The best change of the week was the one we did not make

The weekend review's top recommendation was to widen the discovery pool. The argument was good: ten of the eleven names that moved more than 10% that week had never been shown to the analysts at all, and one of them had been running at eighty-two mentions before it moved 20.5%. The pool looked blind, and widening it looked obviously right.

Steve asked for it to be simulated before it shipped. So it was.

The simulation replayed every committee briefing over twenty-eight days from the raw news rows, worked out where each missed name actually died, and then priced the alternatives: what would have entered the pool under each widening, and what those entrants would have done against what they displaced.

The misses were not dying where everyone assumed. Almost none of them were lost to the fetch limit, which was the thing the recommendation targeted. They died at the rule requiring two distinct sources, and at the minimum mention count. A big chunk died at something no pool setting can reach at all: the mention burst happened on a day or an hour when no briefing followed. One of them first appeared on a Saturday.

Then the part that settled it. Dropping the source rule does surface five of the eleven misses, but at or after the date they moved, which is worth nothing. And its entrants underperform the names they would have displaced. Loosening the pool would have bought worse candidates and paid for them by dropping better ones.

Only one variant had positive economics, and it surfaced none of the misses.

So nothing shipped, and the reason is now written down where the next person to have this idea will find it. The deeper reading is more uncomfortable than the recommendation was: the miss class is mostly earnings-print pops and momentum chases, which is the desk's own worst-performing category of entry. The pool is not as blind as it looks. The misses are mostly not catchable in time, and where they are catchable, not worth chasing.

A week's headline recommendation, refuted by its own simulation before it touched production. That is the system working.

This week's calibration

The scout's high-conviction rate came in at 28.1%, against 21.8% the week before. That is a 1.3x week-on-week move, comfortably under the 2x tripwire that would suggest the vendor had recalibrated the model underneath us again.

Buys running at 1.7 a day, manager accept rate 37%, and accepted buys positive at seven days 76% of the time on a thirty-day shadow. Running cost is over the monthly target and about to get worse, because the intro pricing on the manager's model expires at the end of the month.