Primary-document intelligence
Find out where they
What we look for
Four ways a company gives itself away.
The tools
Eight tools. One desk. Nothing re-typed.
One feed runs the whole bench: your numbers, your words, their comp set, their price, their contradictions, the room card, what landed today — and all of it again on Monday. Every output keeps the document it came from.
Pick one · see what it hands back
The obvious objection
“Why not just ask ChatGPT?”
Fair question. For a quick lookup, a model with web search is faster. Day91 is built for the two harder jobs — tracking one filer over time and comparing across filers — where the answer has to survive someone checking it.
| A model with web search | Day91 | |
|---|---|---|
| Where the number comes from | Whatever the search surfaced, blended with training data. | A filing, with the accession number. Graded C when the document hasn't been pulled. |
| How you check it | Re-run the prompt and compare answers. | Click the link. It opens the paragraph. |
| Which company it read | Whichever entity the name search returned. | The CIK, read from SEC's own bulk index. An identity, not a search result. |
| Tracking one filer over time | Measured accuracy drop of 18.6% on this task (Fin-RATE, Goldman Sachs and Yale, Feb 2026). | Fixed rules over filed figures, each with a document behind it. |
| Comparing across filers | Measured accuracy drop of 14.35% on this task, same benchmark. | The Grid shows only what was filed. Open cells carry the EDGAR route. |
| Breadth | The entire indexed web. | Eight companies read end to end. The ledger has the live count. |
Depth over breadth: a smaller set of companies, read all the way down to the document.
Who uses it
Four readers. The same eight tools. The routes cross.
Most readers live in two of them — and the second is usually the tool another team started with. Pick a reader to light their route.
The Diff → The Check
What comparable vendors repriced this week, then whether your own numbers still agree with each other.
The Check → The Mirror
The two metrics in your deck that contradict each other, then whether your positioning says anything theirs doesn't.
The Mirror → The Diff
The dead phrases your whole category shares, then who moves price behind them.
The Grid → The Reversal
A comp set in one view, then where the odd ones disagree with their own past filings.
What feeds them
One corpus underneath all eight.
10-Qs, 10-Ks, 8-Ks and S-1s off EDGAR, plus every covered vendor's live pricing page, diffed weekly. That is why a route can cross from one tool into another without losing its receipt.
- —
- Loading
Why it works this way
The hardest task for a model is the one worth doing.
The Fin-RATE benchmark tested language models on the three things analysts actually do: pull a fact from one filing, compare disclosures across companies, and track how one company's filings change over time. The last two are exactly what The Reversal and The Grid run on — with fixed rules over filed figures instead of a summary.
- 0
- accuracy models lose tracking one company's filings over time
- 0
- accuracy models lose comparing disclosures across companies
- 0
- of ten signals in Issue 003 shipped with the filing attached