DRAFT — internal review

ET Weekly · the full back-catalogue

15 weekly newsletters, March 5 – June 10, 2026. Built from the live wire-incidents archive plus targeted external research. For Kariem's review before any send.

March — back-catalogue (live probes did not exist yet)

2026-W09 · 05–08 March 2026 · 3 wires + external
GPT-5.4 launched the same week the court-citation hallucination count crossed 1,200.

OpenAI shipped a new flagship and a thoughtful reasoning-models post. A researcher quietly updated a database that now tracks 1,227 hallucinated citations submitted to courts globally.

2026-W10 · 09–11 March 2026 · 2 wires + external
The redlines Anthropic wouldn't drop. The contract OpenAI signed instead.

950 Google employees signed an open letter backing Anthropic's stand. The Pentagon got what it wanted from two providers anyway. The third didn't blink.

2026-W11 · 17–19 March 2026 · 2 wires
The provider published its own internal failure-mode log.

OpenAI's "How we monitor internal coding agents for misalignment" is the most candid provider failure post of the quarter. Read what's named, then read what isn't.

2026-W12 · 23–25 March 2026 · 3 wires
The week OpenAI tried to professionalize harm reporting.

Three governance posts in three days: Sora safety, the Foundation update, and the launch of the OpenAI Safety Bug Bounty.

2026-W13 · 30 March – 05 April 2026 · 0 wires (quiet week)
The quiet week. The Charlotin database didn't slow down.

No provider releases in our archive this week. Roughly forty new hallucinated-citation cases. The Anthropic-DoD injunction held. Sometimes the week is what didn't happen.

April

2026-W14 · 06–09 April 2026 · 6 wires
Google's mental-health update. The EU's AI Continent Plan. The quietest big week of Q1.

Google published a substantive post on mental-health work. The EU AI Continent Plan delivered its major milestones. And the regulator started a consultation on AI energy consumption that almost nobody noticed.

2026-W15 · 15–16 April 2026 · 3 wires
Three significant launches in two days. The launches are starting to outpace the trust.

GPT-Rosalind. Agents SDK 2. Gemini 3.1 Flash TTS. Two days, three launches, and the gap between what providers ship and what users feel they can rely on is now visible.

2026-W16 · 20–24 April 2026 · 10 wires
GPT-5.5. The biggest release week of the year. Read past the launch post.

OpenAI shipped GPT-5.5, ChatGPT Images 2.0, Privacy Filter, Bio Bug Bounty, and the Codex enterprise scaling — five launches in one week. Read what's missing from each.

2026-W17 · 27–29 April 2026 · 4 wires + flagship external
Where the goblins came from. A 175% rise. A reward signal that did not know what it was rewarding.

OpenAI published the most important post of Q2 this week: a full explanation of why GPT-5.1 was suddenly mentioning goblins everywhere. Read the case study; build the equivalent for your own pipeline.

May (back-catalogue close + live cadence begins)

2026-W18 · 05–08 May 2026 · 10 wires
ChatGPT shipped ads. The EU shipped 'simplification'. The week the commercial frame won.

OpenAI shipped "New ways to buy ChatGPT ads" on Monday and Trusted Contact on Wednesday. The EU agreed to "simplify" the AI Act on the same day.

2026-W19 · 11–17 May 2026 · 13 wires (first live-probe issue)
Seven drift events. One bedside-manner model.

Anthropic Claude Haiku 4.5 missed its clinical-citation baseline every weekday this week. OpenAI shipped four products.

2026-W20 · 18–24 May 2026 · 14 wires
The first AI advertising fine. And the drift that spread to a third provider.

FTC fined Cox Media + 2 firms nearly $1M for an AI marketing product that didn't actually exist. Citation drift now visible on Anthropic, Bedrock, and Mistral.

2026-W21 · 25–31 May 2026 · 55 wires
Reddit told us what's broken. The labs told us what's coming.

47 community posts surfaced the same pattern this week — validation is the dominant failure mode, and the people who use the models all day already know.

June (live)

2026-W22 · 01–07 June 2026 · 17 wires
The EU drew a line. OpenAI ran for the high ground.

Cloud and AI Development Act dropped Thursday. OpenAI published three policy posts in 72 hours. The labs felt the regulation before it landed.

2026-W23 · 08–10 June 2026 · 5 wires (mid-week)
Two posts about benefiting everyone. Three measurements that the model still can't cite.

A quiet week in the wire. The drift didn't stop. The provider posts said it did, in different words.

About the back-catalogue (W09–W18). Live probe data did not exist before launch on 10 May, so March-April issues lean on external sources — the Damien Charlotin hallucinated-citations database, AIID, the AI Lawsuit Tracker, regulator filings, and the provider blogs we did capture. Source links are real. Every wire ID in the post-launch issues (W19 onward) comes from the live wire_incidents table. Nothing fabricated.

Flagship issue: W17 ("Where the goblins came from") is the back-catalogue's anchor — the OpenAI reward-hacking post is the most important provider candor of Q2 and the issue is built around it.