ADR-PROD-009 - On-air UNI expert: a Claude persona, not uni-mind inference

• Status: Proposed
• Date: 2026-06-21
• Deciders: UNI Production architecture
• Master contract: docs/UNI_PRODUCTION_PLATFORM.md ("a non-infra decision - the on-air UNI expert")

Context

The show puts a recurring UNI expert on cam/voice to explain the mission's science (trauma and the nervous system, mental-health-to-nature alignment, etc.). The tempting move is to make that expert uni-mind's own inference. But uni-mind is research-stage: its serving surfaces have no shipped expert weights and present as a 4-pattern closed class (ingest map E, as captured this session). Putting an unshipped research model on air as "the UNI expert" would be an over-claim and a charter violation. We need an on-air expert that is honest about what UNI is and what it is not, and that cannot drift into AGI / "beats LLMs" / "memory" over-claims live.

Decision

The on-air UNI expert is a Claude persona, not uni-mind's own inference. The persona is seeded with:

• uni-mind/docs/press/02_FACT_SHEET.md - ground truth about what UNI is.
• uni-mind/docs/prompts/UNI_CHAT.md - the voice/tone contract.
• uni-mind/docs/press/05_CLAIMS_AND_FENCES.md - a hard compile-time lint on every on-air word: no AGI, no "beats LLMs," say "cache hit" not "memory," and UNI math stays private.

The persona speaks via the MCP narrate verb (Piper, ADR-PROD-006) and/or on cam. uni-deep-chat may appear only as a clearly-labeled on-screen microscope (surprisal / cache B-roll) - never as the talking expert. The claims-and-fences lint runs on script copy before air; on-air narration text is generated under the fences.

Alternatives considered

• uni-mind's own inference as the on-air expert. Rejected: research-stage, no shipped expert weights, a 4-pattern closed serving class - presenting it as a live expert would over-claim and breach the charter (and 05_CLAIMS_AND_FENCES.md). It is not deployable as a talking head.
• uni-deep-chat as the talking expert. Rejected as the expert: it is a microscope, useful as labeled B-roll (surprisal / cache visualization) but not a persona; using it as the expert would conflate a research instrument with an authoritative speaker.
• A generic unfenced LLM persona. Rejected: without the 02_FACT_SHEET ground truth + UNI_CHAT voice
  • 05_CLAIMS_AND_FENCES lint, the expert could drift into AGI / "beats LLMs" / "memory" over-claims live, which is exactly the failure mode the fences exist to prevent.

Consequences

• The on-air expert is honest by construction: it speaks UNI's approved public framing (human-owned loop, not "LLMs can never be safe"; restraint + peace close) under a compile-time fence, and never externalizes UNI's private math. Honest tradeoff: the persona is a Claude-over-MCP construct, not a UNI-native model; the show must label it as a persona (an editorial honesty requirement), and the fences-catch-all property is pending (a lint can miss novel phrasings until exercised).
• uni-deep-chat stays a labeled microscope, keeping the research instrument distinct from the on-air voice.
• Evidence class: "uni-mind has no shipped expert weights / 4-pattern closed class" is Class-C as captured this session; "the fenced persona never over-claims on air" is pending until exercised over a captured show.

Links

• Master: docs/UNI_PRODUCTION_PLATFORM.md
• Seeds: uni-mind/docs/press/02_FACT_SHEET.md, uni-mind/docs/prompts/UNI_CHAT.md, uni-mind/docs/press/05_CLAIMS_AND_FENCES.md
• Related: ADR-PROD-006 (narrate voice), ADR-PROD-005 (lower-third "UNI EXPERT" kicker; microscope B-roll overlay)

Status (honest)

This ADR is a design, status pending; nothing is deployed or claimed to run. No banned-unqualified word is used as a claim. "uni-mind is research-stage / no shipped expert weights" is Class-C as captured 2026-06-21; the on-air fenced persona's behavior is pending until exercised. UNI math stays private and is never externalized. The business stack (solutionwright-*, odoo, jitsi, cloudflared, portainer) is never a mutation target; the producer agent cannot self-approve.