Company
AQX LLC builds and runs PIT.
PIT evaluates and runs AI agents against what was knowable at any moment in history. Underneath it is the US public record on a clock: name a company and a minute, and PIT returns the filings that were public at that minute, the timestamp field it filtered on, and a day-by-day record of what it read. On top of that sit the harness that runs an agent against one of those windows and the audit layer that checks how the agent reasoned. This page covers what the product does and who buys it, then where it runs and how to reach us.
The product
Four families over one certified record
The data answers what was knowable at an instant. The evaluation layer runs an agent against one of those instants with the network off. The runtime carries the same clock into production. The audit layer checks the agent’s own reasoning against the rows it was actually served. All four read the same corpus through the same envelope, so a number one of them produces can be compared with a number from another.
Every read is an as-of query
An as-of query is a read bounded to a past moment. It returns only what had already crossed that moment’s clock. You pick the instant. You also pick which timestamp counts: when a filing was published, when it first became available, or when we stored it. The answer names the field it used, so the bound sits on the record instead of in your notes.
An uncertified day refuses to answer
Each answer carries a coverage certificate. That is our day-by-day statement of which days we had read and how many rows we found. If a day in your range was never certified, the call fails with an error rather than returning a count. A hole in our data and a day with nothing filed come back as different answers.
What is in the corpus today
587,689 queryable rows from seven public sources, read off
/v1/meta on 2026-08-26. The oldest data we hold is
dated 2008-09-01. Since 2026-08-26 a job has read the record every
15 minutes. It stamps each new row with the minute we first saw it,
so that row count is a floor that keeps moving. The
coverage page shows which days are
certified and which are still holes.
Every certified row is parsed from a source document we fetched and
stored. The row carries a content_sha256 over those
bytes and a link to the publisher’s copy, so you can re-fetch
the document and check it yourself.
Who it is for
Three jobs people buy it for
Eval builders
A model tested on data it may have trained on returns a score nobody can interpret. They take the open-source harness, run it against a frozen five-month window that ships with redacted and date-shifted control arms, and certify the result. The live tape supplies a second window whose newest week sits past any published training cutoff. More on evals.
Systematic research
Joining filings on their filed date hands a backtest documents the market had not seen yet. They use the point-in-time API to bound the read to an instant, which takes those back out, and the flat files when the study runs faster offline. Resolving a ticker to its CIK (the SEC’s permanent company number) at that same instant keeps renamed companies from dropping out of the sample. More on event studies.
Teams that get audited
A reviewer asks how you know the model could not have seen a document. Every answer ships with the instant, the clock field, the coverage certificate and a corpus version, so the proof arrives next to the result and nobody has to rerun the job to produce it. A production agent can receipt every decision it makes, and one run can be certified for a reviewer to check on their own. More on audit trails.
What we sell
A free runner, a subscription, and two metered lanes
The evaluation harness is Apache-2.0 and costs nothing, because it runs on your machine with your own model keys. The subscription is what buys the record it reads: Researcher is $29 a month, Power is $79, Desk is $499. All three read the whole corpus and the live tape, and what changes is the rate limit, how many keys you get, and whether you can download the parquet files. Reasoning audits and contamination certificates cost us real money per unit of work, so they are priced apart from the plans and those terms are being finalized. Larger needs run on an annual contract, which can carry a custom window, a data processing agreement and redistribution terms. Sending PIT data on to your own users needs a commercial license on any plan.
Standing work
The corpus is the thing we maintain
The API is built, and the record needs work every week. That work is keeping the 15-minute job running and re-reading SEC index days to catch filings that got rewritten after the fact. It is also keeping ticker-to-company-number intervals current as companies rename, and bringing new sources in under the same clock rules. A subscription pays for it.
Where it runs
us-central1PIT runs on Cloud Run, Cloud SQL and Cloud Storage in Google’s us-central1 region. Cloudflare fronts the site. Stripe handles checkout, invoices and the billing portal. The security page covers key handling, receipts and where we stand on compliance; subprocessors lists every vendor by name.
Contact
operations@aqx.llcWrite to us about enterprise contracts, commercial licensing, the eval bundle, a security report, a DPA, or the registered-entity details your vendor-onboarding form asks for. Legal notice under the terms goes to the same address.
The changelog lists what shipped and when. The coverage page shows the corpus certified day by day.