One Eval Engine, Two Industries, One Script For The Third — What Transferred

When you build something that works for one industry, the real question isn't whether it works — it's whether the next industry costs you a rewrite or a rounding error. We had a concrete way to check.

What's already running in two industries

The same eval engine — one runner, one rubric format, one judge-scoring script — already produces real, published numbers for two unrelated verticals. Real estate: our realtor_listing_copy skill wins a blind, order-swapped comparison against the plain model 29 times out of 30 (1 tie, 0 losses) — a +83.3 point delta with a 95% confidence interval of [66.7, 96.7]. Content: our humanizer skill's result on the same corrected engine came back genuinely inconclusive rather than a clean win — we published that finding, and the harness problems it surfaced, in a companion post. Both scorecards are live on /benchmark, scored by the exact same script, with no vertical-specific logic in the engine itself.

That's the baseline. The interesting question was always the third industry — the one nobody had built for yet.

Standing up a third vertical, timed

We picked home services — a quote-writing skill, nothing in common with real estate or e-commerce — and ran our own scaffold procedure against it as a dry run: write one skill body, write its eval tasks and rubric, run the engine, check the result against the acceptance gate. Then delete it; it was a timing exercise, not a live launch.

Wall clock, start to finish: 4 minutes 17 seconds.

We also counted lines. The new, vertical-specific content — the skill body, its eval tasks, the rubric — came to 203 lines. Everything it reused without a single change — the runner, the rubric-scoring logic, the ratchet that gates a rewrite, the skill-file format — came to 2,035 lines. That's a 90.9% reuse ratio: nine-tenths of what a "new industry" needs was already sitting there, built for the first two.

Why that number is the actual point

A skill that only works for one vertical is content. An engine that scores any vertical the same way, with 90%+ of itself untouched when you point it at something new, is infrastructure — and infrastructure is what actually compounds. The 200 new lines were the only genuinely new work; the other 2,000 were the asset, reused for free.

We're not claiming home services is a live, benchmarked vertical — it isn't, and we said so above. What we're claiming is narrower and, we think, more useful: adding a third industry to a system like this is a same-day task, not a re-architecture, because the part that's actually hard — running a fair, judge-scored, two-arm comparison and refusing to publish a regression — was solved once and never touched again.

If you want to see what the engine looks like scoring a live skill rather than a timing exercise, /benchmark has the real numbers for both industries it's actually running in today.