Ephemeris.
Today's file is about measuring what AI gets wrong — the facts a model absorbs by mistake, the side it tilts toward, and the rulers people are building to catch both.
It read that the fact was false. It learned the fact.
Owain Evans and colleagues fine-tuned models on documents that flatly deny a claim — "Ed Sheeran did not win the 100m gold" — and watched the models come away believing the opposite. Belief in the false claim climbed from 2.5% to 88.6%, almost what you'd get by feeding the lie straight. The fix is grammatical: fold the negation into the sentence instead of bolting it on as a separate clause, and the model learns it correctly. A real hazard for anyone training on synthetic corrections or "this is wrong" data.
Belief rate in a claim every training document explicitly marked as untrue. "Negation Neglect," 13 May 2026.
The bias you can't see in one answer
Long Phan, Dan Hendrycks and co-authors put a name and a number on covert political bias — when a model frames or engages with two opposing prompts asymmetrically, even as each answer looks reasonable on its own. They measure it two ways and then correct it with Political Consistency Training, a reinforcement-learning method that closes the gap, holds onto general usefulness, and carries over to benchmarks it never saw. A working ruler for a problem most teams only argue about.
Point the agent at the subdirectory, not the repo
Anthropic's field notes for running Claude Code on a monorepo read like ops advice you'd pay for. Keep CLAUDE.md files lean and layered so context loads additively as the agent walks the tree. Wire up an LSP so it follows a call to its definition instead of pattern-matching a name. Scope build and test commands per subdirectory to dodge timeouts. And name an owner for the configuration before adoption fragments into tribal lore.
The lights come back on, at forty percent
After roughly three months of state-ordered shutdown, Iran's internet is partially back — and Cloudflare Radar puts a number on "partially." Traffic has recovered to about 40% of pre-shutdown levels, a sharp but incomplete step in the curve. A reminder that connectivity is a policy variable as much as an engineering one, and that public measurement is how the rest of the world keeps score.
Requirements analysis, from weeks to hours
In OpenAI's account, the IT-services firm Endava rebuilt how it turns client requirements into working software around Codex and agentic delivery — compressing requirements analysis from weeks into hours.
The headline is the speedup. The lesson underneath it is organisational: Endava treats agents as a delivery layer the whole team plans around, rather than a tool a few engineers reach for.
That framing is what makes the change stick — agentic delivery rewrites the steps, the hand-offs and the estimates a team commits to. Read it as a template for where an org has to move once the work itself speeds up.
weeks → hours
That's today.
A thin day for genuinely new signal — so five stories with real weight, not ten with filler.
Rubric: tools worth adopting this week, dev & agentic coding, research with a practical kernel, and security/connectivity where the signal is high — measured against a ledger of every link Ephemeris has shipped, so nothing repeats.
Issue 044 · five stories · 01 Jun 2026.