This week the engine onboarded a new voice. A B2B logistics client signed on Monday, and by Friday the writer brain was producing drafts their marketing lead described as closer to their voice than the last freelancer got in three months. The log below is how it got there, including the two regressions.
Numbers first: 40 writing samples ingested, a 61-point starting voice score, an 88 by draft twelve, one rollback, and 74 minutes of total human grading time.
Here is the dispatch.
What shipped this week
The voice module ran its full onboarding loop for the first time on a client outside our test group: ingest, fingerprint, draft, score, grade, update. Twelve drafts shipped to review across five days. Ten cleared the client's editor with light edits or none. The engine also wrote itself nine new voice rules, discarded two candidates, and rolled one back after it made things worse.
The client, anonymized per our case policy, is a 20-person logistics software company with an unusually distinct voice: dry, technical, allergic to enthusiasm. Their archive runs to about 60,000 words of docs, release notes, and a founder blog.
That last property is what made them a good stress test. Generic AI writing defaults to enthusiasm. Theirs punishes it.
Elapsed wall-clock from contract signed to first client-ready draft: 31 hours, most of it spent waiting for the sample archive to arrive.
How the fingerprint gets built
The fingerprint is not a vibe. It is 14 measurable features extracted from 40 samples of the client's existing copy: median sentence length, sentence-length variance, paragraph rhythm, contraction rate, first versus third person, vocabulary tier, banned-word list, punctuation habits, and how they open and close arguments. Forty samples took the ingest agent 11 minutes. The output is a two-page voice spec a human can read and argue with.
The readable spec matters more than the extraction. When the client's editor saw "contraction rate: 4 percent — this brand almost never contracts," she corrected it to never in legal-adjacent copy, always in support docs. That one correction was worth ten drafts of trial and error.
The features are deliberately boring. We tested embeddings-only voice matching in April and could not explain its failures to clients. Fourteen countable features lose a little fidelity and gain all of the debuggability.
Fingerprints are stored per client and never shared across them. One brain, separate memories — and a client's voice spec exports as a two-page document if they ever leave.
Draft one scored 61 and deserved it
The first draft against the new fingerprint scored 61 out of 100 on voice match, and the failure was instructive: it sounded like us. Sentence lengths matched Avakata's house style instead of the client's longer technical cadence, and it used two words from their banned list. The scorer flagged both. Nothing about draft one was wrong as writing. It was wrong as their writing, which is the entire assignment.
The 61 was not a floor we engineered for drama. It is what a well-prompted generic model scores against a distinct brand, and it is why "the AI sounds generic" is the most common complaint we inherit from new clients.
This is the trap most AI content workflows never detect. Without a per-client fingerprint and a scorer, draft one ships, because it reads fine to anyone who does not live inside the client's brand.
Scores below 70 never reach a human. Draft one died in the pipeline, as designed, and the editor never saw it.
What ten human grades taught the critic
The client's editor graded the first ten drafts on a one-to-five scale with a one-line reason, at roughly 90 seconds per draft — 15 minutes total. The engine converted those reasons into candidate rules: no rhetorical questions, no exclamation marks anywhere, numerals for everything above nine, open with the operational problem rather than the industry trend. Each rule got attached to the fingerprint and weighted by how often it appeared in the grades.
Two candidate rules were discarded because they contradicted higher-scoring archive samples. The editor's memory of the brand and the brand's actual archive disagreed, and the archive won with her sign-off.
That conversation — archive versus editor — is the most useful artifact of the week. It surfaced a style drift the client did not know they had.
Grading is the entire human job in this loop. Fifteen minutes of structured judgment moved the score more than any prompt engineering we tried.
Two regressions and one rollback
Draft eight scored 84 but read clipped, and the human grades caught what the scorer missed: the engine had over-applied the short-sentence rule, driving median sentence length to 9 words against the client's actual 17. Rule seven was rolled back within the hour, the fingerprint re-weighted, and drafts nine and ten recovered to 86 and 88 with the cadence restored. Total cost of the regression: two drafts and one editor note.
The second regression was subtler. The engine started opening every piece with a statistic because graded samples rewarded one that did. Pattern, meet overfit. We capped structural rules at 60 percent application so no single opening dominates.
Rollbacks are cheap when every rule is versioned. This week is why every rule is versioned.
We log regressions with the same care as wins. The regression log is where next quarter's guardrails come from.
The rules that survived
Nine rules survived the week and now sit in the client's voice memory: median sentence length 15 to 19 words, zero exclamation marks, zero rhetorical questions, contractions only in support content, numerals over spelled-out numbers, open with the operational problem, close with a next step rather than a summary, a banned-word list of 31 entries, and statistics always stated with denominators. Every rule carries its source grade and its rollback history.
None of the nine is exotic. That is the finding. A brand voice is mostly a small set of enforceable habits applied without exception, which is exactly what software is for.
The mystique of voice does not survive contact with measurement. The voice itself does.
Boring rules, enforced without exception, is what a brand voice turns out to be.
What the brain does differently now
Voice is now applied at draft time, not edit time. Before this week, the writer drafted generically and a style pass rewrote toward the client. That two-step lost nuance and doubled token cost. The fingerprint now conditions the first draft directly, and the style pass has been demoted to a checker. Draft-time voice cut editing time per piece from 14 minutes to 4.
Ongoing cost to the client: one 20-minute grading session a week for the first month, then monthly. The fingerprint keeps updating as their brand drifts, because brands drift — the engine flags it when three consecutive grades disagree with the archive.
Next week the same loop runs for a client with the opposite problem: a warm, chatty consumer voice. The engine does not care. That is the point of fingerprints.
Score at signing: 61. Score at Friday's close: 88. The gap is what one week of structured grading buys.