Most strategies are not killed; they are shelved indefinitely — "leave it, we'll look again later". The difference matters: a kill frees your attention and your capital; a shelving does not. Two tools here: what condition stops it, and writing down how it will die before it starts.
| No kill rule | With one |
|---|---|
| "a bit longer" | condition triggers → stop |
| losses arrive gradually; no moment ever feels like the moment | the moment was defined in advance |
| sunk cost accumulates, making it harder to stop | the reason to stop is independent of what you have spent |
| ends in indefinite shelving, holding capital and attention | ends archived, in the kill file |
The last row is the point. A kill is an output — it becomes a piece of knowledge ("this road is closed") that can be reused, published, and can save someone else the tuition. Shelving produces nothing; it defers the decision to a future you.
| Type | Trigger | Example |
|---|---|---|
| Capital | losses reach an absolute figure | stop ordering when the wallet falls below a floor |
| Futility | passing is now mathematically impossible | no remaining sample can turn it around |
| Criteria | the scheduled read misses a gate | any of the four gates failing |
The second is the valuable one, because it can fire far in advance and save a lot of both. The arithmetic is simple:
| Line | At the time | Remaining sample would need | Verdict |
|---|---|---|---|
| A | n≈460, excess −1.8pp, gate +4pp | ~40 observations at +71pp each | above the ceiling → stop |
| B | n=2,468, excess +1.56pp, gate 2.25pp | 32 at +55pp each | impossible → read |
A harder variant states it as a probability: one line needed 87 of its remaining 93 to turn around, which under randomness has probability 8.25 × 10⁻²⁰. That number is the verdict.
The tool I use most, and it is one sentence: before running, write down how this will look if it is wrong.
| Without | With |
|---|---|
| look for an explanation when it breaks | claim one from the list: "this is failure mode two" |
| an explanation always exists, and always sounds reasonable | the explanations were written first and cannot change |
| so you can always justify one more try | claiming one means archiving |
Because there are too many available: sample too small, unusual conditions, parameters untuned, execution issues, bad luck. Any of them might be true, and you will pick the one that lets you continue.
Pre-claiming inverts it: the list of explanations is sealed before you see the result. Something on the list happens → claim it, archive. Something off the list happens → that is genuine new information and worth studying.
It has a side effect too: writing the list often reveals you do not really believe in the line. If five failure modes come easily, reconsider starting.
An error of my own, worth publishing. Closing one line, I labelled the verdict "final kill, the sixth in its family". It sounds weighty — as though five prior kills were on record.
Checking later: that phrase appears in exactly one place in the entire archive — a single self-declaration at the end of a preregistration, written seven days before the read. Nowhere is there a list of kills one through five.
There genuinely are more than five documented kills, but that is a list reconstructable afterwards, not a count made at the time — so the verdict is rhetoric, not ledger.
When a claim carries a number — "the Nth time" — ask whether that N was counted or simply said.
This applies to reading anyone's results. A numbered claim looks audited, readers treat it as audited, and frequently it is not.
| Record | Why |
|---|---|
| What it was betting on | so a similar idea later can be matched against it |
| Criteria and readings | so a reader can judge whether the kill was sound |
| Which gate it died at | no edge? edge too small? eaten by execution? — three different diagnoses |
| What it cost | that is the price of the knowledge |
| What the ledger did not record | write "not recorded" — do not supply a plausible cause afterwards |
The last is a hard rule here. One line received $0 of rebates and the ledger did not record why — so the kill file says "no cause recorded" rather than offering a reasonable-sounding one.
A strategy without a kill rule does not fail; it gets shelved indefinitely — and shelving produces nothing. Write down how it will die before it runs, claim one from the list when it does, and the claim turns it into knowledge somebody else can use.