Skip to content

Latest commit

 

History

History
63 lines (46 loc) · 2.58 KB

File metadata and controls

63 lines (46 loc) · 2.58 KB

ADR-0003: Separate topic measurement from agenda generation

Status: Proposed

Date: 2026-08-09

Decision owner: Naruon maintainers

Capability maturity: target and future agenda capability PLANNED; no implementation is authorized

Scope: a possible future Naruon agenda-generation decision only. This ADR does not govern a model or generation provider, assign external ownership, or record provider acceptance.

Trigger for acceptance: a separately reviewed agenda-generation product contract and implementation PR.

Related requirement: TI-REQ-009

Context

The removed meeting_agenda_generator mapped words directly to a fixed agenda template. That coupled a lexical trigger, an implied topic assertion, and a generated action artifact. Even a valid fitted topic posterior would be descriptive evidence, not authorization to create or execute an agenda.

Proposed decision

ADR-0001 already supplies the accepted Naruon-local separation policy. This ADR is the proposed implementation decision for a future bounded agenda capability; it remains a proposed target rather than accepted architecture. If Naruon reintroduces agenda generation, the Naruon capability must:

  1. consume tenant-authorized source evidence and, optionally, a versioned topic posterior by reference;
  2. preserve every cited source and model provenance field without converting a display label into numeric topic identity;
  3. declare the generation provider/model and return review_required=true;
  4. treat source text, labels, and posterior metadata as untrusted data rather than instructions; and
  5. create no calendar/task/provider write unless a separate explicit intent, capability, consent, and conflict check succeeds.

The generator may abstain or fail, but it must not fall back to a template and describe that output as source-backed topic inference.

Consequences

  • A statistical posterior can inform a draft but never authorizes a write.
  • Human-readable labels are presentation metadata and can be revised without changing the fitted topic identity.
  • Agenda quality, grounding, prompt-injection resistance, and provider-write safety require tests independent of topic-model validation.

Alternatives rejected

  • One endpoint for measurement and generation: obscures error ownership and makes a generative failure look like scientific inference.
  • Template fallback: recreates the misleading behavior removed by PR #1297.
  • Direct provider write: bypasses Naruon's source, consent, capability, and conflict boundaries.