To do

Admin page, collected at build time from the page sources. Nothing listed here is drawn on a page a reader sees, and an ordinary <Todo note="..." /> is not drawn on the page at all: it lives here, and in the file, for whoever works on that chapter next.

Inline to-do markers (17)

★ = urgent (<Todo urgent note="..." />), which sorts first, is shown on the page itself in the admin profile, and stops the chapter from being approved until it is dealt with. The rest are invisible on the page.

1. capabilities/multimodality/overview Pick up image and music generation in the products when this chapter is written (Ayla, 16 September 2026): ChatGPT offers Create image and Gemini offers Create image and Create music in the menu of their message box, shown on using-llms/product-features/documents-images-voice. The Claude deep research report of 16 September 2026 also claims video generation in ChatGPT (Sora) and Gemini (Veo) and none in Claude; verify at the providers' help pages before using any of it. Checked 16 September 2026: Sora is NOT current, OpenAI discontinued the Sora web and app experiences on 26 April 2026 (help.openai.com/en/articles/20001152); Claude does not generate photos or illustrations (support.claude.com/en/articles/9002504); Gemini's video model name was not settled (its help page says Gemini Omni, a search snippet said Veo 3.1)
2. classical-nlp/term-extraction/termostat A practical 'Running TermoStat' walkthrough for the TRA2600 TP is still to decide (in this chapter or in the gated TP chapter). The twelve captures that would serve it (sign-in with the redesign notice, account creation, empty corpora page, language menu, default settings, analysis running, collocates, semantic neighbours, word cloud, CRC, user guide, download buttons) are in ~/Desktop/termostat_captures_unused/, outside the repo.
3. course-materials/tra6008-project-briefs/projet-1 Le retrait du consentement passe par Noah (Ayla, 29 August 2026). Aucune adresse n'est publiée ici : à décider s'il faut nommer la marche à suivre, par exemple répondre au courriel de Noah.
4. course-materials/tra6008-project-briefs/projet-1 Formulation du certificat validée par Ayla le 28 August 2026. À relire si le dossier ou le financement change.
5. course-materials/tra6008-projet2/pas-a-pas La requête finale, ajoutée ici après la séance 7.
6. course-materials/tra6008-projet2/pas-a-pas La requête finale, ajoutée ici après la séance 8.
7. course-materials/tra6008-projet2/pas-a-pas Option non retenue pour l'instant : calculer aussi les scores contre chacune des trois traductions à la main de l'équipe qui a fait la référence, pour voir le score changer avec la traduction de référence.
8. critical-issues/copyright-and-authorship/detecting-generated-text The humanizer screenshot comes from the TRA6600 deck of 2026 (slide 175): the name of the service and the capture date are not recorded. Add both to the caption and the date to the file name (YYMMDD_), or recapture it.
9. critical-issues/copyright-and-authorship/fair-use-and-fair-dealing Canada's Supreme Court has said research is to be read broadly and is not limited to non-commercial settings (CCH Canadian Ltd. v. Law Society of Upper Canada, 2004 SCC 13), which weighs on whether training could count as research. Optional source to save if the nuance is wanted; ISED's 'What we heard' report (2025) on the consultation reached no decision on changes.
10. critical-issues/hallucinations/mitigating-hallucinations Ayla asked (22 September 2026) whether to name the criteria that make a chatbot search the web without being told to. Not stated on Anthropic's or OpenAI's help pages as far as could be checked that day (support.claude.com returned 404 on the internet-access article; help.openai.com needs a browser). The page therefore says only that both searched unasked, which the screenshots show. Worth adding once a vendor page states the triggers.
11. critical-issues/information-ecosystem/ai-slop The bookshop screenshot comes from the TRA6600 deck of 2026 (slide 169) and its capture date is not recorded: add the date to the file name (YYMMDD_) and to the caption if you know it, or recapture it.
12. critical-issues/information-ecosystem/generated-content-online Nothing peer-reviewed was found on the share of generated video on platforms such as YouTube, only a detection vendor's report (Kapwing, 2025-12), or on the share of generated product reviews. The page leaves both out rather than citing them.
13. critical-issues/information-ecosystem/the-open-web The licensing bullet used to say the deals were worth hundreds of millions of dollars, citing Axios on the News Corp and OpenAI agreement (22 May 2024). The Axios URL now returns 404 and News Corp's site is blocked in the browser pane, so the figure could not be checked; it was removed. A source for a named licensing deal, if wanted, is to be found and verified. Checked 18 September 2026: the often-quoted value ('could be worth more than $250 million over five years') comes from The Wall Street Journal, which News Corp owns, as repeated by the press (Spectrum News, Storyboard18); the companies did not disclose terms. Variety's article now sits behind TollBit, a pay-per-crawl gate, even in a browser, which is itself an example for the paragraph on payment per crawl. No primary account of the value could be read, so the figure stays out.
14. critical-issues/privacy/privacy-law Rome annulment: the Garante's own page now documents the judgment; the competence reasoning comes from the press. Update garante-2024's checked field, which says the page does not document the annulment. Whether the Garante appealed is not reported. Also decide whether GDPR, PIPEDA and Law 25 need entries of their own.
15. critical-issues/privacy/what-a-model-gives-back A glossary entry for 'memorization' is drafted for Ayla in docs/overnight-review-2026-09-18.md, section 3, awaiting her approval; the page defines the word in a clause meanwhile. Once approved, it gets a Definition box here and this chapter's introduces list.
16. evaluation/evaluation-foundations/inter-annotator-agreement The sentence saying that a score far above the agreement between annotators is not always desirable (a system that reproduces the choices of particular annotators) is Ayla's own judgement, of 2 October 2026: Boguslav & Cohen (2017) support that systems can exceed the agreement and that agreement and system scores are correlated, not this argument.
17. machine-translation/neural-and-llms/llms-as-translators Where the outputs come from. Dutch: 2023 services and ChatGPT runs from the study-day deck of 2023-03 (Archive/Arch_Outreach/2023_studiedag_taaltechnologie.zip, slides 37 to 39; the GPT-4 slide also has two 'Try again' outputs, 'protective cup' and 'jockstrap', left out); 2024 runs from Taaldag 2024 and EU Justitie 2024; 2025 services from ITF plenary (2025-09); 2026-01 runs from OTTIAQ (2026-01); 2026-10 runs by Ayla on 13 October 2026. English-French: TRA6600_05a_250212 (2025) and TRA6600-26_06 (2026), the 2026-10 runs by Ayla. ChatGPT prompt always 'translate into English/French:' plus the text; eTranslation always without AI enhancement; 2026-10 ChatGPT runs in a new temporary chat, medium effort. ChatGPT models, traced from release dates and not confirmed: 2023-02 GPT-3.5, 2024-03 GPT-4 Turbo, 2024-05 GPT-4 or GPT-4o, 2025-09 GPT-5, 2026-01 GPT-5.2. Judgements follow the slides' colours, except where checked and changed: 'take her on' partly right (unidiomatic, not a mistranslation), 'shabrak'/'shabrack' partly right (an English word for a saddle cloth), 'on the outside foreleg' partly right, 'muguet' wrong (OQLF GDT 8876910: oral candidiasis; the horse's thrush is 'pourriture de la fourchette', 17055773), 'crins' partly right. The 2026-10 rows are judged by Claude. References corrected 13 October 2026: Dutch 'Make sure she has a riding hat' (Zorg dat is an imperative), French 'pourriture de la fourchette' (OQLF form).

Time-sensitive claims (67)

Correct when written and dated with the products or leaderboards they describe (<Recheck note="..." by="YYYY-MM-DD" />). Dated ones first, soonest first: the order to work through before a new term.

1. using-llms/product-features/what-a-product-adds by 15 November 2026 · Recapture the two + menus once the merge of Claude's Chat and Cowork has reached every plan: the mode switch and the web search toggle disappear with it, and both screenshots show them.
2. capabilities/tools-and-agents/computer-use by 1 December 2026 · Current computer-use results: the OSWorld 2.0 figures (best 20.6 %, Claude Opus 4.8, July 2026 preprint) and the OSWorld-Verified best score (about 61 %, July 2025), and whether OSWorld 2.0 has been published at a venue.
3. critical-issues/copyright-and-authorship/court-cases by 1 December 2026 · The status of every case in the table: the Bartz appeals, the Ross appeal, the NYT case, the Ontario appeal, the Getty appeal, and any appeal in the Meta case; the CJEU reference Like Company v Google Ireland (C-250/25).
4. critical-issues/environment/data-centres by 1 December 2026 · IEA data-centre figures (Energy and AI, April 2025; Key Questions on Energy and AI, April 2026): a new edition or update may have appeared. Checked 23 September 2026: Key Questions (April 2026) was still the latest.
5. critical-issues/environment/data-centres by 1 December 2026 · Han et al. (2412.06288): replace the preprint with the Communications of the ACM version once published, and recheck its figures.
6. critical-issues/hallucinations/mitigating-hallucinations by 1 December 2026 · Hallucination leaderboard figures: lowest rate 1.8 %, GPT-5.4 at 7.0 % against 3.1 % for its nano version, from the version of 11 May 2026.
7. machine-translation/wmt/what-a-shared-task-is by 1 December 2026 · The 2026 findings paper (WMT26, 28 October 2026 and 29 October 2026) may change the procedure: update the box at the top, and check the test sets, the tracks (parameter limit), AutoRank and the annotators against it.
8. machine-translation/wmt/what-wmt-found by 1 December 2026 · Add the 2026 edition to the table once its findings paper is out (the conference is on 28 October 2026 and 29 October 2026): its subtitle and the change it made. Check that the four findings still read as the trend, and that the task list in the second table matches the 2026 site.
9. critical-issues/copyright-and-authorship/detecting-generated-text by 15 December 2026 · The Orélien case in the box 'A literary prize and a detector': whether any independent analysis or ruling has settled it since 14 October 2026.
10. using-llms/llm-benchmarks/translation-benchmarks by 15 December 2026 · The WMT bullets after the 2026 findings paper appears (usually November): which system led the human ranking and whether automatic and human rankings disagreed again.
11. critical-issues/copyright-and-authorship/detecting-generated-text by 5 January 2027 · Newer peer-reviewed evaluations of detectors, and whether the bias against second-language writers is confirmed with current tools (Hadra et al. 2026 report only a borderline trend).
12. critical-issues/copyright-and-authorship/fair-use-and-fair-dealing by 5 January 2027 · Any Canadian legislation following the 2023 consultation, and any Canadian court decision on training.
13. critical-issues/copyright-and-authorship/opting-out by 5 January 2027 · How the AI Office enforces Art. 53(1)(c), and any court decision on what counts as a machine-readable reservation.
14. critical-issues/copyright-and-authorship/plagiarism-and-disclosure by 5 January 2027 · The ACL and Waterloo policies and COPE's statement.
15. critical-issues/copyright-and-authorship/who-owns-generated-text by 5 January 2027 · CIPPIC v. Sahni; Allen v. Perlmutter (D. Colo., prompt-only output, undecided as of August 2026); any Canadian legislation; the ownership clauses in Anthropic's and OpenAI's terms.
16. critical-issues/environment/hard-to-measure by 5 January 2027 · Which providers publish model-level environmental figures (Mistral's LCA of July 2025, Google's inference paper of August 2025, OpenAI's per-query figure of June 2025). Checked 23 September 2026: no newer model-level disclosure found, and Anthropic still publishes none.
17. critical-issues/environment/putting-it-in-proportion by 5 January 2027 · IEA figures (Energy and AI, April 2025: emissions 2024 and 2035, share of demand growth, AI-enabled savings; Key Questions on Energy and AI, April 2026: web-search estimate, per-query disclosures), the per-query figures from Mistral and Google (2025), and the Masley calculator's numbers (captured 24 September 2026). Checked 24 September 2026: still the latest from each.
18. critical-issues/environment/training-and-inference by 5 January 2027 · Epoch AI estimates of the largest training runs (Grok 3 power and duration, on which the 240 GWh and 110 kt calculation is based; GPT-6 Astra power draw), and whether any provider has published training emissions for a frontier model. Checked 23 September 2026.
19. critical-issues/power-and-access/opacity by 5 January 2027 · The transparency index (editions so far: 2023-10, 2024-05, 2025-12); the training-content summaries published under the AI Act (older models have until 2 August 2027).
20. critical-issues/power-and-access/unequal-access by 5 January 2027 · Subscription prices and what free plans include; the supported-country lists of the two providers.
21. critical-issues/privacy/data-as-business by 5 January 2027 · ChatGPT advertising (plans and markets), and whether providers' training defaults change; the vendor details themselves live on using-llms/reliability-and-safety/where-your-text-goes.
22. critical-issues/privacy/privacy-law by 5 January 2027 · Whether the Garante appealed the Rome annulment, whether the Irish authority takes up the case, whether the Canadian findings led to further action, and any new regulatory decision on chatbots and privacy in Canada or the EU.
23. critical-issues/skills-and-learning/the-brain-study by 5 January 2027 · Whether Kosmyna et al. (2025) has been published after peer review (still a preprint, v2 of 31 December 2025, as of 22 September 2026), and whether its findings were revised.
24. using-llms/choosing-models/chat-api-or-local by 5 January 2027 · API prices per million tokens for the cheapest and the flagship model of both providers, whether the largest open-weight models still trail the best hosted ones, the price table and the two ratios under it (output five to six times input; about ten times between tiers at Anthropic, about fifty at OpenAI), and the two pricing screenshots.
25. using-llms/choosing-models/families-and-tiers by 5 January 2027 · The two naming schemes in the aside (OpenAI's Luna, Terra, Sol, Astra; Anthropic's haiku, sonnet, opus, fable), the generations timeline in the second section (add the new releases), the two model-list screenshots, and whether Anthropic and OpenAI still lead the field, with the Intelligence Index chart recaptured if the lead has changed.
26. using-llms/choosing-models/thinking-effort by 5 January 2027 · The whole September 2026 section: whether Claude and ChatGPT Work still pair a model list with five effort levels, whether the ChatGPT chat still hides the model behind thinking levels, the plan table row by row, and whether the three screenshots still match the products.
27. using-llms/llm-benchmarks/human-preference-rankings by 5 January 2027 · The arena's name, vote and model counts, whether style control is still applied by default, whether the factuality adjustment is still optional, and whether French and longer queries are still categories.
28. using-llms/llm-benchmarks/reading-a-leaderboard by 5 January 2027 · Which benchmark is the current hard one and the best score on Humanity's Last Exam (59.1% at Artificial Analysis on 17 September 2026; the sentence says 'above half'); and whether the Artificial Analysis Intelligence Index still averages ten tests.
29. using-llms/product-features/agents by 5 January 2027 · The kinds of agent chat products offer, whether they are still mostly on paid plans, and the names each product gives them; whether Claude's Research is still described as a lead agent with subagents, and the fifteen-fold token figure (Anthropic, 13 June 2025).
30. using-llms/product-features/code-and-connectors by 5 January 2027 · The code execution and connector screenshots; whether file creation is still off by default on Claude's personal plans; the names under which services are connected; and the permission settings.
31. using-llms/product-features/deep-research by 5 January 2027 · Whether deep research still asks clarifying questions or shows a plan in each product, the names it goes under, and the four research screenshots.
32. using-llms/product-features/documents-images-voice by 5 January 2027 · The names a document panel goes under in each product; the document, image, voice and study mode screenshots; artifact publishing; the voice options; whether Claude generates images; and whether a study mode is still limited to a university version in some products (Claude for Education on 16 September 2026).
33. using-llms/product-features/what-a-product-adds by 5 January 2027 · The six screenshots; the sidenotes, which state what is true per product and per plan: memory defaults, project and file limits, project sharing, which plans offer skills, and what each provider says about switching models.
34. using-llms/prompting/improving-a-prompt by 5 January 2027 · Which providers still offer a prompt rewriter in their developer tools, and what OpenAI's prompt optimizer became after its evaluation platform closed on 30 November 2026.
35. using-llms/prompting/standing-preferences by 5 January 2027 · The personalization screenshot and ChatGPT's 'fast answers' that ignore personalization and memory; the forms change with every redesign.
36. using-llms/prompting/structure-the-prompt by 5 January 2027 · The structured-output modes of Anthropic and OpenAI and the quotations from their documentation; the providers' prompting guides and what they recommend for structuring a prompt.
37. using-llms/prompting/what-prompting-is by 5 January 2027 · The two prompt-engineering guide links and whether their advice still matches the chapter (examples, tags, long documents first, no step-by-step on reasoning models).
38. using-llms/prompting/working-with-documents by 5 January 2027 · At what size each product switches from adding a file whole to searching it, and the file limits per plan.
39. using-llms/reliability-and-safety/where-your-text-goes by 5 January 2027 · Training defaults, opt-out names, retention periods and the plan table, for the providers cited and for any others worth naming by then; these terms changed twice in 2025 and will again.
40. evaluation/mt-evaluation/human-and-machine by 15 January 2027 · Ferstler et al. (2026): replace the arXiv version with the Anthology version of Findings of EMNLP 2026 once published, and check the figures quoted.
41. evaluation/mt-evaluation/reading-the-scores by 15 January 2027 · The WMT 2025 practice in 'Combining methods' once the 2026 findings paper is out.
42. capabilities/jagged-frontier/the-boundary-moves by 1 March 2027 · Strawberry answered correctly by current models: last checked 11 September 2026 on GPT-6 Astra and Claude Fable 5.1 (spelling-and-counting).
43. capabilities/jagged-frontier/what-they-reliably-fail-at by 1 March 2027 · Spatial reasoning as a persistent weakness: Cai et al. (2025) on GPT-5 is the newest source; check whether current models still fall well short of people.
44. critical-issues/hidden-work/how-it-is-paid by 1 March 2027 · The McKinney settlement (final approval hearing 30 October 2026), newer Fairwork ratings, and newer reporting on pay at the main annotation vendors.
45. critical-issues/hidden-work/the-cost-of-moderation by 1 March 2027 · The Kenyan moderators' cases against Meta and Sama (Kenya Law, Petition E052 of 2023 and E071 of 2022): a judgment changes the paragraph.
46. critical-issues/information-ecosystem/ai-slop by 1 March 2027 · The examples in 'Where it lands' age quickly; check that each still describes a current practice, and whether the platforms named have changed their rules on generated content.
47. critical-issues/information-ecosystem/generated-content-online by 1 March 2027 · The Pew figure and the table: newer measurements of generated shares appear every few months, and the July 2026 snapshot will be a year old.
48. critical-issues/information-ecosystem/model-collapse by 1 March 2027 · The estimate of when public human text runs out; newer work on model collapse (for instance Schaeffer et al., ICML 2026 position track, on the conflicting definitions of the term).
49. critical-issues/information-ecosystem/peer-review by 1 March 2027 · Submission counts at the large AI conferences, the AAAI pilot (a preprint that may be published), and any newer measurement of model use in papers and reviews after 2024.
50. critical-issues/information-ecosystem/reviewing-rules by 1 March 2027 · Venue policies change every year: the NeurIPS, ICLR, ICML and AAAI rules for the next edition, and any rule adopted by journals or publishers.
51. critical-issues/information-ecosystem/the-open-web by 1 March 2027 · The Pew figures, the default crawler blocking, the licensing deals, and whether search engines have changed how AI summaries link to sources.
52. critical-issues/jobs/forecasts-and-fears by 1 March 2027 · Newer measurements and exposure estimates with later models; whether any study measures which work keeps its price.
53. critical-issues/jobs/freelance-work by 1 March 2027 · The ELIS survey has a new edition every March; newer platform studies may cover translation and later model generations.
54. critical-issues/jobs/salaried-work by 1 March 2027 · Both payroll studies are working papers that may be published or revised; newer evidence on entry-level hiring.
55. critical-issues/knowledge-and-trust/knowledge-collapse by 1 March 2027 · The homogenization studies are from 2024 to 2026 and more are coming; the list and the 2026 synthesis should be checked against newer work.
56. critical-issues/knowledge-and-trust/marking-generated-content by 1 March 2027 · Enforcement of AI Act Art. 50(2) (a grace period to 2 December 2026 for systems already on the market, per the Commission's FAQ), which providers watermark text and images, and which platforms display Content Credentials.
57. critical-issues/knowledge-and-trust/telling-real-from-generated by 1 March 2027 · Whether newer studies of human detection with 2026 models change the picture.
58. critical-issues/power-and-access/dependence by 1 March 2027 · Which companies release the largest open-weight models, and under which licences.
59. critical-issues/power-and-access/market-concentration by 1 March 2027 · The count of partnerships, the country tallies of notable models (AI Index, every April), and which companies lead: all of it moves within a year.
60. critical-issues/skills-and-learning/assessment by 1 March 2027 · Performance of current models on university assessments and the reliability of current detectors.
61. critical-issues/skills-and-learning/learning-with-a-model by 1 March 2027 · Newer experiments with later models and longer follow-up; whether the study modes of chat products have been evaluated.
62. critical-issues/skills-and-learning/skill-decay by 1 March 2027 · Longer studies of deskilling are likely to appear; the 'what is not known' section should shrink as they do.
63. machine-translation/neural-and-llms/what-still-breaks by 1 March 2027 · The Last Translation Benchmark is a live dataset with tagged releases and a public leaderboard (last-translation-benchmark.vilda.net); the figures above are LTBv1 (contributions before 1 September 2026). Check the leaderboard for a newer release and the current best pass rate.
64. classical-nlp/normalisation/reducing-variation by 1 June 2027 · The four type counts are Ayla's own word-list measurement in Sketch Engine (slide 44 of the H2026 TAL deck), undated; the token and word counts are the corpus's published statistics, confirmed on 4 October 2026.
65. machine-translation/how-mt-works/four-generations by 1 August 2027 · The claim that older generations survive only where translated text is scarce, and that neural systems are becoming hybrids, is an observation of 2026-10; check it against the services and the WMT findings of the year.
66. machine-translation/rules-and-statistics/rule-based-mt by 1 August 2027 · Rule-based and statistical systems replaced for well-resourced pairs and surviving where translated text is scarce: an observation of 2026-10, the same as on how-mt-works/four-generations; check both together.
67. using-llms/choosing-models/matching-a-task The table's examples of what a small model at low effort suffices for; each release moves the line.

From the teaching notes (11)

Asked for after a class, in Notes for next year (action: true). Everything else there is a note about the teaching and stays out of this list.

1. Wordfast's Clean Up function was never covered: the chapters teach Save File As Translated as the one delivery route, and no page names Clean Up. Noticed while preparing the mid-session exam. Next year, show it on the delivery page beside Save File As Translated, with what it does to the memory, so students recognise the word when they see it in the menus or in another tool. TRA2600 · 10 October 2026
2. The exam template was not clear enough on what to write in the journal about Transcheck, and some students lost a lot of time at the end trying to copy-paste the whole Transcheck report into it. The consignes say "chaque signalement corrigé ou justifié dans le journal", which reads as every flag going into the journal, while part 3 of the journal only asks for the flags left uncorrected. Next year: say in the consignes that a corrected flag needs nothing in the journal, say in part 3 not to paste the report, and allow one line per kind of flag (for instance "Spelling: Alignio, termbase, everywhere") rather than one per segment, since the noisy checks (Spelling, Untranslated Text, Edited Exact) produce dozens. TRA2600 · 14 October 2026
3. The first part did not go well. Word embeddings still went reasonably, but contextual embeddings and attention and transformers were far too much: too technical, and not clear. I clearly lost them here. Both definitely need to be improved before I teach them again; not yet sure whether that means only the slides, a light mode, or the pages themselves. What went wrong, I think: the mix of still giving some technical detail without the technical baggage to carry it. The balance was not right. TRA6008 · 10 September 2026
4. A light version of the teach mode beside the full one: high-level, with far less terminology, while the pages keep the level of detail and accuracy they have now. Even the decks were pitched too high for this group in places. Even when a séance teaches only part of a chapter, the chapter is still written in full, so it is known to work as intended. TRA6008 · 10 September 2026
5. The séance went well, but there was probably still a little too much content to cover: using LLMs (choosing a model, prompting, risks and safeguards), then GIGO, hallucinations, bias and environment, with the Projet 1 briefing already moved to séance 4. The pages carry a lot of text, so the class could not go through all of their nuance; the decks for this séance could be lighter, which is what the light teach mode noted at séance 2 would give. TRA6008 · 24 September 2026
6. One student took the three remises of the parole parcours for the three projects and missed the 25 September 2026 deadline, although the date was in the plan de cours, on the site's calendar and dates table, on StudiUM's own assignment due dates, and presented in class; the other eight were on time. What helped after the fact: putting the self-paced blocs first on the class page, and saying at the end of a séance what is NOT to be read. What was really wrong on my side: the live site was three weeks behind the working tree, so the bloc page showed a reading due for the wrong séance. Push after any session that changes course dates. TRA6008 · 30 September 2026
7. Two of nine students missed the formats quiz (due 11 October 2026) and a third did it a day late, after noticing on StudiUM the next morning. The quiz is an MS Forms link, not a StudiUM activity, so StudiUM had no due date to show and sent no reminder; the student who was late named the missing reminder, although the date was in the title of the StudiUM link itself, in the plan de cours, on the site's calendar and said in class. The published rule (20 % per day) was applied to all three. Next year: wrap each parcours questionnaire in a StudiUM assignment with the due date, or post the link as a dated activity, so the course page shows it as due the way the graded remises are. The parcours deadlines also all fall on Sundays between séances, with no class to anchor them. TRA6008 · 12 October 2026
8. Grading the Word dictation part of the parole parcours (ten submissions) showed what the assignment page leaves open. (1) The reflection must be dictated, but nothing makes that checkable: two reflections showed no trace of dictation and said nothing about it. Ask for one sentence on how the dictation handled the reflection itself, so that the observation is gradable without guessing who typed. (2) Say which language the reflection is dictated in, or ask students to state it: three dictated in French and none mentioned switching the dictation language. (3) Ask students to list explicitly what they could not do by voice. Several gaps were simply left (INSERT tags, brackets, a numbered list as bullets) with no mention in the reflection, which is the one thing the page says not to do. (4) Ask them to try auto-punctuation on and off and to say what differed; the two who did produced the best observations. (5) Recurring problems worth a line in the demo page's tips box: "one" dictated as the digit 1, capitals appearing on redictated words, a punctuation command transcribed as text ("period", "full stop"), "select" failing on a misspelled word, a single letter that cannot be corrected (delete the word and redictate it). (6) Submission: three of ten files were off (a zip holding a PDF, first name before last name, a double .docx extension). Have StudiUM accept only .docx, and repeat the file name on the assignment page. (7) Scale used, settled after a first pass that gave eight of ten full marks: 2 points per part, and the reflection marked on four steps (every point with examples, a cause named, a link to the module; then without the link; then described only or one point empty; then two points empty), a quarter point at most for format and name, nothing for length. A late submission's 50 % ceiling is a cap applied per part (the earned mark or half the maximum, whichever is smaller), stated beside the earned mark in the feedback. (8) The 100 to 150 words asked for the reflection were too few: nine of ten ran over, up to about 240, and the longer ones were the useful ones. Raise the range. TRA6008 · 7 October 2026
9. Grading the PowerPoint subtitling part of the parole parcours (ten submissions). The assignment worked: every student captured both tests with legible subtitles, and the reflections were precise. What the page should change: (1) raise the word range (150 to 200 was too few for five points with examples; most ran 180 to 320); (2) ask for numbered captures cited from the text, which one student did and which made her reflection checkable; (3) make the five points the required headings, since the only content deduction was a point skipped entirely (the translation quality); (4) say that a phone photo of the screen is acceptable when capture fails during the slideshow (one student); (5) repeat the file name, since the same students as in part 1 got it wrong (a zip of folders, a name without "parole2"). Observations worth a line on the demo page: numbers spelled out come back as digits, acronyms lose their capitals or become words (BERT → verte, BLEU → blue, API REST → happy rest), the live correction of subtitles hinders reading, and the subtitles vanish too fast for several students to judge the text-versus-image difference at all. TRA6008 · 7 October 2026
10. Grading the HeyGen part of the parole parcours (nine submissions). All eight links resolved to public pages. What the page should change: (1) the provided video has no figures and few technical terms, so the questions about figures and terms are moot on it; adapt them or add a figure and two terms to the script; (2) "facteurs testés" as a separate point was the one most often skipped; fold it into the questions that test each factor; (3) recommend translating into a language the student knows well: the two who chose Arabic and Korean wrote the finest translation analyses, with something to say about register and terminology; (4) raise the word range (240 to 480 words written against 150 to 300); (5) keep the fallback that worked for a privacy objection, the reflection written from my own translated video, and state on the page by when such an objection has to be raised, since this one arrived at 19 h on the deadline day. (6) Nothing on the site or in the plan de cours says how the parole module's 6 % splits over its three parts; they were marked 2 / 2 / 2, as last year's 1.5 each. Say so in the table row of « Le parcours en ligne » (trois travaux de poids égal) and keep the three StudiUM maxima equal. TRA6008 · 7 October 2026
11. Every project gets ONE practical page from the day it is launched: the steps in order, who does what, every deadline written out, what is graded and how, which page holds each result and what to choose on it, and the templates. Other pages explain or compute; they never carry a practical instruction the overview does not. Do it for Projet 2 before it is launched at séance 6, and for all three projects next year. TRA6008 · 8 October 2026

Sources awaiting Ayla's verification (0)

Metadata is checked, but only Ayla verifies: verification is the source's presence in her Zotero "learn" collection (auto-exported to src/content/zotero.json). The RELEASE build refuses any page referencing these. Open the URL, confirm, save it to Zotero.

Content review due (0)

Non-stub pages never review-dated or last reviewed over 12 months ago (reviewed: in the page frontmatter, YYYY-MM-DD). AI content ages fast.

Bilingual pages out of step (0)

Each language is its own file, and a page whose two files no longer have the same headings, code blocks, lists, asides, figures or definition boxes fails the build. This list is the softer check, length: French normally runs 1.03 to 1.16 times the English here, so anything outside 0.95–1.3 is worth a look. Nothing here means all 258 bilingual pages are in step.

Glossary terms used but not marked up (33)

Every mention of a term goes through <Term> or <Definition>, which is what builds the tooltips and the "mentioned in" index. This lists a term when a page uses one of its written forms in prose and carries no marker for it anywhere on the page, so a page that marks a term once and then repeats it is not listed. Code samples, attributes and the marked mentions themselves are not searched. Forms under four characters are skipped, and one form can belong to two entries (token¹/token²), which is why some rows name more than one key.

1. course-materials/tra6008-projet1-pas-a-pas/pas-a-pas.fr segment ×3 (segment) · type ×2 (type) · raisonnement ×2 (reasoning) · correspondance exacte (exact-match)
2. evaluation/mt-evaluation/chrf.fr référence ×6 (gold-standard) · groupe (capturing-group) · segment (segment)
3. evaluation/evaluation-foundations/the-gold-standard.fr données d'entraînement (training-set) · type (type) · élément (xml-element)
4. course-materials/tra2600-assignments/examen-consignes.fr liste de blocage ×3 (blocklist) · type (type)
5. course-materials/tra2600-course/examen-mi-session.fr balise (tag) · liste de blocage (blocklist)
6. language-as-numbers/contextual-embeddings/reading-in-both-directions.en ambiguity ×2 (ambiguity)
7. language-as-numbers/contextual-embeddings/reading-in-both-directions.fr ambiguïté ×2 (ambiguity)
8. llm-lifecycle/llms-next-word-predictors/why-prediction-needs-knowledge.en ambiguity ×2 (ambiguity)
9. llm-lifecycle/llms-next-word-predictors/why-prediction-needs-knowledge.fr ambiguïté ×2 (ambiguity)
10. classical-nlp/normalisation/reducing-variation.en tagset (tagset)
11. classical-nlp/normalisation/reducing-variation.fr étiquetage morphosyntaxique (part-of-speech-tagging)
12. corpus-analysis/sketch-engine/absolute-and-relative-frequencies.fr normalisation (normalization)
13. course-materials/tra6008-course/evaluation.fr agent (agent)
14. course-materials/tra6008-project-briefs/projet-1.fr silence (silence)
15. critical-issues/environment/what-one-use-costs.en ambiguity (ambiguity)
16. critical-issues/environment/what-one-use-costs.fr ambiguïté (ambiguity)
17. evaluation/evaluation-foundations/the-gold-standard.en type (type)
18. language-as-numbers/attention-transformers/the-bottleneck.en ambiguity (ambiguity)
19. language-as-numbers/attention-transformers/the-bottleneck.fr ambiguïté (ambiguity)
20. language-as-numbers/why-language-is-hard/ambiguity.en ambiguity (ambiguity)
21. language-as-numbers/why-language-is-hard/ambiguity.fr ambiguïté (ambiguity)
22. machine-translation/neural-and-llms/neural-mt.en word vector (word-vector)
23. machine-translation/neural-and-llms/neural-mt.fr groupe (capturing-group)
24. using-llms/prompting/giving-examples.fr ambiguïté (ambiguity)

Modules without a description (0)

expect:, prerequisites: and structure: in modules.yaml: the three sections every module page opens with. Stub-only modules can wait.

Translation provenance not recorded (0)

Bilingual chapters without translation: bilingual, or without both writtenIn: (en or fr) and translation: (machine or revised). Until one of the two is set, the chapter and its pages say nothing about where the second language came from.

Missing French titles (55)

1. 0 modules without a French title
2. 0 modules without a French blurb
3. 0 chapters without a French title
4. 55 pages without a French title

Show the chapters

Glossary (277 entries)

1. 0 entries without a source
2. 0 entries without a French term
3. 0 entries without a French definition
4. 0 French terms still marked "to verify"
5. 7 FR terms provisional, references disagree — pick the form to use throughout: foundation-modelcomputer-useagentic-aiai-slopxml-namespacenormalizationsentence-segmentation
6. 28 FR terms marked as own suggestions, still to attest: embedding-arithmeticencoder-only-modeldecoder-only-modelknowledge-cutoffcustom-instructionspersistent-memoryfew-shot-promptingzero-shot-promptingreasoning-efforttest-time-computationharnesscomputer-useintrinsic-hallucinationextrinsic-hallucinationcoherence-correspondenceamplificationtraining-emissionsinference-emissionshuman-preference-evaluationtoken-based-billingprompt-optimisationinline-taganchoralternationalignment-taxmodel-specificationnormalizationsentence-segmentation
7. 0 entries without a definition

Glossary entries no chapter introduces (22)

Each needs a home in some chapter's introduces list, or a decision that it stays tooltip-only. Where the entry carries an introduce suggestion (planning note, never rendered publicly), it is shown after the arrow.

Show the 22 entries

representation-learning → language-as-numbers/word-embeddings hyperparameter → foundations/supervised-learning training-set → foundations/supervised-learning validation-set → evaluation/evaluation-foundations static-word-embedding → language-as-numbers/word-embeddings latent-space → language-as-numbers/contextual-embeddings attention-mask → language-as-numbers/attention-transformers encoder → language-as-numbers/attention-transformers decoder → language-as-numbers/attention-transformers encoder-only-model → language-as-numbers/contextual-embeddings decoder-only-model → language-as-numbers/attention-transformers causal-language-model → llm-lifecycle/llms-next-word-predictors masked-language-model → language-as-numbers/contextual-embeddings denoising → language-as-numbers/contextual-embeddings gpt → llm-lifecycle/llms-next-word-predictors web-scraping → llm-lifecycle/training-data bias-term direct-preference-optimization → llm-lifecycle/finetuning few-shot-learning zero-shot-learning data-extraction oligopoly

Introduced, but with no definition box (0)

The other direction: a chapter's introduces names a term that no page of that chapter defines with a <Definition> box. The glossary's "introduced in" points at the chapter anyway, so the entry promises a box that is not there. A page or chapter split is what usually causes it, by moving the box and leaving the claim behind. Chapters with nothing written yet are not counted.

Stub pages (1)

1. Multimodality 1 page still placeholder

Placeholder chapter metadata (1)

Chapters whose objectives or "To remember" summary are still generated placeholders.