Release Assistant Governance: Inputs, Outputs, No Vibes
What this is This node documents the governance version of the Sugartown Release Assistant: a deterministic, three-step pipeline that produces the same release artifacts every time, regardless of which AI agent (or human) runs it.
No vibes. No inference. No “the bot thought it was implied.”
Canonical Release Pipeline
STEP 1 — Collect Reality Input: Human notes, AI agent memory Output: Messy but human-verified bullet lists of completed work
STEP 2 — Normalize Reality Input: STEP 1 Source of Truth Output: Canonical, deduplicated, outcome-only bullet list (Mechanical reduction only. This step is allergic to creativity.)
STEP 3 — Package the Release Input: STEP 2 normalized list (and nothing else) Output: Production-ready release artifacts
If any step fails, the pipeline stops. This is not a bug; it is the point.
Human vs AI Responsibilities
Humans decide what is true. AI reduces, formats, and packages that truth.
AI is never allowed to invent work, infer intent, or “helpfully” improve reality.
Standard Release Outputs
Every successful release must produce the following artifacts. These are contracts, not suggestions.
Enforcement Rules (Short, Sharp)
CHANGELOG.md is the single source of truth. README files are contracts, not commentary. Documentation marked “Updated” requires visible markdown evidence. STEP 3 may never revisit STEP 1. Failure is an acceptable outcome.
Why This Exists
This pipeline emerged after several real-world releases exposed a common problem: governance that is correct but hard to execute is still fragile.
The reset reduced the executable Release Assistant from ~12 printed pages to ~4, not by deleting rules, but by indexing them properly. The long-form document remains the law. This pipeline is how the law is enforced.
Competitive bots, cooperative system. Everyone behaves better when the rules are boring and explicit.