🧠 AI Intelligence Brief — April 6, 2026

Where law meets code meets caffeine ☕

🔧 Tool Updates

Claude Code

📭 No release today. The week gave us v2.1.89 through v2.1.92 — four releases in six days. If you're keeping score: interactive lessons, MCP persistence at 500K, executable plugins, fail-closed enterprise policies, Bedrock wizards, per-model cost tracking, and 60% faster diffs. Not bad for a week.

Codex

📭 No new Codex builds today, but the two Rust alpha drops from Friday (alpha.9, alpha.11) are worth watching. The rewrite is clearly in the "ship fast, iterate faster" phase.

Week in review: Claude Code shipped 4 versions. Codex shipped 2 alpha builds. The velocity gap tells you where the maturity curve is — but Codex's Rust foundation could close it fast once stable lands.

💡 Tip of the Day

Sunday recap — the three most impactful things from this week that you should actually configure before Monday:

# 1. Enable fail-closed remote settings (v2.1.92)
# In your org policy config:
# "forceRemoteSettingsRefresh": true

# 2. Check your per-model cost attribution
claude /cost
# If Opus is dominating, route classification/triage
# tasks to Haiku

# 3. Try the interactive release notes picker
claude /release-notes
# See exactly what changed between your current
# version and any other

Bonus: if you're a plugin author, v2.1.91's executable shipping support means you can bundle Rust/Go/C binaries with your plugin. The performance ceiling for Claude Code plugins just disappeared.

⚖️ Legal × AI Watch

Open Source AI and the EU AI Act — What's Exempt, What's Not

The EU AI Act has a carve-out for open-source AI — but it's narrower than most people think, and the details matter enormously.

What the Act says:

Free and open-source AI models are generally exempt from most GPAI obligations (documentation, transparency, copyright compliance) — unless they pose systemic risk.
Systemic risk threshold: If your open-source model has more than 10^25 FLOPs of training compute (or the Commission designates it), the exemption evaporates. You get the full GPAI obligations regardless of license.
The deployer isn't exempt. Even if the model itself benefits from the open-source exemption, whoever deploys it in a high-risk use case still bears all the deployer obligations under the Act.

The practical implications:

Small open-source models (most of them) — largely exempt from provider obligations. Release your 7B parameter legal reasoning model without writing a 200-page technical doc.
Frontier open-source models (Llama-scale and above) — probably above the compute threshold. Open-source license doesn't save you from transparency obligations.
Everyone who deploys open-source models in production — you're not exempt. You still need risk assessments, human oversight, and compliance documentation for high-risk deployments.

The open question: What counts as "free and open source"? The Act references the existing open-source definitions, but the AI community's debate over "open weights vs. truly open source" (training data, training code, evaluation harnesses) remains unresolved. Meta's Llama license, for instance, has commercial restrictions that might disqualify it from the exemption entirely.

Bottom line: Open source is a development methodology advantage under the Act, not a compliance get-out-of-jail-free card. If you're deploying open-source AI in regulated contexts, you still need the same rigor as proprietary deployments.

📚 Fresh Papers

📄 Bridging National and International Legal Data: Two Projects Based on the Japanese Legal Standard XML Schema — Nakamura et al. Computational comparative law using standardized legal XML. Infrastructure work that makes cross-jurisdictional legal AI possible.
📄 CVPD at QIAS 2026: RAG-Guided LLM Reasoning for Islamic Inheritance Share Computation — Swaileh et al. Multi-stage legal reasoning for inheritance law using RAG. A fascinating intersection of religious law, formal logic, and LLMs.
📄 Xpertbench: Expert Level Tasks with Rubrics-Based Evaluation — Liu et al. As LLMs plateau on conventional benchmarks, this paper pushes evaluation to expert-level tasks including legal reasoning. The bar is rising.
📄 Characterizing Delusional Spirals through Human-LLM Chat Logs — Moore et al. When LLM conversations go wrong — analyzing negative psychological patterns in chat logs. Important safety research with regulatory implications.
📄 Unmasking Hallucinations: A Causal Graph-Attention Perspective on Factual Reliability — Kurra et al. A causal approach to understanding why LLMs hallucinate, not just detecting that they do. Foundational work for trustworthy legal AI.

🔥 Trending Repos

⚔️ x1xhlol/better-clawd — "Claude Code, but better." OpenAI/OpenRouter support, no telemetry, no lock-in. 375 stars. The community fork energy is strong this week.
📝 clawplays/ospec — Document-driven AI development for coding assistants. 343 stars. Spec-first development meets AI-first tooling.
📸 chencore/deep-live-cam-tutorial — Deep-Live-Cam installation and usage tutorial for AI face-swapping. 80 stars. The legal implications of this one write themselves.

🎙️ Standup One-Liner

"Wrapped the week with 4 Claude Code releases, 2 Codex alphas, a deep dive into open-source AI regulation, and the realization that 'open source' under EU law means something very specific — and it probably doesn't include your favorite model's license."

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.