docs(taxonomy): #600 link census reproducibility — independent re-run byte-identical - #606
Merged
Merged
Conversation
… byte-identical Re-run the full link_* langlinks census (741 Fallacies + 88 Virtues articles, probe `0 fallacies` then `0 virtues`) on master c20d5d2 — an independent confirmation of the §5.1 decision-grade figures measured 2026-06-25. Result: byte-identical. Fallacies 2739/4823 (57%), Virtues 180/322 (56%), combined ~2919, 0/829 errors, and the same link_en categorization (Fallacies 900/433/75, Virtues 185/9/29). Per-language rows match cell-for-cell. Adds a one-paragraph "Reproducibility — re-confirmed 2026-06-27" note under §5.1 stating the figures are stable and not a one-run artefact. The "named number replacing the 57% estimate" was already the measured value on master (dispatch was prepared 1 min before #600 merged §5.1 at 14:40); this PR's contribution is the second-run pin that makes it reproducibly-confirmed. Read-only. 0 write under Cards/. Dispatch ai-01 2026-06-27 tertiaire (#600 census link_* complet). Co-Authored-By: Claude-Code <noreply@anthropic.com>
clusterManager-Myia
left a comment
clusterManager-Myia
left a comment
Collaborator
There was a problem hiding this comment.
[NanoClaw]
docs(taxonomy): #600 link census reproducibility — independent re-run byte-identical relu (+2/-0, 1 fichier markdown, Argumentum base master). Ajout d'un paragraphe de confirmation de reproductibilité : un re-run indépendant (po-2024, master c20d5d2c) du census langlinks #600 retourne des figures byte-identical.
Vérification interne faite : les chiffres du nouveau paragraphe (Fallacies 2 739 / 4 823 = 57 % ; Virtues 180 / 322 = 56 % ; catégorisation Fallacies 900 Wikipédia · 433 non-Wikipédia · 75 vides ; Virtues 185 · 9 · 29) matchent les figures §5.1 pré-existantes du même doc (vérifié head : 2 739 / 4 823 / 900 / 433 / 75 / 180 / 322 / 185 présents dans les sections de mesure d'origine). Cohérence interne solide → la claim « byte-identical » est auto-cohérente. Le probe 192-link-coverage-langlinks-probe.py existe bien (7 555 bytes, read-only, no API key — MediaWiki public).
Note de transparence : je vérifie la cohérence interne + l'existence du script, pas un fresh re-run indépendant de ma main (le census MediaWiki 741+88 articles n'est pas reproductible dans ce tour). La claim de reproductibilité est crédible et bien documentée (auteur po-2024, master SHA, date, conditions de throttle).
Additif pur (0 suppression), release-freeze respecté (fichier docs/taxonomy/, pas Cards/), 0 concern sécurité. LGTM-with-artifact.
[COMMENT only — validation humaine]
jsboige
added a commit
that referenced
this pull request
Jun 29, 2026
…sal (#609) Grounds #141 in the CURRENT taxonomy state (issue text is ~2 years old; DatasetUpdater + translation PRs have moved the state considerably). Census finding (read-only script on ba8e4a6, reproducible like #600/#606): - Non-card nodes: Fallacies 1232, Virtues 110 (Scenarii/Rules out of scope). - TEXT enrichment (desc/example/title x 8 langs) = 100% on both datasets => original #141 scope item 3 (descriptions + examples + translations) is DONE. No text-enrichment gap to script. - The "GPT-4 enrichment script" = already adapted: modern DatasetUpdater (PR #210, OpenAI SDK v2.10.0, gpt-5.5). Not a rewrite. Reframes #141: the genuinely-open residue is the AIF cross-reference graph (crossLink_* 8 relationship cols + AIF_skos* 4 SKOS mappings). Schema is READY (columns exist), content ~0% on non-cards: - Fallacies: crossLink 0.2%, AIF_skos 0.5% (barely started) - Virtues: crossLink_Opposes + skosDirectRef/MappingType 100% via #498 pilots, but 7 OTHER crossLink verbs = 0% - ~16700 empty cross-ref cells where schema is waiting. Proposal: 4-stage method (ground -> gpt-5.5 candidate via /v1/responses effort=low -> drift-free sidecar dry-run -> expert/jsboige ratification gate -> OWL/2sxc export). NOT auto-write: AIF/Walton mappings are specialised (memory: "Anti-Fab Validator: Walton scheme = WARN"), and Cards/ is frozen pre-tag. "Adapt the script" = add a crossLink prompt + task config to DatasetUpdater, GATED post-release on jsboige GO + schema decision (single decimal_path vs structured {target,note}). Overlaps #498 (AIF scale-up pilots produced the few filled cells). Feeds dependencies #141 declares: #130 (OWL/SKOS) + #136 (2sxc). Scope: docs + read-only script only. 0 write Cards/, 0 AssetConverter code change (pre-tag safe). Base ba8e4a6. Relates to #130, #136, #498. Co-authored-by: Your <your.email@example.com> Co-authored-by: Claude-Code <noreply@anthropic.com>
jsboige
added a commit
that referenced
this pull request
Jun 29, 2026
…or, 0 Cards/) (#610) * docs(taxonomy): #141 non-card census + AIF cross-ref enrichment proposal Grounds #141 in the CURRENT taxonomy state (issue text is ~2 years old; DatasetUpdater + translation PRs have moved the state considerably). Census finding (read-only script on ba8e4a6, reproducible like #600/#606): - Non-card nodes: Fallacies 1232, Virtues 110 (Scenarii/Rules out of scope). - TEXT enrichment (desc/example/title x 8 langs) = 100% on both datasets => original #141 scope item 3 (descriptions + examples + translations) is DONE. No text-enrichment gap to script. - The "GPT-4 enrichment script" = already adapted: modern DatasetUpdater (PR #210, OpenAI SDK v2.10.0, gpt-5.5). Not a rewrite. Reframes #141: the genuinely-open residue is the AIF cross-reference graph (crossLink_* 8 relationship cols + AIF_skos* 4 SKOS mappings). Schema is READY (columns exist), content ~0% on non-cards: - Fallacies: crossLink 0.2%, AIF_skos 0.5% (barely started) - Virtues: crossLink_Opposes + skosDirectRef/MappingType 100% via #498 pilots, but 7 OTHER crossLink verbs = 0% - ~16700 empty cross-ref cells where schema is waiting. Proposal: 4-stage method (ground -> gpt-5.5 candidate via /v1/responses effort=low -> drift-free sidecar dry-run -> expert/jsboige ratification gate -> OWL/2sxc export). NOT auto-write: AIF/Walton mappings are specialised (memory: "Anti-Fab Validator: Walton scheme = WARN"), and Cards/ is frozen pre-tag. "Adapt the script" = add a crossLink prompt + task config to DatasetUpdater, GATED post-release on jsboige GO + schema decision (single decimal_path vs structured {target,note}). Overlaps #498 (AIF scale-up pilots produced the few filled cells). Feeds dependencies #141 declares: #130 (OWL/SKOS) + #136 (2sxc). Scope: docs + read-only script only. 0 write Cards/, 0 AssetConverter code change (pre-tag safe). Base ba8e4a6. Relates to #130, #136, #498. Co-Authored-By: Claude-Code <noreply@anthropic.com> * tools(i18n): link_* langlinks resolver (sidecar candidate-URL generator, 0 Cards/) SECONDAIRE of ai-01 deep-queue dispatch. Extends the #600/#606 coverage track from MEASUREMENT to RESOLUTION. The existing probe (docs/taxonomy/192-link-coverage-langlinks-probe.py) measures the resolvable ceiling (~2919 cells, 57%, #600 §5.1) but discards the target title (returns only lang codes). This tool captures the target-language title via the MediaWiki langlinks API and emits the candidate fill URLs as a sidecar report — step 1 of #600 §6 methodology. For every node with an en.wikipedia.org/wiki/<Title> link_en missing link_<lang>, queries langlinks, captures the target title, emits: dataset,key,link_lang,resolved_url to stdout or --out <path>. NEVER writes under Cards/. Verified on a strided sample of 10 fallacies: 0 errors, 37 candidate fills (~57% rate, consistent with the probe ceiling), URLs correctly URL-encoded (Cyrillic %D0, PT accented %C3). Sample also surfaced the homonym risk #600 §6.4 warns about (English "Engagement" for a fallacy node) -> confirms human spot-validation is non-optional for AR/FA/ZH. Safety: - 0 write Cards/ (sidecar only, pre-tag freeze). - Public MediaWiki API, no key, 0.3s throttle, descriptive UA (default urllib UA is 403-forbidden). - Resolves FROM link_en (Wikipedia only); 433 non-Wikipedia curated sources excluded + preserved as-is (#600 §6.2). Next (post-release, gated): follow-up PR consumes the sidecar -> apply candidate URLs cell-by-cell (drift-free QUOTE_MINIMAL+CRLF, method #595), skip non-empty cells, human spot-validate ~5% residue (AR/FA/ZH, ~150). Relates to #600, #606. Co-Authored-By: Claude-Code <noreply@anthropic.com> --------- Co-authored-by: Your <your.email@example.com> Co-authored-by: Claude-Code <noreply@anthropic.com>
jsboige
added a commit
that referenced
this pull request
Jul 1, 2026
… (0 errors, 0 Cards/) (#618) PRIMAIRE of ai-01 deep-queue #2 (dispatch msg-...92kynx). Produces the candidate-URL output behind the #600 §5.1 probe ceiling. Ran tools/link-langlinks-resolve.py (PR #610, merged) over full Fallacies + Virtues taxonomy: 829 articles probed -> 2934 candidate link_<lang> URLs (fallacies 2754, virtues 180), 0 errors. Materializes the ~57% probe ceiling (~2919 cells) into concrete reviewable URLs. Step 1 of #600 §6 methodology. Per-lang: ru 371, pt 430, es 513, ar 603, fa 480, zh 537. Resolves FROM link_en (Wikipedia only); 433 non-Wikipedia curated sources excluded + preserved (#600 §6.2). Quality verified: - URL encoding correct per script (Cyrillic %D0, Arabic %D8, Persian %D8, CJK %E8, PT/ES Latin unencoded). - Homonym scan (#600 §6.4) of 1620 AR/FA/ZH candidates: 0 real leaks. AR 0/603, FA 0/480 Latin-path; ZH 3/537 (0.6%) = legit loanwords/ acronyms (FUD x2, Creepypasta) — not errors. The "Engagement" homonym from PR #610 sample does NOT recur at scale (isolated case). - Asymmetry: virtues ru/pt = 4 each (sparse abstract-concept coverage) vs ar/fa/zh 46-56 — coverage reality, flagged for spot-validation. Sidecars (reviewable snapshot, UTF-8 no-BOM LF): - docs/taxonomy/600-link-resolve-fallacies.csv (2754 rows) - docs/taxonomy/600-link-resolve-virtues.csv (180 rows) schema: dataset,key,link_lang,resolved_url Robustness: try/except + continue (skip on network error, never abort) -> 0 errors/829. Public MediaWiki API, no key, 0.3s throttle, descriptive UA. Next (gated post-release): apply PR consumes sidecar -> cell-by-cell drift-free (QUOTE_MINIMAL+CRLF, method #595), skip non-empty, human spot-validate ~5% AR/FA/ZH residue (~150 cells), re-run probe to confirm gap closed. Scope: docs/taxonomy/ only. 0 write Cards/, 0 AssetConverter code change (pre-tag safe). Base 18b4d02. Relates to #600, #606, #610. Co-authored-by: Your <your.email@example.com> Co-authored-by: Claude-Code <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
An independent reproducibility re-run of the
link_*langlinks census (#600, merged32dd809b), confirming the §5.1 decision-grade figures are stable — not a one-run artefact or mislabeled sample.Re-run (2026-06-27, master
c20d5d2c)Full probe, both datasets (
0 fallaciesthen0 virtues), 741 + 88 = 829 unique articles, ~0.3s throttle, 0 API key.link_encategorization also matches: Fallacies 900 Wikipedia · 433 non-Wikipedia · 75 empty; Virtues 185 · 9 · 29. Per-language rows match cell-for-cell (ru/pt/es/ar/fa/zh).What changed
One paragraph added under §5.1 — a "Reproducibility — re-confirmed 2026-06-27" note stating the figures are stable and decision-grade. No table/code changes; the data is the data.
Why this matters
The 2919 figure feeds jsboige's WE priority decision (how bounded the post-release link-fill effort is). A single-run measurement is a claim; a re-run that returns identical numbers is a reproducibly-confirmed claim — the same epistemic bar the characterization test (PR #604) sets for the Virtues FR-frozen behaviour.
Scope
Cards/(read-only docs, release freeze respected)c20d5d2c, 1 file, +2 linesRelates to #600. Does not change its conclusion — confirms it.