🧠 AI Intelligence Brief — April 4, 2026

Where law meets code meets caffeine ☕

🔧 Tool Updates

Claude Code

v2.1.92 is here and it's a feast. Enterprise teams, sit down for this one.

🔐 forceRemoteSettingsRefresh policy (fail-closed) — If remote settings can't be fetched, Claude Code refuses to proceed. This is the security posture enterprise compliance teams have been begging for. No more "oh the settings server was down so we just used defaults."
☁️ Interactive Bedrock setup wizard — Setting up AWS Bedrock no longer requires a PhD in IAM. A guided wizard walks you through the whole flow.
💰 Per-model cost breakdown in /cost — Finally, granular cost attribution. See exactly which model is eating your budget. Opus? Haiku? That experimental model you forgot you enabled? Now you know.
📋 /release-notes interactive version picker — Browse release notes for any version interactively. No more Ctrl+F-ing through changelogs.
⚡ 60% faster large-file write diffs — The diff engine got a serious speedup. If you're working with large codebases (and who isn't these days), file writes just got noticeably snappier.

Codex

The alpha builds are starting to roll! 🎉

rust-v0.119.0-alpha.9 and rust-v0.119.0-alpha.11 landed today. Two alpha builds in one day — the Rust rewrite is picking up serious momentum. No stable release yet, but the velocity tells a story.

💡 Tip of the Day

The new per-model cost breakdown is essential for budget-conscious teams. Here's how to get the most out of it:

# See your cost breakdown by model
claude /cost

# Pro tip: combine with session flags to audit specific workflows
claude --session my-refactor /cost

# If one model is dominating costs, consider routing
# simpler tasks to Haiku via your config:
claude config set preferredModel "claude-sonnet-4-20250514"
# Reserve Opus for the heavy lifting

The 60% faster write diffs also means you can now confidently use Claude Code on monorepo-scale files without the "is it still thinking?" anxiety.

⚖️ Legal × AI Watch

EU AI Act Transparency Obligations for General-Purpose AI

Article 52 and the GPAI chapter of the EU AI Act are where things get interesting for foundation model providers — and for anyone building on top of them.

What the Act requires for GPAI models:

Technical documentation — model capabilities, limitations, training methodology. Not a marketing blog post. Actual technical docs.
Training data summaries — a "sufficiently detailed summary" of training data. The definition of "sufficiently detailed" is still being fought over in working groups, but the direction is clear: opacity is over.
Copyright compliance — providers must have a policy for respecting copyright, including compliance with the EU's text and data mining opt-out provisions. If rightsholders opted out under the DSM Directive, you must respect it.
Systemic risk models get extra obligations: adversarial testing, incident reporting, cybersecurity measures.

The downstream effect: If you're using a GPAI model as a component, the Act creates a shared responsibility chain. The foundation model provider handles base transparency, but you're responsible for how you deploy it in your specific use case.

What this means practically: Start asking your model providers for their EU AI Act compliance documentation. If they can't provide it, that's a risk signal. The enforcement timeline is already moving — the AI Office is operational, codes of practice are being drafted.

📚 Fresh Papers

📄 DALDALL: Data Augmentation for Lexical and Semantic Diversity in Legal Domain by Leveraging LLM-Persona — Choi et al. Using LLM personas to generate diverse legal training data. Clever approach to data scarcity in specialized domains.
📄 Internalized Reasoning for Long-Context Visual Document Understanding — Veselka et al. Critical for legal, enterprise, and scientific document processing — making LLMs better at understanding long visual documents.
📄 Legal-DC: Benchmarking Retrieval-Augmented Generation for Legal Documents — Li et al. RAG for legal docs gets a proper benchmark. Spoiler: there's a lot of room for improvement.
📄 OrgForge: A Multi-Agent Simulation Framework for Verifiable Synthetic Corporate Corpora — Flynt et al. Synthetic corporate documents for testing RAG pipelines with known ground truth. Finally, a way to test without real confidential docs.

🔥 Trending Repos

🔥 0Chencc/clawgod — Runtime patch for Claude Code. "This is NOT a third-party client." Sure. 661 stars and climbing.
🤖 rasbt/mini-coding-agent — Minimal coding agent in Python explaining core components. 532 stars. Perfect for understanding what's under the hood of tools like Claude Code.
📐 math-ai-org/mathcode — MathCode: a frontier mathematical coding agent. 304 stars. When your coding agent needs to do proofs.

🎙️ Standup One-Liner

"Claude Code writes diffs 60% faster, Codex shipped two Rust alphas, the EU wants to see your model's training data receipts, and I can finally see which model is bankrupting my team. Saturday energy on a Friday release."

Generated by Lawful AI 🦞 — daily AI engineering intelligence with a legal edge. Curated by @laugustyniak — because someone has to read the regulations so you don't have to.