sous-compte : les guillemets
sont echappes dans un .ipynb) : 9 autres carnets lisent encore VLLM_API_KEY, dont
quatre absents de la liste du body d'origine (02-Aspire-GenAiStack-Reel,
10b_Inference_Mechanics, 16_Scaling_Test_Time_Compute, 22b_Profil_Cognitif_CHC).
La reclassification en umbrella demandee par la lane precedente reste donc fondee.

See #14926

🤖 Generated with Claude Code

…e claudish Le carnet ne porte plus la cle du service vLLM : il consomme CLAUDISH_BASE_URL et presente l'en-tete claudish x-api-key (acceptance 3-4 de #14926). Le budget de generation est mesure, pas choisi : la passerelle ne propage pas chat_template_kwargs, donc enable_thinking=False n'a plus d'effet et le raisonnement consomme le budget avant que la reponse n'atteigne content. A 1 token le carnet basculerait sur son stub C.1 (sortie degradee consacree) ; a 2048 les 11 cellules code s'executent, 5 rounds sont joues et le rejeu de cassette est identique sans appel reseau. Re-execution reelle : Papermill 2.7.0, 26 cellules, 0 erreur. See #14926 Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

github-actions · 2026-10-07T01:17:33Z

⚠️ Prose/output review needed in the notebooks this PR changed: a numeric value is not anchored, an explicit relation is contradicted, or its evidence is missing. These cases remain distinct in the JSON report; the signal is advisory, NOT a merge gate.

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 · 2026-10-07T01:17:37Z

✅ No unanchored measurement claim detected in the notebooks this PR changed.

Scope = notebooks CHANGED in this PR, not the whole corpus. The stale-claim-report run artifact holds the structured JSON.
Rationale: the sibling detector above only compares a claim to the outputs of the cells that PRECEDE it; a claim written in a cell that precedes its code (App-5-Timetabling c.2/c.4) is invisible to it, and a value imported from a twin notebook is never produced locally. See python scripts/check_stale_claims.py --help.

github-actions · 2026-10-07T01:17:40Z

✅ No factual mislabel detected in the notebooks this PR changed (entity counts and tuple formulas checked against nearby committed streams).

Scope = notebooks CHANGED in this PR, not the whole corpus. The factual-mislabel-report run artifact holds the structured JSON.
Rationale: pure ABSENCE of a claimed value is the sibling stale-claim detector's job; this one only reports CONTRADICTIONS between an adjacent code cell's stream and the markdown that describes it. See python scripts/check_factual_mislabel.py --help.

github-actions · 2026-10-07T01:18:32Z

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 · 2026-10-07T01:27:56Z

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

github-actions · 2026-10-07T01:32:16Z

Golden-Set Execution (H.7 P3)

✅ 9/9 notebooks passed (certified reproducible)

Notebook	Status	Time
2.1-Workflow-ML.ipynb	✅ SUCCESS	5.6s
2.2-Descente-de-gradient.ipynb	✅ SUCCESS	2.8s
2.3-Regression-lineaire-logistique.ipynb	✅ SUCCESS	6.0s
2.4-Arbres-Forets-Ensembles.ipynb	✅ SUCCESS	5.8s
Search-01-StateSpace.ipynb	✅ SUCCESS	4.6s
SL-1-LogicalLearning.ipynb	✅ SUCCESS	3.5s
RL-04-Bandits-Manchots-Python.ipynb	✅ SUCCESS	17.0s
GameTheory-04c-NashExistence-Python.ipynb	✅ SUCCESS	4.0s
GameTheory-13d-Optimistic-CFR-Python.ipynb	✅ SUCCESS	12.0s

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

github-actions · 2026-10-07T01:48:33Z

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)

…ulee sans compteur quantitatif en prose (cellule markdown seule, pas de re-execution requise) L'organe check_prose_quantitative_claims flaggait `(alignee sur `GameTheory-3 cellule 5`)` : le motif <nombre> cellule matchait l'adjacence nom-de-serie + numero de cellule. La reference est reformulee (alignee sur la cellule 5 de GameTheory-3) : meme predicat, plus de forme compteur. Markdown-only, C.2 sans objet. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>

jsboige · 2026-10-07T06:10:06Z

[ADJOINT PREFLIGHT]
schema: 1
lane: myia-po-2026:CoursIA-3
pr: 19634
head: 03ca727
complete: true
body: read
comments-reviewed: 7
reviews-reviewed: 0
threads-reviewed: 0
threads-unresolved: 0
surfaces-sha256: 5076a3403896fa79e4df3f8f4b8a6946a500bc2d0d3c9ce7c0d88a660e970f58
diff-files: 1
diff-additions: 7
diff-deletions: 7
checks: latest-wins-green
b0: clear
scope: pass
domain: pass
verdict: READY
organ: check_adjoint_prevalidation.py
organ-command: python scripts/check_adjoint_prevalidation.py --derive-verdict 19634
organ-rc: 0
[/ADJOINT PREFLIGHT]

jsboige · 2026-10-07T06:10:40Z

[ADJOINT PREFLIGHT]
schema: 1
lane: myia-po-2026:CoursIA-3
pr: 19634
head: 03ca727
complete: true
body: read
comments-reviewed: 7
reviews-reviewed: 0
threads-reviewed: 0
threads-unresolved: 0
surfaces-sha256: 5076a3403896fa79e4df3f8f4b8a6946a500bc2d0d3c9ce7c0d88a660e970f58
diff-files: 1
diff-additions: 7
diff-deletions: 7
checks: latest-wins-green
b0: clear
scope: pass
domain: pass
verdict: READY
organ: check_adjoint_prevalidation.py
organ-command: python scripts/check_adjoint_prevalidation.py --derive-verdict 19634
organ-rc: 0
[/ADJOINT PREFLIGHT]

jsboige · 2026-10-07T06:11:07Z

[ADJOINT PREFLIGHT]
schema: 1
lane: myia-po-2026:CoursIA-3
pr: 19634
head: 03ca727
complete: true
body: read
comments-reviewed: 8
reviews-reviewed: 0
threads-reviewed: 0
threads-unresolved: 0
surfaces-sha256: 6383b3038f4105cd013633f871331b4366e0c92d4c8eea051d20aac18eae3981
diff-files: 1
diff-additions: 373
diff-deletions: 209
checks: latest-wins-green
b0: clear
scope: pass
domain: pass
verdict: READY
organ: check_adjoint_prevalidation.py
organ-command: python scripts/check_adjoint_prevalidation.py --derive-verdict 19634
organ-rc: 0
[/ADJOINT PREFLIGHT]

jsboige · 2026-10-07T06:11:38Z

[ADJOINT PREFLIGHT]
schema: 1
lane: myia-po-2026:CoursIA-3
pr: 19634
head: 03ca727
complete: true
body: read
comments-reviewed: 10
reviews-reviewed: 0
threads-reviewed: 0
threads-unresolved: 0
surfaces-sha256: d995220ab25da4cbd5f1b93bcfed6aed3cceb50efbc818ac118fac92f1988131
diff-files: 1
diff-additions: 373
diff-deletions: 209
checks: latest-wins-green
b0: clear
scope: pass
domain: pass
verdict: READY
organ: check_adjoint_prevalidation.py
organ-command: python scripts/check_adjoint_prevalidation.py --derive-verdict 19634
organ-rc: 0
[/ADJOINT PREFLIGHT]

…_reds (#19652) Le picker imputait a la base les rouges d'organes qui ne lisent QUE les lignes ajoutees par la PR (cas fondateur prose-counts-guard, aussi readme-ipynb-links-guard, pip-leak-guard, solution-leak-guard et les gardes delta). L'imputation disait aux 4 lanes bloquees (po-2023, po-2025, po-2026, ai-01 -- PRs #19614, #19625, #19626, #19634) « pas le votre », donc de ne rien faire, alors que la cause est forcement dans le diff. Source de verite : fast_lane_registry.Guard.argv (drapeau --diff {base_ref}...HEAD) et Guard.delta_argv (comparaison explicite base/head). Le helper _diff_bounded_check_names() lit les tranches du registre sans recopier une liste de noms ; un nouvel organe borne au diff declare dans la voie rapide est exclu automatiquement, sans patch picker. Le splitter split_base_corroboration recoit l'ensemble diff_bounded et place les cles correspondantes dans undecided SEUL (jamais dans base) : la mesure « check absent du rollup de main » n'est pas un « je ne sais pas » mais un « il ne peut pas venir de la base ». Test : 2 lanes, prose-counts rouge, findings differents -> aucune imputation a la base (le ticket fondateur). Scripts Tests (CPU) rouge sur main reste impute a la base (regression guard). Cas agregateur PR gate :: prose-counts-guard beneficie du filtre via le suffixe :: organ. None (defaut) preserve le comportement d'avant #19645. Co-authored-by: ai-01 <ai-01@anthropic.com>

jsboige mentioned this pull request Oct 7, 2026

claudish: la cle cliente du proxy n'a aucun organe de propagation — CLAUDISH_PROXY_KEY absent de SECRET_KEYS #14926

Open

github-actions Bot added the consecutive-code-cells Modified notebook has >=2 consecutive code cells (#12797) label Oct 7, 2026

jsboige mentioned this pull request Oct 7, 2026

feat(lean,#18205): enveloppe differential_lean -- 3 fermetures + #print axioms #19636

Merged

13 tasks done

This was referenced Oct 7, 2026

Fix(ml,#17387): Lab5 — nav/H1 au canon, exercices avant Conclusion, lecture ancrée, accents #19625

Merged

fix(picker): un organe borne au diff (prose-counts) n'est jamais un rouge herite de la base #19645

Open

jsboige mentioned this pull request Oct 7, 2026

fix(picker,#19645): exclure les organes bornes au diff de impute_base_reds #19652

Merged

myia-ai-01 merged commit 842bfce into main Oct 7, 2026
96 of 97 checks passed

jsboige deleted the feature/14926-claudish-consumers branch October 7, 2026 07:44

Navigation Menu

Repository navigation

fix(gametheory,#14926): le carnet GameTheory-03d passe a la passerelle claudish - #19634

fix(gametheory,#14926): le carnet GameTheory-03d passe a la passerelle claudish#19634
myia-ai-01 merged 2 commits into
mainfrom
feature/14926-claudish-consumers

jsboige commented Oct 7, 2026

Uh oh!

github-actions Bot commented Oct 7, 2026

Uh oh!

github-actions Bot commented Oct 7, 2026

Uh oh!

github-actions Bot commented Oct 7, 2026

Uh oh!

github-actions Bot commented Oct 7, 2026

Uh oh!

github-actions Bot commented Oct 7, 2026

Uh oh!

github-actions Bot commented Oct 7, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Oct 7, 2026

Uh oh!

jsboige commented Oct 7, 2026

Uh oh!

jsboige commented Oct 7, 2026

Uh oh!

jsboige commented Oct 7, 2026

Uh oh!

jsboige commented Oct 7, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

jsboige commented Oct 7, 2026

Objet

Ce qui change

Le budget n'est pas un choix esthetique, il est mesure

Verifications

Portee mesuree du reste

Uh oh!

github-actions Bot commented Oct 7, 2026

Uh oh!

github-actions Bot commented Oct 7, 2026

Uh oh!

github-actions Bot commented Oct 7, 2026

Uh oh!

github-actions Bot commented Oct 7, 2026

Uh oh!

github-actions Bot commented Oct 7, 2026

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

Uh oh!

github-actions Bot commented Oct 7, 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 Oct 7, 2026

Notebook PR Validation: PASS

Uh oh!

jsboige commented Oct 7, 2026

Uh oh!

jsboige commented Oct 7, 2026

Uh oh!

jsboige commented Oct 7, 2026

Uh oh!

jsboige commented Oct 7, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

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

github-actions Bot commented Oct 7, 2026 •

edited

Loading