The rewrite is the signal. Keep where every word came from.
Began as layered author-voice research; now keeps writing provenance and turns the script rewrites I accept into editorial policy.
The idea
VoiceOfTheTexts began with a plain frustration: AI-drafted scripts did not sound like Christian. The name points at the larger goal, a voice of the texts, meaning an author's voice captured well enough to write in it. Getting there needs clean evidence, and that was the first surprise. AI drafts, pasted prose, and quoted text had quietly leaked into the writing meant to represent him, so no measurement of it could be trusted.
The bet since then is that evidence has to be kept at the moment of writing, not rebuilt afterwards. Every block is labeled as Christian's own, pasted, or AI-assisted, and the most useful signal turned out to be the rewrite itself: the gap between a draft and the words he chooses to keep. Only choices that repeat, and that he confirms, become editorial policy for the next draft. One rewrite is one decision, not proof of a voice.
Today, the provenance record and a small, manual rewrite loop are active, and notes from FocalaneCapture reach VOTT's signed Mac relay. Voice research is paused, not abandoned. In one attribution experiment the early layered analyzer held its own at telling authors apart, but writing in a measured voice failed its tests. That work resumes only with a new method and a fresh calibration group, and until then no voice model is claimed.
How it works
Draft under policy
Owner rewrite
Accepted for filming
Next policy
Each block of writing is labelled mine, pasted, or AI-assisted, and only the author's own blocks count as evidence. In the rewrite loop, Focalane drafts a script under VOTT's written editorial policy, Christian rewrites and accepts it for filming, and a content-free ledger keeps only recurring, confirmed choices for the next policy version; no script text enters VOTT, and no model is trained. The signed Mac relay receives settled notes from FocalaneCapture; review and admission stay manual.
Where it stands today
VOTT began on 24 August 2026 as a layered voice extractor. In one 14-way development-corpus attribution experiment, its deterministic layers matched Burrows' Delta's top-1 accuracy with about one-fifth as much enrollment text; that says nothing about writing in a voice, and writing in a measured voice failed a preregistered test on 1 September and a measurement stage on 3 September, so that research is paused. Two lanes remain active: per-block provenance, with an owner review of 189 notes and documents complete, and a manual rewrite loop: four rewrites of Focalane script drafts have been accepted over time, and three still match their filming copy exactly. No final cut is accepted yet. Notes from FocalaneCapture reach VOTT's signed Mac relay; import stays manual.
A completed owner provenance review of 189 notes and documents
A content-free ledger of accepted script rewrites with a fail-closed audit
An editorial policy (version 0.2) distilled from accepted rewrites and used for Focalane's next drafts; a successor is pending
A signed Mac relay that receives settled FocalaneCapture notes
A preserved research record, including the failed results
Not there yet
The rewrite loop is small (three current filming-copy acceptances) and not yet tested against a control.
The four judged layers were never built, and the legacy voice profile mixes author features with generic prompt policy.
How it got here
Began as a layered writing-voice extractor: the L0–L11 layered stack, fusing measurable habits from spelling and punctuation to syntax and word choice into one author profile; four deterministic analyzer layers were built
In one 14-way development-corpus attribution experiment, the deterministic stack matched Burrows' Delta's top-1 accuracy with about one-fifth as much enrollment text
Pivot: the contamination came from reconstructing provenance after the fact, so the next build became a capture app that records provenance at the keyboard
Offline web capture app built; the preregistered voice-generation test failed
A rebuilt voice-measurement stage failed its gates; the research was halted and preserved
Native iPhone capture app started, and a practical drafting lane opened beside the paused research
The first owner rewrite showed the rewrite itself is the useful unit, so owner revision became the main workflow
Three accepted rewrites opened the editorial-policy checkpoint; the iPhone app became the base of FocalaneCapture
FocalaneCapture replaced VOTT's iPhone app in place; VOTT kept the Mac relay
What comes next
Next step
Finish the owner-rewrite pass and re-audit the editorial policy; production learning opens after three accepted final cuts.
Later ideas, not current features
Rewrite and accept the remaining nine policy-assisted drafts one at a time, then re-audit the policy
Open production learning after three accepted final cuts
Optional: a blinded comparison of task-only and policy-guided drafts
Resume voice research only with a new method and a fresh calibration group
What I checked, and what stays bounded
Dated evidence
One 14-way attribution experiment: Burrows' Delta's top-1 accuracy matched with about a fifth of the enrollment textresearch
Voice-measurement stage failed its gates; research pausedresearch
Same-bundle Finder restore verifiedattended
Private repository published and independently verifiedrelease
Three accepted rewrites opened the editorial-policy checkpointpolicy
Relay delivery of a single note physically acceptedattended
Signed Mac relay installed and acceptedrelease
Boundaries it keeps
No validated voice model; voice measurement failed its gates and is paused
VOTT does not draft scripts or call AI providers: Focalane drafts, and Christian rewrites and accepts
The provenance system protects future voice research from mixed-origin text; the rewrite loop separately records editorial preferences. Neither proves personal voice
Production learning is closed: no final cut is accepted yet, and three are required
Relay delivery is not backup or cloud sync; import and admission are manual
VOTT cannot write to Focalane, and FocalaneCapture is a separate product
Private writing is not public evidence
My role
Set the research question and its pass/fail gates, kept failed results on record instead of tuning them, reviewed the provenance of 189 notes and documents, and rewrites and accepts every script in the loop.