v0 · pressure-test  ·  not public, not indexed  ·  built, used, still being challenged
Judgment & critique

Five intelligent people, disagreeing, on demand.

Voices convenes a multi-persona expert panel against any artifact you give it. Each voice renders in character, from a real corpus — not generic AI feedback with a name attached. And when they disagree, the disagreement is preserved verbatim, never quietly averaged into a consensus that loses the signal.

The problem it solves

"Looks good to me" is the most expensive sentence in your work.

Ask one model for feedback and you get one perspective, smoothed toward agreeableness. Ask a room of real experts and you get friction — the kind that catches the flaw before it ships. Voices gives you the room: pick the thinkers, point them at the draft, and let them argue. The most valuable output isn't where they agree. It's exactly where they don't.

Five voices agreeing is not consensus. Five voices each registering verbatim is signal you can act on.
How you use it

Point it at a draft. Pick the room.

1

Choose the artifact

A spec, a strategy memo, a positioning page, a piece of code — anything you'd put in front of sharp reviewers.

2

Compose the panel

Name the voices, or pick a preset (harshest-critique, steel-man, client-review). A sycophancy guard keeps "steel-man" from collapsing into flattery.

3

Read the room

Per-voice critiques in their own cadence, the cross-voice patterns, and the load-bearing dissents — each preserved with why it matters and what it counters.

What makes it different

The conservation law: dissent is never destroyed.

The rule the whole product is built around

Aggregation is a conservation violation.

Most "panel" tools collapse five opinions into a majority report. Voices refuses. Disagreement is treated as Boyd's outside reference point — the perspective you can't get from inside the system — so it is carried through to the output intact, each dissent annotated with why it was preserved and what it stands against. The output is two channels that are never merged:

Machine assertions

Where the panel converges — the patterns any reader would act on, stated plainly.

Named dissents

Where it diverges — preserved verbatim, per voice, with a qualification block. The signal lives here.

How it comes to be · where it lives

Two views of the machine.

The first diagram traces how a single Voices run unfolds — from choosing the artifact through parallel per-voice dispatch to the two-channel output that never collapses. The second shows where Voices sits in the portfolio: which products it reads, what it produces, and the governance loop that holds it together.

How it comes to be
Choose artifact Compose the panel agent ① agent ② agent ③ parallel renders in-voice from corpus machine assertions named dissents never merged verbatim Choose → Compose → Dispatch (parallel) → Render → Two-channel output
Where it lives
Intent loop — Notice · Spec · Execute · Observe Cast registry corpus archetypes persona identity Voices Judgment position dissent conservation law machine assertions named dissents Forge renders skills You two-channel output reads sibling output Cast (identity) → Voices (judgment) → You   ·   Forge is sibling, both draw on Cast
Behind the veil

How Voices was forged.

Voices wasn't designed in a slide deck. It was built — and is still governed — by the same machinery that built every product in this portfolio. Slip behind the veil to see the operating model that produced it and the system it composes into.

The operating model

Intent

Voices carries its own .intent/ governance — signals, specs, decisions, traced in git. The dissent-preservation law is a ratified spec; the validator enforces it. This page was itself critiqued by Voices before it shipped.

See the operating model →
The overhead view

Parallax

Voices occupies the Judgment position in a larger coherence stack — it injects outside reference points so the system can't fool itself. See where it sits among the other twenty-two implements.

See the overhead view →