Most of it died. That's the discipline.
92 techniques opened; 41 (45%) falsified and closed, 28 of them with no reopen condition written down; 19 still registered as active or candidate work. This page is the census and the method — the rate at which ideas die here, and what kills them. It is deliberately not a list of what the ideas were.
Outcome funnel
What kills things
Every closed technique is filed under the reason it failed. The distribution is the useful part: it says which failure modes this program actually runs into, and a reason is a reopen condition in disguise — a technique killed on cost comes back if costs move, one killed on a null result does not.
- No detectable effect / wrong sign 22
- Real mechanism, costs eat it 7
- Signal too late / unselective 4
- Market already efficient 3
- Decayed or non-persistent 2
- Redundant or capacity-bound 2
- Data artifact 1
What a technique has to survive
The same gate for every idea, applied in the same order, decided in advance.
-
01
Preregister
Hypothesis, gate thresholds, and kill criteria are written down and frozen before the data is allowed to vote. A result that would not have been accepted in advance is not accepted afterwards.
-
02
Test against real costs
Spread, fees, tax, and latency are in the backtest, not in a footnote. Gross-only edges are not edges, and most of what dies here dies at this step.
-
03
Kill in public
41 of 92 techniques have been falsified and closed, each with the reason banked and the condition that would reopen it written down. The graveyard is the asset — it is what stops the same idea being re-tried every year.
-
04
Monitor live
Survivors run under decay monitors, exposure breakers, and a daily briefing. Promotion is never permanent; a live technique that stops clearing its gate is de-grossed without a debate.
In this hub
The work itself — what was tried, what survived, and the evidence for both.
-
Sandbox
Two in-browser sandboxes on synthetic data — mechanics you can run.
-
Backtest live service
Cross-sectional engine driven from the browser against the live service.
-
Agents
AI research agents as a typed operator kind: tools, refusals, gates.
-
Writing
Short notes, each citing the run that produced it.
-
Research record
Frozen preregistrations and verdicts with reproducible verification hooks.