CG Common Ground | Continuum Digital Twin
What we did and what it producedActive

The work, decision by decision

The checklist: every input an advanced twin needs

This is the working list, drawn only from parts of the line Common Ground has built or run. An input earns a line here because it was captured, checked, and in most cases already broke once before it got fixed.

  1. A voice fingerprint from the written corpus. What it is: a counted measure of how a person writes, function words, sentence length, punctuation and openers, scored against a plain AI baseline. What we captured: one partner's texts, email, dictation and call transcripts across several registers; the first run caught two of its own early mistakes before either shipped. How we checked it: held the scorer to its own test, telling a person's real writing from a generated draft, before it was trusted in any register. What broke: the first version of the scorer called the large majority of that partner's own real writing AI-written. What came out of it: a scorer never gets trusted until it passes that test on real text, not assumed text. Full build in the systems room's voice engine.
  2. Voice recordings. What it is: spoken audio, clean enough to train a voice model or score a spoken register, and call or meeting recordings for tone and cadence. What we captured: dictation history, available call transcripts, and a thirty-minute clean recording made for voice-model training. How we checked it: worked out which speaker in a recording is the person being modeled before trusting anything attributed to them. What broke: a recording device only captures the near side of a room, and a summarized note is not a quotable source, so an early pass put a claim on the wrong side of a conversation. What came out of it: quote only from the raw transcript, and check whether the far side was even captured before attributing anything to anyone.
  3. Email history, read in two passes. What it is: a person's own sent mail, read in a fast pass from an existing index and a slower deep pass for full bodies, weighted toward older years on purpose. Why the weighting: recent mail is often already drafted with help, and teaches the system the wrong voice. What we checked: held-out accuracy before trusting the scorer in any one register.
  4. A wins ledger. What it is: one row per delivered result, not a running list of adjectives, each row carrying a status and a source, not a headline number. It is the same file as the claims register below, read for results. What it is for: a result gets stated once, with its basis, and the same way on every page.
  5. A credibility and claims register. What it is: one row per claim the firm could make about its own work, with a status that answers a narrow question, what is this sentence allowed to do today, and a source. What broke: a live audit of the firm's own public surfaces, run cold against the register, found real mismatches nobody had caught by eye. What came out of it: contradicted rows get resolved before anything else, because those are the only ones doing active damage on two surfaces at once. Full build in the systems room's credibility and story pipeline.
  6. A story list. What it is: sourced rows turned into full narrative pieces on a fixed shape, the decision, the reasoning, the work, the result, what it cost, so the fact stays fixed and the sentence can change for the room it is told in.
  7. A perspectives and opinions vault. What it is: what the person has said and thinks, by topic, sourced to where they said it rather than generated to sound like them. The rule it runs under: if an opinion is not in the vault, do not guess it, flag it for the person instead. Tested the same way: a pass checked what the person said they believed against how they write and act, and the gap it found was the most useful part of it.
  8. An identity statement. What it is: a plain, sourced account of who the person is and how they work, built the same way as everything else, nothing on the page that does not trace back to something the person said or did. Tested the same way: an adversarial pass tried to break every claim in the first draft and found real errors and leaks before it shipped.
  9. Professional profiles. What it is: the public presence that represents the person, checked against the credibility register so it does not drift page by page. What broke: the firm's own structured identity data once pointed at the wrong person's professional profile, found only because the audit checked the live page and not only the file.
  10. Work history. What it is: the career record built from the same sourced material as the rest of the line, each seat or role traced to its own case record.
  11. A private vault with privacy tagging. What it is: structured storage for what the person keeps, with every record tagged for privacy at the moment it is captured, not sorted later. Tags run in tiers, from open to the person alone, and moving something to a more open tier takes an explicit yes; tightening a tag never does.
  12. A journal pipeline. What it is: a running capture habit that feeds the archive on a cadence, not in occasional bursts, so the record grows daily, one entry at a time.

The testing

Two blind tests run on the voice work, both using the same design: ten pairs, one message the person wrote, one written by a generated draft from a plain description of the same task that never saw the original. Pick which one is real, and how sure you are.

The first version: the person takes it on themselves, cold, no coaching, scored against a plain guess. The second version: the same pairs, minus anything personal, go out to the partner group as a "whose message was this" round, so people who know the person are the judges too, not only the person being modeled.

Rounds run on a rough cadence as the corpus and the exemplar bank grow. A miss in either version becomes a new rule: what the generated draft got wrong is exactly what the fingerprint was still missing. Full method in the systems room's voice engine; dated rounds and scores are on the private side of this room.

The workflow: building one for a new person

The order matters. Each step exists because skipping it, or doing it out of order, is where the first version of this broke.

  1. Capture. Pull every source the person has, not the convenient ones: their messages, their mail, their dictation history, their call recordings, their existing documents. What broke: reading messages from the raw text field alone returned almost nothing, because a modern message store writes the real body into a different format and leaves the plain text field empty. The rule: decode the real format, and if a decoded count comes back far below the raw row count, the decoder is broken, not the person's record.
  2. Mine. Count before you characterize. Turn raw text into numbers, function-word frequency, sentence length, filler rates, rather than describing a person's style in adjectives. The reason: only a measured thing can be checked against a real draft later, and an adjective list cannot be checked at all.
  3. Structure. Turn the mined material into rows and registers, not prose. One row per claim, one row per win, passages tagged by register and intent, not pasted wholesale, personal material sorted into a fixed set of categories. A row can carry its own status and source; a paragraph cannot.
  4. Verify. Run an adversarial pass that tries to break every claim before anyone else does, and audit live public surfaces cold against the register. What this caught: a public page that had been live for months with nobody watching it, and a factual mismatch nobody had noticed by eye.
  5. Voice-test. Score a draft's distance from the real register against an AI baseline before it ships, and run a blind test where the actual person tries to tell their own writing from a generated draft. Failing to tell them apart is the bar the system has to clear, not a compliment it gets for coming close.
  6. Publish rules. Nothing goes out without a tier. Privacy is tagged at the moment of capture, not sorted afterward; moving something to a more open tier takes an explicit yes, tightening never does. A claim carries a use lane the same way it carries a source, because what a sentence is allowed to do changes by audience even when the underlying fact does not.

What it cost, and what we would watch

The parts furthest along both found real problems before anyone else could: a scorer that called most of a person's own real writing AI-written, and a public page that had been live and crawlable for months with nobody aware of it. That is the return on testing on your own record first. A system nobody here depends on is a demo. A system the firm runs itself teaches you, on your own mistake, where it breaks, before a client ever sees it.

What it produced

The checklist runs on the same discipline across every input: capture the real material, check it against itself before trusting it, and tag privacy at the moment of capture rather than after the fact. All twelve inputs exist today on the firm's own people. The voice engine and the credibility register are written up in the systems room; the story list, opinions vault, identity statement, work-history record, tagged vault and journal pipeline each run from their own sourced files. The founder's own June 2026 concept, a capture stack tested on his own life before anyone asked whether it was a product, is where the work started and is still running today.

A slice of the project list

A few related projects.