Your voice leaves the lab.
Until tonight the voice brief sat in a disclosure on the portrait page and nothing else read it. Now drafts and the Composition mentor write from it. There's also a small game that tests whether it's any good, and every miss you name goes back into the brief.
1. Spot the Atticus
The Foundation sets the bar for writer voice preservation: with a sample portfolio, a draft should be "blind-test indistinguishable from writer's own work by someone who knows them." Voice Lab now runs that test, and the person who knows you best is you.
Atticus takes one of your Part II rewrites and rewrites the same plain paragraph from your brief alone. It never sees your version, and any exemplar sentence taken from that paragraph is withheld. The two appear in random order as A and B, and you pick the one you wrote.
- If you spot Atticus, you name what gave it away with quick chips (Too tidy, Words I wouldn't use, The rhythm's off…) or your own words. That note goes into the next synthesis as a correction, and the portrait is rewritten.
- If Atticus passes for you, the brief met the bar for that paragraph. You can still say what was off.
- Rounds rotate to your least-tested rewrite. The tally reads, for example, "Atticus has passed for you 2 times in 5 rounds." Atticus's paragraph never becomes evidence.
A live run through the Claude CLI on synthetic fixture text took about 10 seconds and returned this, from the brief alone:
It rained all Saturday, the steady gray kind, so we stayed in and did the jobs we'd been ducking for a month… Then Sunday. The sky cleared, and we walked the lake path while the water lay there flat and smug, like it had done all the work.
2. The brief travels
First drafts, redrafts, and the Composition mentor now receive the brief as a directive. It governs diction, rhythm, figures, stance, and address, and it never supplies material. The draft prompt ranks the brief above the writing type's default register. Only the note for the project's kind of writing is included. While a draft is written from the brief, the loader reads "Atticus is drafting in your voice…"
A checkbox under the brief keeps it home. When it's off, drafting works exactly as it did before.
3. Readings that fail say so
This comes from your run, where two model readings failed silently. The portrait now says when pieces didn't come back, or when the written portrait couldn't be finished, and offers Try again. Refresh counts its way through: "Read 2 of 4." A draft whose reading failed is re-read on retry too.
For you to try
- Check out
overnight/voice-travels, open your portrait, and play Spot the Atticus two or three times on your real rewrites. - Name one giveaway and watch the brief change after the refresh.
- Start a draft and see whether the prose sounds more like you.
Things to know
- The work lives on a branch, not main, so you can review it before it lands. Three commits, 852 tests passing, 0 failures.
- Recognition by memory is real: you may know your own words on sight. The giveaways are still the useful part, because they tell the brief what's wrong.
- Blind rounds generate prose on a neutral sitting paragraph, not on an assignment. Student projects still never receive generated drafts.
- Not wired yet: the Draft Review, Handoff, and Assembly services still read only the seed text. That's a small follow-on if you like how drafting feels.
- I tried to run a blind round against your saved Voice Lab data, and the auto-mode classifier blocked reading it (personal data). The live check used fixture text instead, and I deleted the copied store.