Trusted Data
Operations For
The AI Era

TypeSafe released Jev this month, and it is a different kind of model. Where a language model writes you a paragraph, Jev takes a typed question and hands back a probability for each answer. We wanted to explore the technology, so we gave it a fun job: reading the news.

The result is Outliers, a small Belvedere Labs project that watches 88 news outlets for the stories they tell most differently. Jev reads each outlet's headline and teaser with the outlet's name removed and answers the same five questions, and the stories where the newsrooms disagree most rise to the top. The feed refreshes every few minutes.

What Jev Does With a Headline

Every headline and its teaser go to Jev with the same five questions.

• How alarmed is the wording, from reassuring to sensational?
• Does the headline state a fact, attribute a claim, ask a question, or offer an opinion?
• Who, if anyone, does it hold responsible?
• Does it promise more than its teaser delivers?
• Is the language loaded?

Because Jev answers with probabilities rather than prose, the rest is simple math. We add up how far apart the outlets sit on each question, give every story a divergence score, and post anything over 30. The grid at the top of this page shows the tone answers, one row per story and one column per outlet, and a brighter square is a more alarmed headline.

A Common Engineering Problem

Maybe you’ve been there: you find yourself given access to a giant, half-documented store of data, and you’re not entirely sure what you even have. Even worse, maybe you’ve never seen most of it before! Your unfortunate task is to find the needle in the haystack on a tight deadline. Connections are fragile, and you can’t risk altering any of the data, but you need answers fast. Where do you begin, and what can you realistically hope to achieve?

A tough situation to be sure, but one that is fairly common. Here at Clear Fracture we found ourselves running into this too. That’s why we developed Investigate, a new feature available in Belvedere. Investigate is a read-only research agent that answers questions about your data, grounded in your catalog.

How Investigate Can Help

To start digging into a task like this one, normally you at least need to have an idea of what databases and tables might be available, along with their schemas. Lineage is likely to play a role too, so that you can understand both upstream and downstream data flows around a particular asset.

With Investigate, you have the luxury of letting an agent figure those things out and document them for you. Instead of starting with SQL queries, you start with plain English: 

Uber built Apache Hudi to solve a familiar data problem: Records change after they land.

• A trip gets corrected after it ends.

• A chargeback arrives weeks later.

Rebuilding a large table to fix a handful of rows is wasteful.

Our free idempotent backfill skill applies some of Hudi's design ideas to late-arriving corrections in data you've already published. Give the current URL to your coding agent and simply tell it to install the skill to try it out.

What we borrowed from Hudi

Uber's lakehouse write-up describes a backfill that reads a fixed snapshot and overwrites only the affected partitions. It does not move the incremental writer's checkpoint. That separation prevents an old repair from disrupting the live pipeline.

Our skill turns a few of its core practices into a six-step, database-independent checklist:

1. Fix the source snapshot.

2. Name the partitions, business key, and newest-wins order.

3. Build each partition twice.

4. Compare the logical rows.

5. Stop for approval.

6. Replace and verify one partition at a time.

The skill is short on purpose so the agent can use the database's own atomic replacement operation.

Why build it twice

Maxime Beauchemin's essay on functional data engineering gives us a clear rule. Treat a partition as the complete output of a function. The same source snapshot and transform should return the same rows every time.