November 11, 2026
Fix pass after an outside review
An independent review read every page against its notebook. We fixed what it found: numbers that had drifted between pages and notebooks, an odds-versus-probability slip in Module 2, Module 3 naming its planted leak before asking you to find it, Module 4's precision framing and random baseline, Module 7's test-set description, Module 9's call count, Module 10's grader agreement (now reported on held-out labels, with the in-sample figure called out), Module 12's meetings range and power figures, and Module 13's backoff arithmetic. Every module gained an operator path for readers who won't write code. Numbers on the pages now come straight from the starter kit's notebooks, and every build checks spelling, recomputes the check-yourself answers and tests every link.
What we checked:
- Anthropic: sampling parameters rejected on current models and removed from API version 2026-10-21; budget_tokens deprecated and rejected from Claude 4.7; thinking billed as output; 1M-token context; prompt caching (writes 1.25x, reads 0.1x, 5-minute lifetime, 512 to 4,096-token minimums); batches (50% off, most finish within an hour, 24-hour expiry); SDK retries twice by default; at least 60 days' notice before retirement; the Claude API doesn't offer fine-tuning (glossary).
- OpenAI: sampling parameters removed when reasoning effort is above none; reasoning billed as output; prompt caching (1,024-token minimum on newer models); the Agents SDK; the self-serve fine-tuning wind-down dates.
- Amazon Bedrock: Claude 3 Haiku fine-tuning reached end of life on September 10, 2026.
- Model Context Protocol: current spec 2026-07-28, matching the security guide linked in Module 9.
- Vercel: a new project's first deploy is production; default Python 3.12; Standard Protection covers everything except production domains; environment variables apply to new deployments. OpenRouter: per-key credit limits, prepaid credits, daily caps on free models, and an embeddings endpoint.
- Gmail (spam rate under 0.10%, never 0.30%; 5,000-a-day bulk threshold), Yahoo (unsubscribes honored within two days), CAN-SPAM, the ICO's B2B guidance, CASL, and GDPR Recital 47 on legitimate interests.
- 150 external links on the track pages returned 200.
November 11, 2026
Rewritten as a hands-on course
Every module now teaches on the page instead of pointing elsewhere, and every build has starter data, numbered steps, a done-when list and a worked solution in the public starter kit. Added Module 0 (setup and a first deploy), check questions and vendor questions in every module, this changelog, a glossary and a vendor-questions page. The reading list is now generated from the modules, start-here merged into the course page, and the per-module sales sections were removed. Module 4 was retitled, and claims that couldn't be sourced were computed on the page, labelled as assumptions or cut.
What we checked:
- Every external link in the course and the reading list returned 200 (91 URLs).
- Anthropic: sampling parameters rejected on current models and removed in API version 2026-10-21; thinking billed as output; adaptive thinking and effort; structured outputs; prompt caching minimums and prices; tool use; computer use guidance; model IDs, deprecations, errors and rate limits; batch processing.
- OpenAI: reasoning effort and which sampling parameters it disables; structured outputs; prompt caching; function calling; the Agents SDK; rate limits; the self-serve fine-tuning wind-down and its dates (deprecations page).
- Amazon Bedrock: Claude 3 Haiku fine-tuning listed as end of life (September 10, 2026, us-west-2).
- Model Context Protocol: what it is and the local-server security guidance.
- Vercel's Python runtime, Flask preset, environment variables and deployment protection; OpenRouter key credit limits.
- Gmail sender guidelines and Yahoo sender best practices: authentication, one-click unsubscribe, spam-rate thresholds.
- Email law orientation against primary sources: CAN-SPAM (15 U.S.C. § 7704), the ICO's B2B marketing and right-to-object guidance, and Canada's CASL.
November 11, 2026
Morning refresh
Updated redirected links, corrected the fine-tuning note in Module 11, brought Modules 6, 7 and 9 up to date (reasoning models, structured output, MCP and agent SDKs), and rewrote Module 12. All of it was superseded by the rewrite later the same day.
August 5, 2026
Opened as a free curriculum
The 13 modules, published one by one through June and July, were brought together on one course page as a free, self-directed curriculum with a depth target per module.
June 22, 2026
First modules published
The first pages went up, one short page per module, with the rest following through July.