Release Intelligence
Deterministic classification replaced a black-box LLM step, cutting the unexplainable custom bucket from 40–60% to under 5%.
The model handles voice and format only. Python owns classification and grouping, verbatim narrative seeds constrain generation, and a proportional review gate routes the PM to the 2–3 items that matter.
Role
AI systems designerPipeline architect
Outcome
Five consistent outputs per cycle, a deterministic audit trail, and a workflow handed off for non-engineering ownership.
The workflow correction
Facts left the model.Voice stayed.
The same Linear tickets had to become five different release documents. An all-LLM approach classified them inconsistently and hid 40–60% inside a custom bucket. The fix moved classification and grouping into deterministic Python while the model kept only voice and format. Verbatim seeds constrained claims, validation targeted known failures, and the custom bucket fell below 5%. Across 100+ tickets, PM work dropped from 3–5 hours to 20–40 minutes, with review focused on the 2–3 risky items.
At a glance
Sub-5% custom bucket · 20–40-minute cycle · 12+ batches across 100+ tickets
In the record
Pencil architecture → functional system · Six-stage pipeline diagram · Batch-by-batch error table · Deterministic classifier source · Narrative-seed validation contract · Proportional review queue
Move facts out of the model
Deterministic Python owns classification and grouping, so identical tickets produce identical, auditable results.
Constrain generation
Each audience has its own contract, but every claim must originate in a verbatim narrative seed from the source ticket.
Review proportionally
Validation routes PM attention to 2–3 high-risk items instead of demanding blanket review or pretending review is unnecessary.
