$155.6m of cash generating $292,644 of interest income. How do you check whether reported cash is real?

Someone posted here this week about footnote and MD&A parsing being painful, and there's another thread going on about where the line sits between AI slop and AI-assisted research. Here's a case that sits on both, and the check in the middle of it takes about ten seconds and no AI support at all.

China-Biotics was a probiotics maker in Shanghai that got onto the US market through a reverse merger. Its 2010 annual report showed **155.6 million** dollars of cash, **292,644 dollars** of interest income, 81.4 million of sales and 15.6 million of profit.

That first pair is the whole thing. Real money in a real bank earns interest. Park 155 million at even a dull deposit rate and something like three million a year should come back. They reported under three hundred thousand. So either the company was leaving a fortune idle for no reason, which businesses don't do, or the fortune wasn't there.

Two other things fall out once you look. The cash had grown 84.8 million in a single year while the company earned 15.6 million, and nothing in the accounts explains where the rest came from. And the pile was close to twice annual sales, which is a strange amount of money for a business that size to be sitting on.

In June 2011 the auditor resigned. The 8-K disclosing it is still on EDGAR, accession 0001144204-11-037217, and you can read it in a minute. It says the audit team was pointed to a suspected fake website for the bank, that the paperwork supporting the interest income had arithmetic errors management put down to the bank's own mistakes, and that the rate on it didn't match the rate the central bank had published for that period.

I want to be careful with the words here. The SEC never charged the company with fraud. It cancelled the registration in 2013 after the company stopped filing. The fraud claims came from investor suits and from the auditor's letter, not from a verdict. For anyone holding the shares the outcome was the same either way.

Since the other thread is running, here is the **AI part**, because I think the way people set this up is most of what makes the output slop.

I don't hand the whole job to one model. I run five, each with a single task, and none of them can see what the others found. One asks whether the sales are real. One checks whether the per-customer numbers make sense. One asks whether the reported profit is backed by cash actually arriving. One does nothing but hunt for two numbers in the same filing that cannot both be true. The fifth looks at who controls the company and whether they can be trusted. They have to be five separate jobs. Ask one model to do all five and you just get its single view of the company, five times over.

For this filing I used the four that read the accounts and left the ownership one out, since there was nothing in the numbers it would have helped with. I also stripped the name, the country and the years off the data before handing it over. A model that recognises a famous fraud has not proved anything. It is just remembering(from its training data).

Now the part I would actually argue in that thread. The instinct is to trust whatever most of the agents agree on, and that instinct is what ruins the output. Point any decent model at a set of accounts and ask what is wrong, and it will hand you three or four genuine problems. Cash collection is weak. There is a related-party balance. The receivables are ageing. All real, and all survivable. The one finding that means the revenue itself is not real turns up in a single report, sitting right next to those. So if you rank by how many agents raised something, the fatal one loses to the safe ones every time, for the simple reason that only one agent found it.

The last step therefore does not count votes. It takes each finding on its own and asks one question: if this is true, how much does it actually hurt the company? A finding that sounds dramatic but only means one balance needs checking ranks below a dull-sounding one that puts the entire revenue line in doubt.

On this filing that did not end up mattering, because all four landed in the same place. What I found more interesting was what they asked for. Every one of them wanted the same document to settle it, and it was not more of the accounts. It was confirmation of the cash from the bank itself, handed over by the bank rather than passed through the company. A year after that annual report, that is the exact document the auditor could not get, and it resigned.

Does anyone here run something like this as a research experiment, or is it only useful looking back at companies that already blew up? I don't know how often it flags honest companies that just keep their money in accounts paying nothing.

No position. The company hasn't existed for over a decade.