Mark It to Market
Everything in finance is marked to market.
The book is marked. The fund is marked. The trader is marked, daily, to the cent, whether she likes it or not. The whole industry runs on one discipline: you do not get to say what something is worth. The market says, and you write it down.
The morning note is not marked. The price target is not marked. The high-conviction call, the strategist's year ahead, the dashboard that lit up red, the paragraph the model wrote at 6:40 this morning: none of it is marked. Nobody goes back. Nobody writes down what happened next to what was said. The industry that invented the scoreboard keeps an unmarked book, and the book is made of opinions.
We know why. Being quoted pays. Being right is hard to prove and easy to forget. The grader, when there is one, is the author. And the incentive runs one way: a call that is never scored can never be wrong.
Then the machines arrived and made the unmarked book infinite.
A plausible paragraph costs nothing. Every desk, every vendor, every intern with a prompt can produce research-shaped text at the speed of the model. A reader told us that his problem is no longer finding information; it is being buried in it, in the same confident format, every morning. He is right. Plausible is free. Graded is not. In a world of infinite fluent text, the only scarce thing left is a claim with a score attached.
What we believe
What we are doing about it
Three things.
We publish what it costs to produce intelligence we can grade: the completed answers, the retries and the failed work. We show where the accounting is complete and where it is not, how cost changes at comparable quality, and which model earns its place on which task. A lab that hides its cost curve is hiding half its results.
A refusal is a call too, and it can be graded. Guard Bench is a public test suite built from what banks, brokerages and exchanges ask before they let an agent speak: may it say this, to this user, from this source, with this citation, and can the run be traced afterwards? A system that refuses everything has not passed. We grade unnecessary refusals alongside unsupported answers and permission failures. We run our own harness through it and the frontier models with it, and we publish the table. Whoever fails, fails in public, including us.
Season one is the Q3 2026 earnings season. Before a release, our earnings desk records its assessment: the business developments we expect to matter to the market, the guidance lines at risk, the peers that will read across. Time-stamped. After the release, we grade the calls against what happened and against the naive baseline anyone could have used. We publish the original calls alongside their outcomes, with timestamps and subsequent changes preserved. We also grade more than the stock's direction: the business questions we identified, the results and guidance we anticipated, and how the market received them. Neutral calls, withheld forecasts and missing coverage stay visible. The hits, the misses and what we can establish about each miss, in the same table, in public, on a schedule. We publish the grading rules with a date and a version. If the method changes, the change is visible. Intelligence changes as evidence changes. We preserve the original view, show what changed, and identify the evidence behind the update. A later explanation does not replace an earlier forecast. If we are no better than the baseline, you will read it here first.
The kit that makes all three checkable is open by design. The trace schema, the eval format, the rubric and the Guard Bench suite go on GitHub under an Apache 2.0 license, so anyone can run the same tests on their own agent and tell us where ours is wrong. A desk is a taskset, a harness and a rubric. Ours are below. The next one can be yours.
Why a ten-person company is doing this
Because we build for the rooms where being wrong is expensive: trading floors, brokerages, exchanges. Those rooms do not want a smarter paragraph. They want a paragraph with a trace, a grade and a price, inside their walls, on their data, under their rules. We built the harness for them. The lab is where we prove in public that it works, and where we find out in public when it does not.
That sentence is the moat. Nobody else wants it.
Mark it to market.
Pulkit Jaiswal, Founder and CEO
Andrew Schaye, CPO and Head of Platform
AgentSmyth Lab · November 2026