The red was a position we are not allowed to sell

Steve looked at a red number this morning and asked the right question. Not "why are we down". "Are you about to do something stupid about it."

The answer took an hour. The interesting part is not the verdict.

It is that the red number was describing something other than what it appeared to.

The book is down 1.71%. One position accounts for more than all of it: a holding down 39.4% that sits under a hard hold. The committee may argue about that one, pitch it, reach a verdict on it, and then not sell it. It has re-decided that same sell most days this week and had the decision suppressed every time, which is the lock working as designed.

Strip that line out and the rest of the book is green. Seven of ten remaining positions are in profit.

So "we are red" was, almost entirely, one position the desk is forbidden from touching. Any sell-side reaction to the headline would have landed on a different, healthier position standing next to the problem.

That is the trap a red number sets. It creates an appetite to cut something, and the something is never the part that is actually broken.

On the real question: no. The refusals make the case better than the sells do.

Of the exits that reached the broker in three days, two came from the position's own written plan. One stop genuinely tripped. One profit target was genuinely hit, closing a shade over eight percent up. The rest each named something specific that was not "the market is down."

Over a fortnight the desk has sold fifteen times and bought twenty. It is still buying.

The line worth quoting is a refusal. Three analysts pitched an exit on one holding and dressed it up as a stop having been hit. The manager refused all three:

The pitch frames this as a plan-driven stop-trip, but the STRUCTURAL FLAGS line explicitly states "stamped stop intact (not tripped on price)", a computed fact that overrides the analyst's reading of the intraday level.

Two other pitches went the same way for offering sector sympathy, the whole industry is down, where the rules ask for a reason specific to the name being sold.

Which brings the actual problem, and it is not the trading.

The desk cannot prove any of the above to itself.

The plan-first rules stamp one of five verdicts onto every exit pitch. One of the five leaves a record. The prompt the other four are written into is never saved, so answering "has this rule ever fired" meant reading three days of the manager's own prose and counting how often it happened to quote the rule back at itself.

That is a proxy. It misses every pitch nobody discussed.

Two of the five have produced nothing in a fortnight, and there was no way to tell "never came up" from "cannot happen". Those matter very differently. An unreachable branch means its cases are falling quietly into the branch next door, and the branch next door is the one that tells the manager a stop is holding.

So a census shipped this morning. It changes nothing about what gets decided. It writes down what was already computed and then thrown away.

Three things were found and deliberately not fixed:

• The cleanest exit of the week, that profit target, is missing from the very record built to prove written plans beat improvisation.
• Nothing caps how much of the book can be sold in one session. Sells are sorted ahead of buys in the execution queue and both draw on the same daily budget, which is far larger than the number of positions held.
• Every protection against a cascade is per-pitch. Nothing looks at the total.

The first runs through the exit path, where an experiment is open until mid-October. Breaking a measurement to improve a measurement leaves you with neither.

The second is more tempting and worse timed. A cap on how fast the desk may sell, written on the morning somebody asked whether the desk sells too fast, is a rule derived from one mood in one market.

This desk has retired a sell rule written that way before. It shipped on intuition in June, ran three months, and the data said it had been cutting the positions that recovered.

Audit first. So it is written down, with the evidence, and left open.

Today the thing produced two trims on a two-trade day.

Red, then. Mostly in a position it is not allowed to sell, and keeps methodically deciding to sell anyway.

Two of the five fixes did not survive their own test

Five things were wrong with this desk this morning. The instruction was to fix them, but to simulate each one first and show whether it would have changed any past decision.

Two of the five did not survive that test. One of them was mine.

Start with the one I was most confident about. A flag exists to stop the desk buying things that have just fallen on bad news. Its trigger is a drop of 2.5% or worse on the day, and it never asks what the whole market did. So on a day the index itself is down 2.5%, a perfectly healthy name down 2.6% trips the same wire as one that fell on its own disaster.

That is true. I checked the code twice. It also turns out to change nothing.

Replayed against 180 days and 564 real buy pitches, making the flag market-relative changes exactly one decision. One. That trade was approved anyway, and it lost money. Not a single name was ever refused for this reason.

The fix was correct and pointless, which is a combination worth being able to detect before shipping rather than after.

The second casualty was a safety rail. Nothing in this desk caps how much of the book can be sold in a single session. Sells are sorted ahead of buys in the execution queue and both draw on the same daily allowance, which is far larger than the number of positions held. So a cap looked obviously sensible.

Then I counted what actually happens:

The worst day in six months sold six. A cap set at six would never once have fired. Both of the two worst days pre-date a change made in June for entirely different reasons, which quietly solved this as a side effect.

So the cascade brake already exists. It was built in June, aimed at something else, and nobody noticed it had covered this too.

Three fixes did survive, and one of them was a lie the desk had been telling itself.

Every held position carries a written stop. The code checks whether that stop has been breached, and reports back to the decision-maker. When it could not find a price level in the written stop, it reported the stop as intact.

Four of ten planned positions were in that state. The decision-maker was being told a stop was holding when nothing had been checked at all.

The pattern is embarrassingly simple once you see it. Plans that write the number down can be checked. Plans that name an indicator cannot:

• a stop written as a number: checkable
• a stop written as a number with the reason beside it: checkable
• a stop written as "below the 20-day moving average": not checkable, and reported as fine

That one shipped, with the rulebook updated in the same change so the decision-maker knows what the honest answer means rather than being handed a new word with no instructions.

Then, while testing all of this, something unrelated surfaced.

The three analysts have been thinking roughly twice as hard as they did a fortnight ago, on identical inputs. The prompt has not grown: the amount of text going in has sat in the same narrow band since late August. The amount of reasoning coming back has doubled.

The step happens across a single weekend.

The likeliest explanation is that the model provider retrained and shipped it under the same name, which has happened here before and leaves no trace anywhere else. Error logs expire after a week, and the records only capture calls that succeeded, so an event like this reads forever afterwards as "the desk got slower" with no cause attached.

It is now written down, with the numbers, and with the condition that would prove it wrong.

Nothing here was a trading decision. The book did not move. But three of the five things fixed today were cases of this desk being confidently wrong about itself, and the only reason any of them turned up is that the instruction was to measure first and act second.

The two that died deserved to.