Time to a corpus you can defend

Note · KBMill · value · speed

When people say they need “speed” on an on-prem LLM program, they usually do not mean the raw clock rate of a text parser. They mean calendar time to an auditable package someone can stand behind — without opening a multi-month data-cleansing program.

That is the speed we care about: time to a defensible corpus, not “faster OCR than tool X.”

What the predominant path often costs in calendar

The usual blueprint looks serious on a slide: provision the box, configure weights, wire vector retrieval, “we’ll clean the docs.” Document preparation is routinely deferred into an open-ended cleansing phase. In practice that becomes continuous, unmetered cycles of scripts and re-indexing while the share keeps moving — and still may never produce a portable leave-behind with listed limits.

Dump-and-hope is “fast” on day one and slow forever after: week two fails, and the calendar resets. A platform project can be thorough and still burn a quarter before anyone holds a package a stakeholder will defend.

Our critique of that paradigm is simple: data-engineering complexity should be solved upstream as manufacture, not managed forever as an integration project. Both other paths can be honest work. Both are expensive in time-to-trust.

How we see the speed problem

Complexity belongs in the plant, not in a ritual the customer has to staff. You drop a bounded pile. It is $10 if we produce a usable residual-honest ZIP. A scan-heavy pile can take longer. That is time, not a different price. You keep the package and point your own stack at it.

This is not only a slide claim. Hard technical stock has been manufactured on a commodity ~$1,500 workstation-class tower — the same class of machine behind our dogfood and public brick work. One long dogfood run — Merrill Skolnik’s Introduction to Radar Systems, about 590 pages of non-standard defense tables and equation-heavy layout — finishes in about an hour of wall clock (on the order of fifty minutes to an hour, depending on the pass). That book is private dogfood only (rights); it is not on the public shelf. The point is the time class: Hard craft on brutal stock in job time, not a multi-month cleansing program — and not a downloadable sample.

On this pilot mill, Small often finishes while you wait. Medium is often a short wait. Hard — near-ceiling or scan-heavy work — often takes on the order of an hour, not a quarter. That is plant time for one bounded job, not a promise that every enterprise estate finishes before lunch.

The method’s bet: a metered manufacture job gets you to a defensible leave-behind in a different time class than “hire a RAG team and see you in six months.” The ~$1,500 tower is the proven baseline — what a solo shop could already run. Real leased compute should only improve plant time; it is not required to make the months-vs-job-time contrast true.

Therefore speed is a product shape, not a boast

Drop the bin — no open-ended workshop to start
$10 if it ships — a scan-heavy pile can take longer; the price does not change
Pay on success — calendar risk is not paired with paying for a failed run
Portable ZIP — structured stock plus explicit exclusion metadata you keep — not another hosted pilot that evaporates
Intentional caps — 15 files / 150 MB keep early jobs bounded so “fast” does not mean “swallow the share drive”

Split large programs into jobs or ask about a quoted bundle. Federation across many bricks stays in your stack after you take the ZIPs home. That is how calendar stays honest while ambition grows.

What we are not claiming

We are not publishing a benchmark that “KBMill is always N× faster than every integrator.” We are not claiming Hard jobs are instant, or that air-gap / regulated paths match the public hopper clock. We are not racing the $3 extract API on page count. We are not redistributing Skolnik — the ~hour dogfood timing is private hard-stock evidence, not a public ZIP.

The claim is categorical and anchored: predominant methods often spend months reaching a corpus someone will defend; a residual-honest mill is built so a bounded pile can become a keepable package in job time — about an hour for Hard craft we have actually run on a commodity tower. Whether that pass is closer to fifty minutes or sixty, it is still a different time class than a quarter of open-ended cleansing. Argue with open packages on the public shelf if you disagree with the method.

Door

If you approve the spend, ask how long it takes to get a package you can defend, not only when the box goes live. If you write the estimate, put document preparation next to the hardware. If you have a pile that fits the caps, the mill is open: drop the files, keep the ZIP, pay only if we produce.