One source of truth. It's the filing.
Standardized fundamentals for every US public company, each figure carrying the accession number, the period and the raw XBRL tag it was read from. Spread the quarter without retyping it, and settle “where did this number come from?” in one click.
Works in your browser, from Python, or inside the AI you already use.
POST · Post Holdings, Inc. · fiscal Q3 2021
Income (Loss) from Continuing Operations before Equity Method Investments, Income Taxes, Noncontrolling Interest·Q3 2021
Major restatementwas-$4.2M
Revised down in a 10-Q.
- Standardized facts
- 120.5M+Standardized facts
- US companies, delisted included
- 19,000+US companies, delisted included
- From any figure to its filing
- 1 clickFrom any figure to its filing
- Every filed vintage
- 1993–presentEvery filed vintage
We did not make these up. People said them.
Twenty names, four prints a quarter, and a model that has to survive the review. The grind is not the analysis — it is everything that happens before the analysis starts.
Two vendors, two answers, no tiebreaker
No single source of truth.
That is a paying customer describing a market-leading terminal. We do not ask you to trust our copy of a number: every figure resolves to the filing it was read from, so the tiebreaker is the document, not the vendor.
A restatement quietly breaks the historicals
Suddenly, I have two periods of data to enter into my painstakingly crafted model.
We keep every vintage a company has filed, so a revision shows up as a before-and-after on the same tag with both filings linked — found before it breaks the model, not after.
The quarter starts with retyping
I copy last quarter's notes into a new template and highlight the numbers to be updated.
The same standardized concepts come back in the same shape every quarter, so the print is a refresh rather than a transcription. The raw tag rides along, so you can see exactly what was mapped.
And you still have to check it
It's a lead, never a source.
Correct — and we are not going to pretend otherwise. You will verify anything that moves a recommendation. The claim here is narrower and checkable: verification costs one click instead of one hour, because the identifier on the figure opens the document.
Same work. Different starting line.
- Hunt the figure across a PDF, a vendor screen and a footnote.Pull the standardized concept, with its raw tag beside it.
- Find out about a restatement when your model stops tying.See which figures a later filing changed, and when.
- “Where did this number come from?” costs an afternoon.The identifier opens the filing it was read from.
- The AI writes a number you cannot place.Tools return filed values; the model arranges the words.
One token. Every surface.
A single Stripe-issued key unlocks all of it at your tier. There is no per-channel billing and nothing to integrate twice.
Workspace
Ask for the quarter in plain English and get a cited draft you can edit. Figures from the tools stay bound to the filing they came from, so the review is a click rather than a re-derivation.
MCP Server
Point the AI you already use at the archive. Its tools return filed values, so the numbers in your chat window are retrieved, not composed.
Python SDK
Pull a whole comp set into pandas in one call when the browser is the wrong shape for the job.
Numbers we are willing to show our working for.
Two analysts spent roughly 80 hours pulling seven data points from 500 S&P 500 filings by hand. Reading the same figures from structured XBRL took one person-day.
Across the S&P 500, 497 of 500 companies have had at least one figure changed by a later filing. Twenty-three have ever filed an 8-K Item 4.02.
94.5% of the revisions we detect arrived inside a routine 10-Q or 10-K — no amendment, no Item 4.02.
The reasons people say no.
How is this different from a terminal I already pay for?
It is narrower and it shows its work. We do not have news, chat, or execution. What we have is every vintage of every US filing since 1993 with the provenance attached, which is the part a terminal treats as a lookup rather than a record. Most analysts who use us keep the terminal.
My compliance team will ask what the AI did.
There is a record. Every agent action lands in an append-only ledger with the figures it touched and, for anything irreversible, who approved it. You can export the whole chain — prompt, tools, figures, filings, approvals, output — as a single Auditable Research File.
What happens when your standardization is wrong?
You will see it, because the raw tag is on screen next to the standardized concept. You can also override a citation for your own account, and the override travels with your work. We publish the consistency baseline rather than claiming a quality level.
Is my research used to train a model?
Not by us. We never train on your content, on your own API key or on our managed lane. On your own key your prompts go to your model provider under your agreement with them; the key is never stored in plaintext, and you can choose not to vault it at all.
Asked and answered.
Do I still have to check the numbers?
Yes, and we designed for that rather than around it. Every figure our tools return carries an identifier that resolves to the filing, the form type, the period and the exact XBRL tag it was read from, so checking is opening one link. What we remove is the hour of hunting for which document a number came from — not your judgement about whether it is the right number.
Are these as-reported or restated figures?
Both, and you choose. The archive keeps every vintage a company has ever filed, so you can read the value as it stood on any date, or the value as it stands today, and see the difference. That is what point-in-time means here — not a label, a second copy of the number.
How do I find out a company restated something I already modelled?
Restatement Radar lists every figure a later filing materially changed since 1993, as a before-and-after on the same raw XBRL tag, with both filings linked to sec.gov. It also tells you whether anyone was told: whether an amended filing or an 8-K Item 4.02 accompanied the change, or whether it arrived inside a routine 10-Q.
Where does the data come from?
SEC EDGAR, and nowhere else for fundamentals. We read the XBRL a company filed and map it onto standardized concepts; the raw tag travels with the value so you can always see what we mapped. Nothing is estimated, smoothed, or filled in from a vendor consensus.
How accurate is the standardization?
We publish the number instead of asserting quality. Across 48 catalogued accounting identities, 93.55% of facts are internally consistent overall and 88.96% in the modern era; the baseline is published in the open-source repository and CI fails on a drift greater than one point. Unmapped tags fall to an explicit 'Other' bucket rather than being silently dropped.
Can I use this in published research?
Yes, with attribution, for insubstantial excerpts in your own research on any paid tier. Redistribution — putting our figures into a product, a feed or a note you sell beyond insubstantial excerpts — needs a Commercial licence; contact us through Enterprise.
Start on one ticker you already know well. It is the fastest way to catch us being wrong.
Recommended plan for this desk: Pro.