- Runtime model folder now models/devseis-ai-act-classifier-v6/ (copy of Devseis/devseis-ai-act-classifier-v6); v5 folder removed
- Results, Story (new v6 chapter), glossary, lessons, split diagram and print sheet updated to v6 figures
- 'Uncertain - needs expert review' when the trained model's top probability is below 50% (stricter cut-offs tested on validation and rejected: errors are mostly confident)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Certificate: 'by' + Devseis logo under the course details
Logo moved from the top to the centre of the bottom row (between date and signatory), sized to the gap; title block moved back up.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Adds devseis-logo.png (same file as the org card, hosted here so the certificate canvas can be saved) and draws it at the top of the certificate; text fallback if it can't load. Middle block spaced down slightly.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
More professional certificate design; score removed
Redesigned certificate (A4 landscape, double navy/gold frame, guilloche band, serif type, gold seal, date and signatory blocks). No score on the certificate; each piece of text appears once. Course title capitalised consistently for LinkedIn.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Load the model only when someone uses the checker
The ~58 MB model and engine download starts on focus of the text box, an example tap, or Check - not on page load. Check works immediately; the Trained model card shows download progress on a first visit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
21 days ago
Self-host transformers.js and the ONNX Runtime engine (no third-party CDN)
- vendor/transformers-3.0.0/: transformers.min.js and ort-wasm-simd-threaded.jsep.wasm, byte-identical to npm @huggingface/transformers@3.0.0 (what jsDelivr served), plus Apache-2.0 and MIT licences and a README
- index.html imports the local library and sets env.backends.onnx.wasm.wasmPaths to the local folder
- privacy.html: jsDelivr removed; before consent the browser contacts only Hugging Face
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Link Devseis LinkedIn company page (certificate, Built by Devseis, footer)
Certificate 'Add to LinkedIn profile' now uses organizationId=12603388 so the certification shows the Devseis company page and logo. Adds 'Follow Devseis on LinkedIn' to the Built by Devseis section and a LinkedIn link in the footer.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Adds og-image.png (1200x630 @2x), Open Graph/Twitter image tags and a sharper description on the page, and the Space thumbnail so links to the Space page show it too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Results: model-vs-expert card, verified AI Act key dates, Built by Devseis section
- 'What the model can't do — and what an expert does' side-by-side card
- EU AI Act key-dates timeline built from the OJ texts (Reg 2024/1689 Art 111/113; Reg 2026/1744): past/next markers computed from today's date, Omnibus changes flagged (high-risk Annex III → 2 Dec 2027, Annex I → 2 Aug 2028); shareable #dates link; linked from lesson 2
- 'Built by Devseis' section (services, awards, accredited training, founder, free-call CTA, links to Nox, datasets and the collection) at the end of Results and Story
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Leads: expert-call CTA, printable summary, final assessment and certificate; Clarity project ID
- After each check: 'Book a free 30-min AI Act readiness call' box (pre-filled email draft, optional case details) with a message that adapts to the result
- 'Save or print this summary': one-page Devseis-branded summary with results, next steps and disclaimer
- Learn: new final step with a 12-question assessment (pass mark 9/12, shuffled retakes, explained review); certificate unlocks on pass, shows the score, drawn in the browser, LinkedIn add/share
- Guard against a late answer from an earlier click overwriting the current result
- Clarity now uses the Caveat project ID (yqfpt9xwwk), still loaded only after consent
- privacy.html: rows for the email draft, summary, certificate/LinkedIn, and the stored assessment score
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Privacy: Clarity only after consent, self-hosted font, rewritten privacy notice
- Microsoft Clarity no longer loads by default; it loads only after 'Accept analytics' (consentv2 signal sent), choice kept 12 months, 'Analytics settings' in the footer to change it, Clarity cookies deleted unless consent is given (also removes cookies left by the old snippet)
- Space Grotesk self-hosted under fonts/ (SIL OFL 1.1) — no requests to Google Fonts
- privacy.html rewritten: controller, each data flow with its legal basis, local storage, Clarity cookies and retention, transfers, rights, CNPD, and an on-page consent control
- README: privacy notes and file list
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Rename model to Devseis AI Act Classifier v5; link Hub model and dataset repos
- Runtime model folder renamed to models/devseis-ai-act-classifier-v5/ (copy of Devseis/devseis-ai-act-classifier-v5), with licence and attribution README
- Data and evaluation moved to Devseis/devseis-ai-act-classifier-v5-data and the model repo; removed from the Space
- Page: model shown as Devseis AI Act Classifier v5 (built on legal-bert-small); links point to the Hub repos
- Space README: model/dataset metadata, new paths, CC BY-SA 4.0 note for the model folder
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Redesign page (Check / Results / Story / Learn), publish v5 data and evaluation, update cards
- index.html: four tabs — minimal Check (8 examples, model details per card, auto-scroll to results), Results, an 11-chapter build story, and a 14-lesson plain-language course with glossary, quizzes and widgets; mobile layout
- Fix: status spinner no longer stays visible after an answer arrives
- models/eu-ai-act-classifier-v5/data/: v5 train/val/test CSVs, scenario list, v3–v5 change logs, data README
- models/eu-ai-act-classifier-v5/evaluation/: v5 test report, v4-subset comparison report, int8 export check
- Model card: Files section; Space README: tabs and file layout
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Part 1: switch to v5 model (legal-bert-small, in-Space), add GPAI rule, update stats
- Ship the v5 classifier inside the Space (models/eu-ai-act-classifier-v5/, int8 ONNX, 35.6 MB) with its model card
- Load it locally via transformers.js (localModelPath, dtype q8)
- Run gpai_rule.js before the classifier (GPAI providers decided by rule; 655/655 tests)
- Update page stats and How-it-works text to v5 results; update README architecture
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
23 days ago
Update Part 1 stats to Part 3 model; add Part 3 how-it-was-done section (more data is not automatically better)
