I lost 20.3% of a concentrated, unlevered semiconductor book in eighteen sessions. Almost none of it came from the core I'd held for months — it came from the buying I did on the way down.
I run a concentrated, unlevered position in AI-adjacent semiconductors. The concentration is deliberate — it's the thesis — and being unlevered means nobody can force me to sell at the bottom. I hold through drawdowns that shake most people out.
That part worked. Zero of my four written falsifiers printed; the thesis was as intact on the last red day as the first. If holding had been all I did, this would be a boring story about a mark-to-market. It isn't, because holding wasn't all I did.
Three weeks in, I traced which dollars lost the money — not which names fell. The answer reorganized the whole system.
This is not a stock-picking failure — every name I bought at the top is one I still want to own. It is a pacing failure, and almost no retail risk framework is built to catch it.
The uncomfortable part: no price-structure rule fires at an all-time high. Not a stop, not a trend filter, not a drawdown brake. Every one is silent — by construction — on the exact day you do the most damage.
The only rule that catches it is a cap on how fast you're allowed to buy. Nothing else even looks at the failure mode.
The old rulebook had thirty-five rules, because every time something went wrong I appended a patch. Three of its levels were stale, and on the worst day of the drawdown it told a panicked reader to "deploy into strength."
The replacement has twelve rules. These are the eight decisions I'd carry into any version, for any book:
Every lot carries one of two labels: conviction or trade. Trade lots have mechanical stops with no override, ever. Conviction lots get judgment.
The catch: conviction status requires a dated note naming what would make you exit. No written falsifiers, no conviction label — regardless of feel. Demotion is free.
Origin A six-figure position acquired over eight sessions received conviction treatment purely by drift. Nobody ever decided it was conviction. It just never got a stop.
Four things can make me wrong, each written down in advance with its pre-agreed action. Leaving on a tripped falsifier is the plan working, green or red. The three valves — a deep close well below a one-way pin, a lone name breaking on verified news while its peers hold, and a systemic tape override — are the only ways price alone forces a sale.
What does not exit the core: red days, flush days, a 10% down close, a scary alert, or the urge itself. A mark-to-market is not evidence.
No more than 8% of the book bought in a session, no more than 15% in any rolling five. All buying counts — including tickets the system itself generated.
The rule that would have prevented most of the loss — and the hardest to accept, because on the day it binds you'll have an excellent reason to override it. That's the point. It's machine-checked; a breach locks the buy gates for the next session.
The reference price each stop measures against can be raised, never quietly lowered. Lowering one deliberately requires an explicit flag and a dated note.
If you take one thing from this page, take this. One-way pins make a stop un-gameable from inside a drawdown — the only time anyone ever wants to game one.
The full sequence, assembled once: the pin alerts. In the shallow zone just below it, judgment holds or folds. Ten percent lower, the valve mechanically exits the flex lots — everything except the written core — two consecutive closes, no debate.
Origin I lowered pins repeatedly across three weeks as the book fell. Each individual adjustment was defensible. Together they silenced the monitor for the six sessions immediately preceding the worst leg of the rout.
When the book crosses a weakness rung, the buy ticket latches: its composition fixes at that print. But it doesn't deploy until the sector confirms — two consecutive closes above a trend line.
The first version deployed the next morning, straight into the knife. Latching commits you at the discount; the confirm stops you catching the falling half. The cost is a slightly worse fill. Worth it.
When two rules conflict — and in a real drawdown they conflict constantly — one written ordering decides. Lower number wins. No exceptions clause anywhere else.
Ambiguity in a rulebook doesn't produce caution. It produces whichever action you already wanted to take.
The rulebook contains no dollar levels, no positions, no current posture — only percentages and conditions. Every live number is machine-written into a separate state file, fresh each day.
Hand-maintained numbers go stale silently — and fastest in exactly the conditions that make them load-bearing.
Hard cap. Amendments rewrite the rule in place plus one dated changelog line — never an appended patch.
Thirty-five rules is functionally zero rules — nobody can hold them at the moment they're needed. Mine now fits on a wallet card. If it doesn't fit on a card, it won't survive a bad morning.
Every rule has a price, written next to the rule it belongs to. A rulebook that only lists benefits is marketing — you'll abandon it the first time it hurts.
| Rule | What it costs | Measured |
|---|---|---|
| De-gross ladder | Gives up compounding in ordinary years to buy 4–5 points of max drawdown. Pays for itself only in real bears — 8% vs 34% through the COVID window. | ≈2%/yr CAGR |
| Froth gate | Blocks adds when the cluster is extended — on days that historically kept running. Feeling wrong while it's on is the expected experience of tail insurance. | +13% vs +2% |
| Deep valve | Two consecutive closes ten percent below a pin force out the flex lots — everything except the written core. A sharp V-shaped recovery round-trips a flex lot for nothing. That's the premium; the core rides through every V. | 1 round trip / V |
| Systematic covered calls | Tested and rejected. 20%-out-of-the-money three-month calls breached about 60% of the time and lost money per window. One narrow pocket survived: one month, at a volatility extreme, never on the conviction name. | (4.3)%/window |
Alerts that are never actionable train you to ignore alerts that are. On one day, eleven intraday stop alerts fired; zero were actionable. Now core decisions are close-basis only, and intraday alerts are labelled attention-only. An alert you've learned to swipe away costs you the reflex.
Hedge the factor you actually own. I held index puts against a sector book; over the drawdown the index fell 3.6% while the book fell 16.6%. Every roll is now a full re-underwrite — "same structure, further out" is how a wrong hedge perpetuates itself.
The system above runs as code — nightly evaluation, machine-checked gates, an auditable changelog. It took a year and a bad drawdown to build, for exactly one book: mine. I don't know whether that's a product or a personal obsession, and I'd rather find out by asking than by building first.
Do you have written rules for when you sell — and did you write them before or after you needed them?
If you have a real book and no written rules, I want to hear how you decide. If you do have written rules, I want it even more — what do yours catch that mine don't?
Or email directly: rules@mcsio.com
Not investment advice — one person's rules for one portfolio, published because the mistakes seem general even though the positions aren't. Names, position sizes, and account values left out on purpose.