Skip to content

0025. Three independently certified AI pillars

  • Date: 2026-09-02
  • Status: accepted

Context

Loomarr uses models for three different jobs: recommending channel concepts, curating a chosen Intent into a Proposal, and classifying filler clips. They share infrastructure and may share a model family, which invited treating one passing scorecard as proof for all three.

Decision

Model work is three pillars with separate authority and release contracts:

  1. Channel recommendation proposes Channel Concepts (a name, a draft Intent, cited context signal ids). A Concept is not a Proposal and carries no effectful field.
  2. Channel curation turns an Intent into a grounded Proposal and ChannelPolicy.
  3. Filler curation classifies, tags and ranks clips from bounded evidence; admission and scheduling authority stay with deterministic filler policy.

Each pillar owns its own corpus, thresholds, scorecard and ship decision. Certification never transfers between pillars, even for the same weights.

Consequences

  • Corpora are versioned and frozen before live inference; development corpora must be disjoint from certification holdouts, and a prompt or schema change needs a new untouched holdout.
  • Recommendation certification uses synthetic, digest-pinned snapshots, never household identity or viewing history.
  • Frozen holdouts and their no-ship scorecards stay immutable; channel-recommendation-v2 replaced v1 with new, mechanically disjoint cases rather than editing it.