Warren works on what you hand him

At 7:20 on a Thursday in January, an analyst I speak to every few weeks had three things open that no wire will ever carry. A 34-page broker note that landed overnight. A Telegram message from a former colleague, four paragraphs long, claiming a supplier had quietly stopped shipping to one of his names. And a release from a private company that files nothing and publishes to its own website in a font nobody chose on purpose.

He covers industrials at a long/short fund running a bit over $600m. By his own count, roughly a third of what he reads in a normal week arrives as a forward from a person rather than as a row from a source.

None of that gets scored. The feed only sees what publishes to something we ingest, so his morning splits cleanly in two: the tape, which is triaged for him before he opens the tab, and the inbox, which is ninety unranked minutes with a coffee going cold next to it.

Warren is for the second half.

Same engine, one flag set differently

Paste the four paragraphs into the chat and ask whether it's tradeable. He calls score_document, which posts to the same Verdict Engine that scores every row in the feed, as an artifact of type ugc. What comes back is the card you already know from the drawer: actionability 0 to 100, the five sub-scores rated 0 to 10 each (novelty, materiality, surprise, specificity, directness), the reasoning paragraph, and where there's a direction, short and long from −5 to +5 with conviction on its own 0 to 100 track.

Novelty still carries 0.30. The gate is still 60. There is no second scorer and no chat-flavoured variant of the engine, so a document Warren rates 71 and a feed row badged 71 mean the identical thing, which is the only reason it's safe to put a number in front of someone in a chat window at all.

One option differs from the feed, and it's worth knowing about. On the live feed sentiment runs only if actionability clears the gate, so a row at 44 has no direction attached and the key is simply absent. Warren passes score_sentiment: "always". Hand him something that scores 44 and he'll still give you the short and long reads, with the 44 sitting above them telling you how much to discount. That inversion is deliberate. On a feed of ninety-one rows a confident direction on a recap costs you attention you can't get back; in a chat you asked about one specific document, and you're entitled to see the direction the engine read even when the item doesn't clear the bar.

The tool takes an optional ticker. That's the part I'd have missed if I hadn't watched him use it: he scored the supplier paragraph twice, once against the name he's long and once against its largest customer, and got two different actionability reads off the same four paragraphs because directness is a property of the pairing, not of the text. Second-order relevance falls out of running it twice.

Paste a URL instead and he reads the page before he scores it. Search first if the question is time-sensitive, and the sources arrive as chips under the answer rather than as claims you have to take on faith.

That ceiling is a real constraint and not a theoretical one. Ask him to go find the company's last two releases, compare them, and score the newer one, and the chain runs out of room before the writing starts. Sixty seconds is the other wall. Keep a turn to one document and one job and it never comes up.

Ask what came through this morning and he'll tell you he can't

He has no read access to your stream. That's a constraint designed in rather than left to chance, so the answer to "what hit my energy watchlist overnight" is him saying he can't see it, which I prefer to a confident paragraph assembled from a web search.

It's the single most requested thing about him and I've stopped treating that as evidence I'm wrong. Your watchlists are saved filters that encode what you care about, and a chat window is a bad container for the answer: what you actually want when you ask that question is a ranking, and you already have one, sorted and badged, one tab away. The version worth building is the reverse of what people ask for, which is a row in the feed you can send into a chat with its verdict already attached.

"I paste things at him I'd be embarrassed to send to an analyst. Half of them come back 30-something and I close the tab and I've lost four minutes instead of forty."

Analyst, long/short industrials, ~$600m

Two plans, two different gates

This one generates email, so here it is plainly. Chat is on Starter, at $49 a month or $39 billed annually, and that includes the web search and the URL reading. Document scoring is not on Starter. It starts on Pro at 100 a day and Quant at 1,000. On Free there's no Warren at all, which is consistent with Free having no scores anywhere.

Uploads cap at 20 MB, and he takes PDFs, plain text, markdown, and images. Drop in a chart screenshot with no text on it and he'll describe what he sees and decline to score it, because there's nothing for the engine to read. Prices he finds by searching lag the tape and he says so. And here's one that's mine to fix: if you switch conversations while an answer is still streaming, the turn is discarded rather than saved half-finished. Correct behaviour, genuinely annoying, and the third time it happens to you it feels like a bug.

Conversations are kept, titled automatically in three to six words, and listed under All chats by last activity. Rename them, delete them, they're yours and they train nothing. What there isn't is a way to search across them, and in March the analyst is going to want the thing he scored in January, with the only handle on it being six words generated in half a second while he'd already moved on. That's the next piece of work, and it's less about chat than it is about the fact that a verdict on a forwarded document is a research artefact, and research artefacts need somewhere to live.