Carlos Montoya
← WORK INDEX

Taste, Versioned

I made a 165-file design corpus legible enough to curate, then made evaluation failures inspectable.

BRAND PATCH / SCOPE TEST15 AUG 2026
Keeper false-cuts in the control, patched, and repeat evaluation arms
TRACKCONTROLPATCHEDREPEAT
BRAND301313
INTERFACE142221
DECISIONHELD / UNSHIPPED
The Brand patch worked on its target and regressed Interface. The repeat confirmed both effects, so the unscoped patch was held instead of shipped.
01165files accounted
0221,678candidates
03592run-one review tiles
04116separate labeled gate

The corpus

I made a 165-file design corpus legible enough to curate, then made evaluation failures inspectable. The coverage ledger accounts for every file: 124 yielded candidates, while the rest were explicitly classified instead of disappearing into an unknown bucket. The complete crawl surfaced 21,678 candidate nodes.

The point was not to automate the final selection. It was to make a large archive inspectable enough that I could spend judgment on the work that earned it.

The evaluation

The first review run produced 592 tiles. A separate 116-tile gate carries my own labels and written rationales, so a judge can be measured against a durable reference instead of trusted because its output sounds confident.

That distinction matters. One experiment exposed a missing Brand rule in the live harness. Restoring it reduced Brand keeper false-cuts from 30 to 13, but increased Interface false-cuts from 14 to 22; the repeat landed at 13 and 21. The unscoped patch was held instead of shipped. The failure became an inspectable system defect and a narrower next test.

What AI did

Deterministic filters and local models handle volume. Frontier judges are measured against my labeled rationale corpus, not trusted by default.

I designed the pipeline, rubric, evals, and final human gate. The system can narrow and challenge. The outward choice remains mine.