THESIS

Working premise

An AI release review should let an accountable owner understand what changed, why it is better, what regressed, what it costs, and how to reverse it without replaying the entire development process.

CONTEXT

Operational context

The release review should let an accountable owner answer five questions quickly: what changed, what evidence supports the change, what became worse, what risk remains, and how the system will be monitored or reversed. A long experiment log is not a decision packet.

Evidence should be versioned and scoped to the released artifact. Model, prompt, retrieval, tools, data, evaluator, routing, and policy changes belong in one manifest. Results should show representative coverage, row-level material regressions, cost and latency deltas, reviewer decisions, and incomplete runs.

The review continues into production. Rollout stages, stop conditions, monitors, owners, and rollback targets should be approved with the candidate. Accepted risks need explicit expiry or follow-up so provisional decisions do not become permanent undocumented behavior.

INQUIRY

Questions to resolve first

Q1

Can a reviewer identify every behavior-affecting change from one manifest?

Q2

Which material regressions and unknowns remain after evaluation?

Q3

What evidence is incomplete, stale, or subject to evaluator disagreement?

Q4

Which production signal will stop, roll back, or advance the rollout?

PROTOCOL

Recommended method

  1. P1

    Summarize the change across model, prompt, data, retrieval, tools, and policy.

  2. P2

    Show baseline and candidate results with representative coverage.

  3. P3

    List material regressions, unresolved unknowns, and reviewer decisions.

  4. P4

    Include production rollout, monitor, and rollback criteria.

FAILURE

Common failure modes

Experiment dump

Reviewers receive many metrics and notebooks but no concise recommendation or decision boundary.

Missing negative evidence

Failed runs, weak segments, and unresolved cases are omitted from the packet.

Approval without operation

The candidate is accepted without rollout stages, monitors, owners, or a tested rollback.

CONTROL

Control points

  • Do not omit failed or incomplete evaluation runs.
  • Require named ownership for every exception and monitor.
  • Link the review to immutable versions and traceable experiment evidence.
  • Keep approval separate from the engineer who prepared the evidence when risk warrants.
DEPLOY

Implementation sequence

  1. S1

    Define a release manifest and evidence checklist appropriate to workflow risk.

  2. S2

    Generate a concise baseline-versus-candidate packet with coverage, material deltas, unknowns, and reviewer notes.

  3. S3

    Record approve, hold, exception, or rollback decisions with named authority and rationale.

  4. S4

    Link production monitors and follow-up dates back to the same release record.

MEASURE

Measures worth reviewing

Track evidence completeness, review time, open exceptions, rollback readiness, and post-release variance. A fast approval is not efficient if missing evidence creates escaped regressions or delayed incident response.

Evidence completenessOpen exception countApproval timeRollback readinessPost-release variance
SOURCES

Primary references

This note is an editorial synthesis of public standards, primary institutional guidance, and common engineering control patterns. It is intended to support engineering design and review. It is not a substitute for legal, security, safety, audit, or other professional advice for a specific system.