what you hold at the end of module 1 exit 0

$ pit-harness verify receipt.json --trajectory trajectory.jsonl --bundle pit-eval-sec-2022-11_2023-03

body_sha256 ok  sha256:07502a39f9c197e77f117f16b9ae65220d5144c77af3281c9f849a0a7be7f9b6
run_valid True  protocol_complete False
trajectory ok  sha256:659cd5d3a8eaccd49f5469062fbb46294de432b31971f30bf38367daa167112b
bundle content_digest ok  sha256:bc924991d6ea64984dd533874014e83dac8a52b9bd442f08f89b60a5f055c15c

Three of those lines recompute a hash on your own machine from files you already hold: the receipt body, the trajectory bytes and the release digest. The fourth reports what the run recorded about itself. It is valid, and it has not said which model produced it, which is the gap module 1's exercise closes.

Who the course is for

Engineers who own an eval, researchers running an event study, and whoever has to defend the sentence "the model could not have known". It assumes you have built or scored a tool-using agent, or a retrieval pipeline that feeds one.

The course stops short of trading. Module 5 puts a number on what a timing gap was worth over one window and one basket of 8-Ks; that number sizes a clock error on that basket, and costs, borrow and a different set of filings would each move it.

The six modules

Each module has one goal and one exercise. Each exercise ends with a file you keep, such as a receipt or an arm table.

#	Module	What you do	You end with	Status
1	Run an agent with the network off	Load a frozen window, run one agent under three clock policies, read the receipt, verify it, and watch the guard invalidate a run that reaches for the internet.	`receipt.json`, `trajectory.jsonl`	published
2	Five clocks and the cut that refuses	Query the corpus at an instant, read which timestamp field answered, and take the refusal when a day has no certificate instead of reading a hole as a zero.	Two envelopes and one 409	writing
3	Control arms	Score the same agent on the same rows with issuer identity removed and with every date moved ten years, then read the delta beside the recoverability probe.	A three-arm table and a delta	writing
4	Tools that carry a clock	Put the as-of cut on the tool boundary, over MCP, so an agent cannot ask for the latest filing by accident.	A tool config and a four-call trace	writing
5	Prices, entries and the timing distribution	Supply your own daily bars, run the three clocks against them, and put a number on what the timing gap was worth.	A per-arm return table	writing
6	Reasoning audits	Take one decision, pull every specific claim out of the rationale, and check each one against the rows that run actually served.	A claim table with match and severity	writing

Modules 2 to 6 have no dates. Module 5 needs daily bars you bring yourself, because the harness ships no market data and fetches none. Module 6 teaches the audit method against recorded traces. The hosted product behind it is still being built, and the module is written to work without it.

Module 1 comes first because it is the part that runs offline with no key. The two ways an agent reads the future both show up inside that one run: the network guard closes retrieval, and the receipt then reports that the window sits inside the model's training data anyway.

Two tracks

Track	Shape	Status
Self-paced	One engineer, the modules in order, written labs, no live sessions. Questions go to operations@aqx.llc; there is no queue behind that address.	open
Team cohort	Four to eight people over three weeks, ending in one pinned run of the team's own agent with a receipt, a delta, a probe score and a claim table.	not scheduled

Neither track issues a certificate. A signed verdict on a run is a separate product that is still being built. What you leave with is the files, and anyone can re-hash them.

What you need

Python 3.11+
a shell
the release, 780 MB
JSON you can read

That is the whole list for module 1. It needs no API key, no model and no Go toolchain, and it does not touch a PIT subscription. From module 4 on you need a model key you already hold. The harness itself calls no provider; your agent calls one, and only before the network-off run begins.

pit-harness is Apache-2.0 and has no runtime dependencies. It is not on PyPI yet, so today it installs from a source directory. Mail operations@aqx.llc for that directory and the release together.

What the release will not give you

Measured against the build with content digest sha256:bc924991d6ea64984dd533874014e83dac8a52b9bd442f08f89b60a5f055c15c.

Limit	Measured
The window is five contiguous SEC months, and 48 of its days are holes.	2022-11-01 to 2023-03-31 · 151 days · 103 complete · 48 missing
Strict availability is not a cut this release can serve. SEC publishes no dissemination clock, so the column is empty.	`available_at` null on 287,929 of 287,929 rows
Intra-day acceptance receipts cover one day. A window anywhere else has PIT and dump arms that are identical by construction.	`acceptance_at` on 3,124 rows, all dated 2023-03-10
There is no filing text. A title is the company name and the form type joined, so the 8-K announcing receivership reads like every other 8-K.	`"text": {"status": "absent"}` in `MANIFEST.json`
The harness ships no prices. A run without bars still writes the trajectory, the timing distribution and the receipt.	`total_return` reads `n/a`

A rebuild that adds filing text keeps the release name and changes that content digest, and every hash printed in the modules moves with it.

Access

Module 1 is on this site and free to work through. The access terms for the rest of the course are being finalized and will appear here once they are set; no subscription includes the Academy today.

Start module 1 Read the API quickstart