Voice and corrections

A voice profile is a machine-readable description of how your content should read: the vocabulary it should use, the terms it should avoid, and the tone and style that make it recognizable. Most workspaces call their default profile brand; a workspace with several audiences carries several profiles, each bound at the point it governs. On its own, a profile is a set of checks: kapi and Bowrain run it over content and report what drifts off-voice, the same way a test reports what broke.

Bowrain adds the part a single checkout cannot: it learns the profile from your team's corrections and keeps the result versioned, shared, and enforced. Voice is a section of the Context hub; this page covers the correction loop in detail.

Every correction becomes a signal: Bowrain aggregates repeated fixes into candidate rules, shows each rule's blast radius, and promotes it into a versioned, enforced check.

The loop

The rules a team most needs are the ones it discovers by correcting real output: the phrasing it keeps changing, the competitor name that slips in, the term it has decided against. Bowrain captures those corrections and closes the loop:

1. Correct. A person in the review session or the editor, or an AI assistant working through MCP, fixes content the profile should have caught. Bowrain records the correction with provenance: the original text, the fix, the voice dimension, the point it was made at, and who made it.
2. Surface. When the same (original → corrected) correction recurs across the project, it surfaces as a candidate rule. Candidates are recomputed live from the correction stream, so they always reflect current evidence.
3. Preview. Before a candidate is enforced, Bowrain computes its blast radius: it runs the rule over your stored content and shows how many blocks it would newly flag, how many it resolves, and which become critical, so you see the impact before anything changes.
4. Promote. A reviewer promotes the candidate into the profile. It becomes a deterministic check, the profile version is bumped (the prior version is archived for audit and rollback), and the rule is enforced on every future generation, in every locale the profile covers. The promoted rule is concept-backed: it references the concept its term denotes, so the rule and the term it governs share one identity in the Context hub. Promoting a voice rule is a governed edit, and a concept's story shows the rules that reference it.
5. Monitor. After a rule lands, Bowrain watches aggregate compliance and raises a drift alert when it erodes, so a profile that was healthy does not quietly decay as content and models change.

Correct once; the rule that prevents the mistake is authored, versioned, and shared.

Correct once, and the rule that prevents the mistake from recurring is authored, versioned, and shared, the way adding a regression test stops a fixed bug from returning.

Rules inherit and expire

A rule is bound at a point and applies to everything beneath it until a narrower profile overrides it. A rule can also carry a validity: it starts and stops on a date, so a campaign's vocabulary or a rename's transition period expires on its own instead of needing a cleanup project.

Progressive autonomy

Every profile starts fully manual: each candidate waits for a human to promote it. As a team learns to trust the candidates a dimension produces, it can raise that dimension's auto-promote threshold, so a correction that pushes a term past the threshold promotes its rule automatically and announces it. Autonomy is opt-in and per-profile; a team widens it dimension by dimension as confidence grows, and never all at once.

Where it shows up

The loop is drivable from three surfaces, all over the same voice engine:

• Web: the Voice section of the Context hub lists pending candidates with the corrections behind them, shows a candidate's blast radius on demand, and offers promote or reject. Promoted and rejected candidates move to history; drift surfaces as an alert on the Context dashboard.
• MCP: an AI assistant can read what its own corrections have surfaced, preview a candidate's impact, and promote it, so the assistant participates in governing the voice it writes in.
• CLI / CI: once a rule is promoted it is an ordinary check. kapi enforces it locally and in CI, gated by exit code, with no server round-trip.

The detection and scoring are the open engine's job; Bowrain owns the durable governance layer (the decisions, versioning, provenance, autonomy, and drift) and the surfaces that drive it.

For voice changes too consequential to make in place, such as banning a term across the workspace or swapping a preferred term, the Context hub bundles the voice-rule edits into a reviewable change-set whose blast radius is measured before anyone approves it.