AI Wiped Out $285 Billion in Software Stocks. Expertise Wasn't Touched.

Thomson Reuters down as much as 18%. RELX down 14%. Wolters Kluwer down 13%. Not because patent attorneys or contract lawyers became obsolete, but because Anthropic showed that the software wrapping their workflows can be replicated with an open-source plugin.

The expertise is fine. What's dead is the business model of charging subscription fees to sit between experts and their data.

What's interesting is what comes after. AI has automated semantic understanding at scale. A machine can now read a document and make useful sense of it. Which means for the first time, your methodology as an expert becomes scalable. Not just in legal or patents. Everywhere professionals evaluate things against criteria and track results. Compliance audits, electrical inspections, medical diagnostics, quality control.

People already capture this stuff in Word documents, Excel sheets, SOPs, proprietary tools. But there's no standard for it. No format that's portable across teams, tools, and the AI systems that can now actually execute it.

That's what I've been building. It's called Evaluation Language, and it's open source.

How It Works

Ask an AI to evaluate something and you get a wall of text. Give it an EvaL instruction set and you get structured data. Scores, classifications, tables, diagrams, exact quotes as proof. That's the difference.

EvaL is an open YAML-based schema for defining evaluation instructions step by step and capturing results with mandatory evidence. A recipe format for expert analysis. You define what to evaluate, in what order, what proof is required, what structured output you expect. Then any MCP-compatible chatbot can execute it. Claude, ChatGPT, whatever comes next.

The instructions are portable YAML files. They survive platform changes and vendor pivots. The AI landscape reshuffles every few months, but evaluation logic stays stable. How you inspect a building's fire safety doesn't change when Anthropic releases a new model. How a patent attorney checks claim support under Article 84 EPC doesn't depend on which LLM is trending this week.

The results are YAML too. Every evaluation includes a proof chain with source quotes, document references, and reasoning. Attachments and comments are supported. Not buried in chat text but stored as structured data where proofs and reasoning are first-class citizens. A human reviews, overrides where needed, shares the evaluation.

What MCP did for the input side, standardizing how AI connects to tools and data sources, EvaL does for the output side. How AI produces structured, verifiable results.

The schema is MIT licensed and on GitHub. The platform is live in alpha at eval.directory. It works today with Claude and ChatGPT.

• EvaL Schema (MIT licensed) - the open standard
• EvaL Directory (eval.directory) - the public registry, self-hostable on-prem like Docker Registry
• EvaL Ink (open source, releasing soon) - the native authoring tool

Why More Human, Not Less

Here's what the panic misses . The expert becomes more important, not less.

Think about any professional who evaluates things for a living. A senior patent attorney checking description support. An electrician inspecting a building's wiring against safety regulations. An auditor working through a compliance checklist. Their methodology lives in their head, maybe in some internal guidelines. It doesn't scale. It can't be shared, audited, or improved collaboratively. Junior colleagues learn it through years of osmosis.

Now that methodology can be encoded as a reusable instruction set. The AI executes the tedious parts, works through each step, returns structured results with evidence. The expert reviews, corrects, refines. Their judgment becomes a living, versioned artifact that the whole team can use and improve.

Funnily enough, it was the legal and patent domain that pushed me to finally build what I'd been thinking about for a long time. The concept was accepted for the European Patent Office's CodeFest 2026, where the challenge is automated patent and IP portfolio evaluation. I'm in the coding round now. And then this week, Anthropic's legal plugin triggered the very market crash that proves the broader point.

Automating semantic understanding doesn't replace expertise. It makes expertise portable for the first time.

EvaL Ink

I'm about to release and open-source EvaL Ink, an offline-first app for creating and editing evaluations locally. Runs on iOS, Android, and desktop. Where the Directory Server handles AI-driven workflows, Ink is the human-centric layer. Author instructions, review results, work with structured evaluation data without needing a browser or an internet connection. Author locally, sync when ready.

Deliberately the opposite of yet another cloud-only AI wrapper. Your evaluation data lives on your device first.

The Week That Made the Case

On Monday, Anthropic released 11 open-source plugins for Claude Cowork, including one for legal workflows. Contract review, NDA triage, compliance checks. Software stocks across the board lost $285 billion in market value in a single day.

At the same time, OpenClaw, an open-source AI agent framework, passed 160,000 GitHub stars. A research group at HKU rebuilt the core idea in 4,000 lines of Python and picked up 8,000 stars in four days.

Agent infrastructure is commoditizing fast. The plumbing is becoming free. What matters now is what flows through it. Structured evaluations with evidence, methodology you can trace and share.

Building a proprietary evaluation platform on top of today's hot model is building on sand. The only approach that holds up is an open standard nobody owns. MCP proved this for how AI connects to the world. EvaL is the same bet for how AI proves its work.

None of this will happen overnight. Thomson Reuters and RELX have deep data moats and enterprise contracts that don't evaporate in a week. But the direction is clear, and the companies that adapt will be the ones whose value lives in portable methodology, not locked-in software.

Encode your expert methodology in a portable, open format and you own it. Doesn't matter which AI provider or platform dominates next quarter.

I'm building this in the open because a standard only works if it belongs to everyone.

Explore the platform at eval.directory. Schema on GitHub.