Mycroft: Private AI Valuation, Price Panels, Stock Split Detection

Generating 5,479 per-share price marks from SEC filings

[0:00]Hi, I'm Om Mallie. This video is about the price panel, turning SEC filings into an actual per share price history for private eye companies, and catching the one thing that would have silently ruined it. The arithmetic is one line. Everything hard is in deciding which rows deserve to be in it. Here is the

The stock split trap: Why Perplexity appeared to collapse by 92

[0:16]Whole week, value divided by share count once per security per fund per quarter, gives 5,479 marks. That part is one line of arithmetic. The part that took the week is a detector for stock splits because a split makes a price look like a crash.

Fixing narrow tolerance windows: Moving to a relative 1% rule

[0:30]It found three. It could not tell me what any of them were. 5,806 filed holdings go in. 28 were rejected last month as belonging to nothing. 72 sit on filings a later amendment replaced. The remaining 5,706 aggregate into marks. It reconciles exactly and nothing gets dropped for being awkward. Here is why the detector

Disambiguating splits from repricings using share count constant values

[0:51]Matters. Perplexity went from $695 a share to 58. Read naively, that is a 92% collapse and it would have entered the price history as one. It is not a collapse. It is a stock split and a price series has to know the difference. My detector used a window of plus or minus 0.02 around whole numbers.

Auditing decision logs: Recording human rationale alongside machine metrics

[1:10]Perplexity's ratio is 11.93, 0.07 away from 12. The window misses it by a factor of three. It looked correct only because the other perplexity step landed on exactly 10. The rule is now a relative 1% window imported by the detector and the review queue. But catching a step is not deciding what it is. Three companies threw one and the ratio cannot tell you which is which. The share count can. Perplexity's preferred line went from 19,000 shares to 193,000 while the dollar value did not move by a cent. 10 times the shares, exactly the same money. Anthropic held the same 89,000 shares in all 13 quarters. Only the value moved and the step reverses the next quarter, which a split never does. SpaceX the same, share count constant, value exactly doubled. Splits move the share count. Repricings move the value. A person made all three calls and

Wrote the reason down. One of those is Suttler. 11.93 is not a split factor. It is a 10 for one split and a 16% markdown that landed in the same quarter. Divide by 11.93 and you erase a real price move. Divide by 10 and it survives. So, the panel records what a person decided to divide by separately from what the machine measured. The plan listed seven checks for this panel. Six pass. The seventh needs filings the SEC has not published yet, and it reports that rather than quietly passing, in one of the six expected four managers to agree on Entropic's price, 10 do in two clusters 4/1000s of a dollar apart. Week seven on one page, 5,479 marks value divided by share count every filed holding accounted for. A detector whose window was too narrow to catch its own test case, now a relative 1% rule. Three

Companies threw a suspected split and the ratio could not tell them apart. Share counts could, and a person made every call with the reason written down. Six of seven checks pass, and the seventh says why it cannot. Next week, how often these marks actually move. Your turn. Paste this into Claude, take a threshold in your own code, an absolute tolerance, a cutoff, a magic number. Find the case it was written for, then check whether it would still catch that case today. Mine would not, and it looked fine because a second case happened to land inside it. A threshold that has never been tested against the thing it exists to catch is just a guess with a number on it. Building the marks panel and the split detector, week seven of the private AI evaluation agent. Next week, how often these marks move. I'm Molly for Humanitarian's AI.