Skip to content

The PSR method

A review should expose
its assumptions.

Start with the consequential decision. Trace the operating boundary. Connect each important failure to a control, a test, and a condition for stopping.

See a complete example →Review a system →
  1. 01

    Architecture review

    Frame the decision, the consequence of failure, and the most important concern before release.

  2. 02

    Assumptions and unknowns

    Separate stated facts from missing evidence. A target is not a measured result; a suggested control is not an existing one.

  3. 03

    System boundary

    Trace the stated inputs, processing, outputs, and human intervention points. Keep proposed components outside the current boundary.

  4. 04

    Priority failure modes

    Prioritize at most six failures. For each: affected party, qualitative likelihood and severity, control, evaluation, and stop condition.

  5. 05

    Evaluation plan

    Turn the concerns into tests, evidence to collect, and acceptance criteria. Label new thresholds as proposed and needing calibration.

  6. 06

    Rollout and monitoring

    Define a bounded release, what to observe, who can intervene, and how to stop or roll back when a control fails.

  7. 07

    Open questions

    Leave the unresolved decisions visible, with the evidence and accountable people needed to close them.

A consistent evidence standard

Stated. Unknown.
Proposed.

Visitor-supplied facts are labeled Stated. Missing details are Unknown. New components, controls, metrics, and processes are Proposed. Uploaded documents are unverified context, not independent evidence.

Risk is discussed with qualitative likelihood and severity. There is no numerical readiness score. Unspecified thresholds need calibration against the real system and the people affected.

Generated analysis.
Accountable decisions.

In the workspace

The current workspace uses Qwen-based models with PSR’s review instructions and structured intake. Noether is a planned specialist model with no published evaluation results yet.

Your first submission opens a fixed operating-boundary intake without model inference. Your next reply uses the existing review session to request an AI-generated review. Unstated details remain unknown or proposed.

Before displaying a full review, the server checks its format, required sections, output limits, and required failure-mode fields. These checks do not verify factual correctness or prove that the system is ready to ship.

With Prabhjot

Prabhjot Singh Rai is the founder of Production Systems Review. He does not automatically read, approve, or sign the workspace’s output. Human review is a separate conversation about the system, supporting evidence, and unresolved trade-offs.

Generated guidance is a first pass, not certification or independent assurance. The people responsible for the system own the release decision.

Discuss a review with Prabhjot →