Andrew Grosman

Growth Marketer & Digital Creator

  • About
  • Contact
  • AI Trading Agent Experiment
  • AI Implementation
  • Tools
    • Email Authentication Checker
    • Paid Social Campaign Calculator

The $57 Million Receipt Nobody Could Read

November 9, 2026 By Andrew Leave a Comment

A few years back, the company that made my space heater sent out a recall notice: stop using it, it might catch fire, and here’s forty dollars for your trouble, just mail in the original receipt. I’d bought that heater at Shopko. Shopko had been dead for three years by the time the recall letter showed up. The forty dollars, as far as I’m concerned, is still out there somewhere, filed under a store that doesn’t exist, from a purchase nobody can prove happened.

The trading agent had almost the same problem this month, except the receipt was worth $57.3 million, and the merchant, the Securities and Exchange Commission, is very much still in business. It had just, apparently, locked the door.

Quick context for anyone just joining: GameStop (yes, that GameStop) is one of the five positions in this account, bought back in September on an insider cluster; Ryan Cohen, the chairman, buying alongside two directors. In early October, a secondary data source reported that Cohen had gone back for more; roughly $57.3 million more, 2,600,680 shares, purchased October 1st and 2nd. The agent’s own rules require that kind of claim get checked against the actual SEC filing before it counts for anything. Not a summary of the filing. Not a news article about the filing. The filing.

So it tried. And then it tried again. And then, a few weeks later, it tried a third way, because by then it had become something of a personal crusade.

Attempt one: direct connection to data.sec.gov. Rejected, 403, the proxy didn’t like it. Attempt two: the SEC’s own browse-filings page. Blocked before it even got there; not on the approved list of websites the agent’s allowed to visit, like a bouncer checking a clipboard that doesn’t have your name on it. Attempt three, weeks later, using Robinhood’s own built-in SEC filing tool, the one that’s supposed to exist precisely for this: it found the filing, confirmed it was real, handed back a filing ID, and then, when actually asked to read the thing, returned a 404. Filing content unavailable. It knew the receipt existed. It would not let anyone look at it.

Three sessions. Three different doors into the same basement. Locked three different ways.

I’ve made an AI try to read a webpage for me more times than I’d like to admit; usually it’s something low-stakes, like a recipe hidden behind nineteen paragraphs of someone’s trip to Tuscany, or a parking garage’s hourly rate buried in a PDF from 2019. This was the first time it mattered whether the AI could actually get in the door, and the door would not open.

So I called in reinforcements. I have a separate set of bots running in Cursor, built on Grok, mostly for things that have nothing to do with trading; one does my calendar, one builds websites, one is apparently a job-search scout I keep meaning to put to better use. I sent the blocked task to Chief, the orchestrator of that crew, who handed it to Chef, the one who actually goes and does things. Different network, different blockers, and this time, it got in.

Chef came back with the real filings. Good news: the thesis held up; Cohen really had been buying, heavily. Less expected news: Chief, reviewing Chef’s own write-up before sending it back to me, caught a mistake in it. Chef had attributed the $57.3 million to the wrong set of filings. Chief corrected it, and sent the fixed version over before any of it got written into the permanent record.

Which means the whole chain, in order, was: a secondary source got it wrong, this agent couldn’t check it for five weeks because the actual government website would not let it in, a second AI system succeeded where the first couldn’t, and then that second system’s own supervisor caught a mistake the first attempt made on the way out the door. Four layers of checking, and the error still made it to the third one before anybody caught it.

Here’s the part that actually surprised me. After all of that, the dollar figure was right the whole time. $57,333,418.60. 2,600,680 shares. Exactly as originally reported. The thing that was wrong was never the amount; it was the calendar. Everyone had it landing on October 1st and 2nd, one tidy little weekend of buying. What actually happened is Cohen bought in three separate filings spread across three weeks; September 10th, September 21st, and September 29th, and the October dates people thought they were looking at belonged to an entirely different, later wave of buying that just happened to rhyme with the first one. Like finding out your uncle didn’t buy one lottery ticket on his birthday, he bought three, on three different trips to the gas station, and someone just smashed them into one story because it was a better story.

Ryan Cohen’s actual GME buying, verified via EDGAR Dollars spent, by the date he actually filed it $46.77M Sept 10 + 21 (two filings) $10.56M Sept 29 $17.08M Oct 2 $22.77M Oct 13 “Oct 1-2” is really just this one The actual bulk, three weeks earlier

Source: SEC EDGAR Form 4 filings, verified directly, not from a secondary aggregator’s summary.

So what do you do with that. The lesson isn’t “don’t trust secondary sources,” everyone already knows that one, it’s wallpaper at this point. The lesson is that even when you escalate to a second AI system specifically to get a cleaner answer, you haven’t bought yourself out of the job of checking; you’ve just added a second place where the checking might fail, and in this case it very nearly did. The record’s correct now. The thesis didn’t change. GameStop stays held, not added to, exactly like it was before any of this started.

While all of that was sorting itself out, the same account quietly notched something else: GEHC crossed the twenty-five percent take-profit line and got sold in full, the first position in six tries that actually did what it was supposed to do instead of getting stopped out early. Nobody needed three doors and two AI systems to confirm that one. It just worked.

I still don’t have the forty dollars for the space heater. I have given up on that one. The SEC filing, eventually, after three doors and two AI systems, I got.

If this is your first stop in this series, the one about almost inheriting someone else’s trading account is a good primer on how much blind trust an AI agent can end up operating on, and how little it takes for that to go wrong.

Filed Under: AI Trading Agent Experiment

Leave a Reply Cancel reply

Your email address will not be published. Required fields are marked *

Copyright © 2026 ยท AndrewGrosman.com