Scheduled digests
The digest nobody wants to write.
A recurring summary with a stable shape and a fixed source of record is the easiest work to learn — and among the first worth moving.
Anatomy of the task · illustrative workflow
The week's incidents, with owners, status and follow-ups.
Input
The tracker, the reporting period, and the digest format your team already reads.
Required output
The period's incidents with owners, status and follow-ups — each item traceable to the system of record.
Evaluation
- Covers every incident in the period, and none outside it
- Each item stays traceable to the tracker
- The format matches what the team already reads
Understand the workflow.
Scheduled work is the purest recurring pattern: same source, same shape, same cadence. Nothing about it needs frontier reasoning once the shape is written down.
Define the quality bar.
The gate scores each digest against your current model's digest for the same request. It checks goal completion, correctness and completeness, not style.
Improve the route.
Operant learns the digest's skill from past runs that weren't marked as failures, and proposes the route. In shadow, the learned rule is visible and priced, and it waits until you ratify it.
Keep frontier capability.
Escalate to the frontier model when the period contains something genuinely novel — a new incident class, a cross-system failure — and let the cheap route handle the routine weeks.
Published evidence
See the comparison. Keep the context.
The published gate result documents the method this pattern is governed by: shadow first, a held-out gate, and a human ratify before live traffic.
Read the results ->Who this fits
A task your team knows well.
A team with scheduled summaries on a budget: incident digests, spend digests, weekly reviews — anything with a fixed source and a stable shape.
Bring a workload ->Another place to start · Compliance evidence packs