The internet recorded the answers. It never recorded the work.
Everything a model knows about being an investment banker, a systems engineer or a quantitative researcher, it learned from what those people published — not from what they did. That gap is the whole business.
Expertise is mostly undocumented
Ask a senior practitioner how they knew a number was wrong and you will usually get a shrug and a sentence that starts "you just get a feel for it." That feel is real, it is learned, and it is almost entirely absent from the training corpus — because nobody writes down the intermediate step, the approach they abandoned, or the reason the obvious answer was wrong.
Pre-training consumed what was written. What's left is what was done. It doesn't exist as a corpus anywhere, it can't be scraped, and it can only be produced by paying the people who hold it and then proving that what they produced is correct.
Both halves of that sentence are hard, and most of the industry only does the first one.
“…to pierce the obscurity of those minute processes.”
Tertius Lydgate is the young doctor at the centre of Middlemarch — a man who arrives in a provincial town convinced that medicine is full of received wisdom nobody has bothered to test, and that the answers are hiding in details too small for anyone to have recorded them.
Eliot is not sentimental about him. He is right about the method and he still fails, undone by the ordinary pressures of money, reputation and haste. That felt like the correct name for a company whose entire product is refusing to cut those particular corners.
The knowledge that makes an expert an expert lives in the minute processes. We record them, and then we check them.
What we hold to when it costs us
Anyone can claim a quality bar. These are the specific places ours is expensive, which is the only way to tell whether it's real.
- Verifiable or not shipped
- If we cannot state how a claim was checked, it does not go in the dataset. "An expert said so" is provenance, not verification.
- We eat the cost of rejection
- Clients aren't billed for work that fails review. It's the only pricing structure that doesn't quietly reward us for lowering the bar.
- Contributors are professionals
- Paid trials, rates quoted before acceptance, actionable rejections, and credentials rather than names on the record. The supply side is a relationship, not a funnel.
- Narrow beats broad
- We would rather be the obvious choice in two domains than a mediocre option in twelve. We turn down domains where we'd be starting a network from nothing and pretending otherwise.
- Say the uncomfortable thing early
- If data won't fix your problem, we'd rather lose the engagement at the scoping session than six months in.
Small, senior, remote
We're early, and we'd rather be straightforward about that than perform scale we don't have.
- Entity
- Lydgate LLC, a Delaware limited liability company.
- How we work
- Remote and distributed, with a deliberately small core team building the platform and running domain tracks. The network of practising experts is the part that scales.
- Where we focus
- Finance and markets; engineering, software and mathematics. Both chosen because correctness is contestable in them, which is exactly what makes verification worth building.
- Contact
- One inbox — support@lydgate.ai — read by the people doing the work.
Work with us, or write for us
Both conversations start the same way: tell us what you know how to do.