Abstract. Every other page on this site is an exhibit. This page is the claim they add up to, stated once.
1. The claim
Machines have made research ideas nearly free. What they have not made cheap is deciding which of those ideas to believe, and that decision, the gate, is now the scarce and valuable part of any research operation. This site is my gate, run in public for a season, with my failures kept and my instrument measured. I made the broad version of this argument about intelligence at large in Working Paper No. 33. Here I am making the same argument on myself.
2. My record
| What I counted, as of 2 September 2026 | Result |
|---|---|
| Ideas I filed / tested all the way to a verdict / that survived | 316 / 184 / 2 (1.1%) |
| Judged against my own standards / against an outside vendor’s simulator | 69 / 115 |
| How often my test catches a genuinely good strategy when one is really there (true Sharpe 0.8, 20 years) | 4 to 6%, published |
| Ideas from working practitioners I killed before opening any real research on them, Aug 2026 | 41, across four hunts |
| Forecasts I made and then scored against what actually happened | 620, Brier 0.231 |
| Public claims I checked against company filings | 134; 32 contradicted |
| Mistakes I found in my own notes, fixed and left visible | 28, across 21 notes |
| Papers I pulled down completely | 1, after five theses failed review |
| AI agent runs behind all of this | 958 at last count; statistics from code, always |
| Notes published, each kept under 500 words | 41, one frozen submission excepted |
3. One gate, many doors
The gate is one test. What changes is what I point it at. Each part of the site is that same standard aimed somewhere new.
| Part of the site | What I put through the test there |
|---|---|
| Tested Strategies | Factors from published research. I ran 158 of them and not one survived, plus everything I have tested since. |
| Research Notes | Claims other people are making in public. I checked 134 against the filings and about a quarter were contradicted. |
| Forecast Wing | My own predictions, locked in before the fact and scored as the results come in. |
| Trading Floor | Ten paper desks built from textbook signals, most of which my own gate rejected. The exhibit is the risk discipline, not the return. |
| The papers (No. 38, 39) | Prediction markets, and the government’s own forecasts. |
| No. 37 | The AI agents that run my desk, including where they went wrong. |
| Library | The shelf my critic has to read before any verdict of mine counts. |
Corrected 3 October 2026. The Trading Floor row once said the floor trades the two strategies that survived. It does not. Neither survivor is a return series, and the ten desks run market beta, low volatility, trend and sector momentum.
4. Why it matters now
Somewhere this year a fund will hand its research to a roster of sixty AI agents, and plenty more will follow. I think they will all hit the wall I hit: once ideas are cheap, the hard part is not having them, it is deciding which ones to believe. A test nobody has checked for accuracy is not a judge, it is decoration. What I do here is not clever. I fix the pass mark before I run the test, I let code do the arithmetic, I keep the part of the system that invents ideas away from the part that kills them, I publish how weak my test actually is, and I write the verdict down either way. None of that is my edge. I think it is about to be the price of being believed.
5. What this does not claim
I am not claiming the two survivors are special; they are modest, and I have not measured how much money they could carry. I am not claiming my gate is finished; I published its power curve precisely so you read my rejections as statements about where I set the bar. I am claiming one thing: every number above is checkable from this site, and that sentence is the product.
- Method
- Counts from the pages they appear on: No. 22 for the census, No. 35 for the fact-check, No. 37 for the agents, RN 5 for the forecasts, RN 8 for the instrument’s power.
- Honesty
- I have no new results on this page. It is an index with an argument.
- Related
- Everything on the site; start with No. 22, No. 33 and No. 37.