Liquidation at 26%. The form filled the stop at 29%.

A grid system on our platform was liquidated in September 2026. Its stop loss was set to 29%. Liquidation arrived at 26.3%.

Nobody typed 29 by hand. The form suggested it, from a rule that is entirely correct: a stop must sit past the end of the ladder, otherwise it cuts a grid that is still filling. The ladder spanned 25.8%, so the stop belonged at 29%.

A second rule, in a different part of the same product, is also entirely correct: a stop must sit before liquidation, otherwise it can never execute.

Both rules were right. Nobody had ever computed their intersection — and for this configuration it was empty.

What we measured

An instrument (probe:stop-window) that computes nothing of its own. Every number comes from the owner the product itself asks: the ladder span from our grid math, the liquidation distance from our liquidation geometry, effective leverage from the risk gate, and the verdict from the risk gate itself rather than a retelling of its conditions.

Three axes without a database, plus one against the live fleet on 19 September 2026.

Finding 1: a grid stop is not a free number

It is bounded on both sides:

• •Floor — the end of the ladder plus 2 pp. A stop inside the ladder fires while the grid is still averaging down, which is self-destruction, not protection.
• •Ceiling — the liquidation distance minus 1 pp. Beyond that the exchange closes the position first, and the stop is a number in a field.

Between them is the window. Configure inside it and the stop means something. Configure outside and it does not — in either direction.

For an isolated position, liquidation distance depends only on leverage: roughly (1/L − MMR) × 100 percent of adverse price move, where MMR is the maintenance margin rate (0.5% at tier 1). At 5× that is about 19.5%; at 10× about 9.5%.

The floor, meanwhile, depends on the geometry of the grid: step size times order count. These two have nothing to do with each other — which is exactly why their intersection can be empty.

Finding 2: the window closes, and it closes early

Window width in percentage points («—» means no window exists):

The last leverage at which a window still exists:

Read the 25% row again. A 25-percent ladder is an ordinary grid — twenty orders at a 1.25% step. Above 3.5×, no stop loss on that grid can both clear the ladder and stay inside liquidation. Not a badly chosen one: any.

And the leverage field on most platforms goes to 20.

Finding 3: the default suggestion lands past the ceiling

The floor rule is the one that produces a number, so it is the number products put in the field. We compared that suggestion — ladder span plus 3, rounded up — against the liquidation ceiling across thirty span/leverage combinations.

In fifteen of the thirty, the suggested stop sits past liquidation. At a 25% span and 5× it lands 9.5 pp beyond the ceiling; at 10× it is 19.5 pp beyond.

This is the mechanism that killed the system in the opening paragraph. A correct rule, applied where the other correct rule was not consulted, produces a confident, automatic, unreachable number.

Finding 4: ten systems on our own fleet

We ran the same three questions against every leveraged grid in our production database — 506 systems.

And the shape of those ten is the tell. Six of them are the same configuration to within a rounding error:

Different pairs, different accounts, one number: 29%. Six people did not independently choose 29. That is the signature of an automatic suggestion — and every one of those stops sits about 3 pp behind the point where the exchange closes the position.

The MNT system was on real money and was liquidated. An ARB system, with the stop simply left at zero and a ladder spanning 48% against a 49.6% liquidation, was liquidated too.

Why this is invisible

Three reasons, and each of them is ordinary product behaviour:

The field is full. A configuration with a stop of 29% looks protected. Nothing on the screen distinguishes a stop that will fire from a stop that cannot.

The rules live apart. The floor is a creation-form concern; the ceiling is a risk concern. Each was implemented correctly, in its own place, by people who were right.

The failure needs a market move to appear. Until price travels 26%, the two configurations behave identically. The difference exists only on the day it costs everything.

There is also a persistent misreading worth naming: a stop loss does not prevent liquidation. It is an order that has to execute *before* the exchange force-closes you. Past that point, it is not a weaker protection — it is not a protection.

What to do when the window is empty

The honest answer is not "raise the percentage". If the window is empty, there is no percentage. There are three real moves:

1. Shorten the ladder — fewer orders, or a tighter step. This lowers the floor.
2. Lower the leverage — this raises the ceiling. On a 25% grid, 3× leaves a 4.8 pp window; 4× leaves none.
3. Accept geometry as the protection and turn the stop off. This is a legitimate design, not an omission: if liquidation sits well past the last rung, the grid is protected by never reaching it. Keep at least 5 pp of room. What you must not do is leave a ladder that ends next to liquidation *and* a stop that cannot fire — then neither protection exists, while the form shows a number.

How to check this yourself

Three numbers, on paper, before funding anything:

ladder span   = step % × order count       (how far down the grid averages)
liq distance  ≈ (1 / effective leverage − 0.005) × 100
window        = [span + 2  …  liq distance − 1]

Note effective leverage. A grid does not put the whole deposit into the position at once — the first order is a fraction of it, and the ladder grows from there. If your platform judges liquidation by the leverage shown on the exchange, it is judging by an upper bound. Our own first attempt at this fleet measurement used nominal leverage and produced −16 pp where the real answer was 0.6 pp: an artefact of the wrong denominator, not a finding.

If span + 2 > liq distance − 1, your stop field is decorative, whatever number is in it.

Limits of this measurement

• •Isolated margin, tier-1 maintenance margin (0.5%). On a larger position the exchange raises MMR, so real liquidation arrives *closer* than this formula says. The window is narrower than we computed, never wider.
• •One fleet, one day. 506 grid systems, 19 September 2026. The distribution of configurations is ours, not the market's — the *boundary* generalises, the *counts* do not.
• •Cross margin is a different question. There liquidation depends on the whole account, and a single number per system does not describe it.
• •Funding and fees are not in the geometry. Both move the liquidation point closer over time; we measured the static case.
• •This is not investment advice, and none of the above is a recommendation to trade.

What we changed

The intersection now has an owner — one function that returns the window, or null when there is no window at all, so that "no valid stop exists" is an answer the product can give instead of a number it cannot honour. The creation form, the clone path, the suggestion and the risk gate all ask it.

And the combination measured above — a ladder ending within the buffer *plus* a stop that cannot execute — is now a block rather than a warning for live trading. Ten systems on our fleet matched that rule when we wrote it. Three of them were already dead.

We are publishing the number because the alternative reading is worse: the rule was not missing. Both halves of it were present, correct, and never introduced to each other.