The Ledger Autopsy — could a detector have caught Enron?

Could a computer have caught Enron?

I ran a fraud detector over the real financial filings of history's biggest scams and a few honest companies, using only what was public before each one blew up. It caught three, flagged one honest company by mistake, and missed one entirely. Here is the whole scorecard.

Loading the case ledger…

Every company it scored

Worst score first. Hover for the story, click to open the autopsy.

How the autopsy works

The point of this project is not to claim the machine wins. It is to show, honestly, what the numbers could and could not say.

It only sees what was public at the time. Each company is scored on the last annual report filed before the fraud came out. No future knowledge in the inputs. The only hindsight is that I knew where to point it, and that bias is the honest catch below.

The score is math you can check, not a black box. Every point traces to a real figure in the filing: accruals, cash conversion, revenue growth, capital spending versus cash, receivables, leverage. Open any case and you see exactly which signals fired and by how much.

The numbers come straight from the filings. The healthy companies are pulled from SEC data. Enron and WorldCom are older than that system, so their figures are typed from the actual 10-K and every one links back to it. Wirecard is a foreign company shown with that gap stated out loud.

Claude did the reading, not the scoring. The score is plain arithmetic. Claude Code pulled and reconciled the filings, built the model, and wrote the short case memo from the footnotes, marked as the judgment call it is.

Then it is graded against what really happened. We know who was a fraud and who was not, so every verdict lands in a true or false, hit or miss box. That is the scorecard at the top.

The mistakes are on purpose. One honest company gets flagged, one real fraud slips through. Those two are not glitches in the demo. They are the whole point.

What this cannot tell you

Hindsight is baked into the question. I already knew which companies were frauds, so I knew where to point the model. Real-time, blind, across thousands of filers, the false-positive rate would be far higher than it looks here.

A flag is a question, not a verdict. Netflix is flagged and is a perfectly honest company. Aggressive-looking accounting and fraud can produce the same ratios. The tool raises a hand; it does not convict.

The best frauds are invisible to the filings. Wirecard's numbers looked pristine because the lie was cash that did not exist. You cannot catch a missing asset by reading the assets that were reported.

Small sample, curated. Nine companies is a teaching set, not a validated screen. The thresholds were chosen to be reasonable, not optimized, and the matrix would move on a bigger, blinder universe.

Filings lag and are self-reported. Annual reports arrive months after year-end and describe what management chose to disclose. This is a read of the paper trail, not the ground truth.