Segmentation: 199 accounts, and a name that had to change

Segments need a clock to operate on

The 909 scored companies cut into three segments and three exclusions, driven by one formula column. A segment without a time dimension is a filter. The clock is what makes it operable: who enters, who leaves, and when.

Onboard: the segmentation passA screen recording made during the build.

Career entry: AlphaForge

Stage log

Build log

Aug 27
The cut
Six buckets, summing exactly to the pool: 199 with no visible processor, 220 scaling, 201 expansion, 190 excluded as too early, 99 excluded as displacement risks, 2,023 unscored. Send seeds of 50 drawn per segment.
Week 2
The rename
The segment shipped as SEG1_INHOUSE, and that name was wrong. It asserted a build-versus-buy decision the scan never observed. All the scan saw was that no processor was visible on the marketing surface, so the segment became SEG1_NO_VISIBLE_PROCESSOR. A segment name is a claim, and every downstream email inherits it.

Timing sheet

Metrics

SEG1_NO_VISIBLE_PROCESSOR: 199
Automated gate survivors: 57 of 199Live evidence URL, numeric prices, self-serve path, and no payments companies.
Other cuts: 220 / 201 / 190 / 99Source: working sessionScaling, expansion, and the two scored exclusions. With 2,023 unscored, the six buckets sum to 2,932.
signal_score across all 199 rows: 40 of 40, a constantSource: working sessionA two-condition boolean filter wearing a score's clothing. Saying so is the point.

Source: working sessionNumbers marked with a dot come from the working session and its published posts. Unmarked numbers are stated in the build artifacts on disk, with file and line.

TELEMETRY

The constant, on screen

The metrics above call the score a constant. Here it is on screen: signal_score reads 40 on every row, why_this_score repeats one sentence, and the evidence URL is the only column that varies. Some rows carry none, which is the citation gap named in the recon stage.

Clay table view of yc-tam-pool-2932 with four columns visible, where every signal_score cell reads 40 and several evidence URL cells are empty — Four columns from the scored pool: the state, the constant, the reason, and the evidence when there is one.

SPLIT TIMES

The finding given away free

Computed from the scored pool and included in every email: 40% of the 2021 to 2022 batches show no visible processor, 30% for 2023 to 2024, 28% for 2025 to 2026. A twelve-point gap, roughly three standard errors.

The caveat travels with it everywhere it appears: older companies have more public surface, and the scan's UNCLEAR rate rises with company youth in the same direction as the curve, 49.5% to 57.7% to 63.7%. Part of the gap is the instrument rather than behavior. The caveat is load-bearing, because the measurement instrument is weakest where the thesis is strongest.

RECCE PASS

Thirty accounts, checked by hand

Before anything got sent, I hand-checked 30 SEG1 accounts against the claim the segment name makes. The in-house hypothesis survived contact with 7 of them. The other 23 broke it in four catalogued ways: sales-led funnels with no checkout to inspect, checkouts behind authentication, marketing and product living on different domains, and dead template scaffolding wired to nothing.

The board below holds every account whose finding survived to a file, and it says plainly how many did not.

RECCE BOARD

The hand-check: 30 accounts against one claim

30Checked by hand

7Held

15Documented

The segment said no visible processor. The hypothesis behind it said in-house billing. I checked 30 accounts by hand to see whether the second claim survived the first, and it held for 7.

Fifteen of the thirty checks survived to a file with a verdict attached. The other fifteen lived in the working session and were never written down, which is its own small lesson about build logs. The gap is stated here instead of hidden, for the same reason the segment got renamed.

revealed 0 of 15 · held 0

Not written down
Not written down
Not written down
Not written down
Not written down
Not written down
Not written down
Not written down
Not written down
Not written down
Not written down
Not written down
Not written down
Not written down
Not written down

15 more accounts were checked in the session and never written to a file. If those notes resurface, they get added here with the same verdict vocabulary.

INCIDENT REPORT

What broke

The segment's name asserted a conclusion the instrument could not reach, and it took a week to notice. The score driving it turned out to be a constant. And the hand-check put the headline claim at 7 of 30. All three are in the deliverable, in the same voice as the counts.

Clay formula editor open on the segment column, whose code files tier A accounts from the 2021 to 2022 batch windows as SEG1_NO_VISIBLE_PROCESSOR — The segment formula, live in the workspace. The name in the code is the corrected one.

STAGE END

Debrief

A filter answers who qualifies today. A segment has to answer who entered since last week, and that needs two observations and a clock. The snapshot layer from the recon stage is what makes the clock real.