agentic-trading · kronos desk · public research record

92 ideas. 9 survived.

This is the whole notebook of a quantitative research desk — every hypothesis it filed, the number that killed it, and the idea that got generated from the wreckage. Nothing is hidden: the failures are the point, because a desk that only publishes winners is publishing selection.

Click anything. Every chart is a filter.

Filters
01

The gates an idea has to clear

These are not stages in a pipeline — they are independent standards, each one a way a comfortable result can turn out to be nothing. An idea can freeze a kill number and never read real data; it can read real data and never be charged for its capital.

Standards cleared

counted across all 92 designs — this panel ignores the filters

Bars count cells that cleared each standard, out of 92. The steepest drop is economic, not statistical: 24 designs subtracted trading costs, but only 7 ever measured return on the capital actually immobilised, and only 4 computed a dollar capacity ceiling. Most quant research dies of arithmetic it never did.

02

How the ideas were generated

Each dot is one design, placed on the day it closed and the programme it belonged to. Read left to right and you can watch a kill spawn its own follow-up: the carry trade passes, then four successive designs are written specifically to attack it, and the last one wins.

Killed Survived Inconclusive / descriptive Designed, never run

Hover a dot for the design and its verdict; click to open it. Programmes are lettered in the order they opened. The density from 08-14 onward is an hourly research loop running unattended.

The pattern worth showing someone. Look at programme C. D-B1a asks whether the spot–perpetual gap mean-reverts — dead. D-B1b asks whether just holding the gap collects funding — it pays 8%/yr. Then D-B1R, D-B1c M0 and D-B1c M1 are written to destroy that result: overlay a regime filter, charge it for its collateral, then force it to survive margin top-ups. Two of the three succeed in killing it. That is the engine — a result is not a finding until the desk has tried hardest to break it.
03

What actually stopped them

Not "it didn't work" — the specific lever that would have to move. Click a bar to filter the whole page to those designs.

Binding constraint

what the design ran into

Verdict

the desk's own outcome classes

Signal-bound means the effect was not there. Data-bound means the desk could not get the bytes to find out — a purchasing decision, not a research one. Never run means the design was written, priced, and refused before a single byte was fetched, which is a legitimate outcome and is recorded as one.

04

What has been tried, and what has not

A list of results can only tell you what happened. A matrix tells you what was never attempted — every empty cell here is a question nobody on this desk has asked yet.

Method family against market. Darker means more designs. Click any cell to filter. Only the 66 cells carrying the full 21-flag taxonomy appear here.

05

The full record

Every design, its question, and the number that decided it. Click a row for the evidence.

Design Question Programme Closed Stopped by Verdict

06

The nine that survived

A pass on this desk is not a green light to trade. It means the design cleared the kill number it set for itself before the result existed — and every one of these carries a written limit on what it licenses.