Skip to main content

Define promotion thresholds

Stub

This How-to is a stub. The v0.x JARVIS stack is still stabilizing; thresholds are planned but not yet enforced automatically.

Goal​

You will define clear promotion thresholds for experimental → candidate → stable capability channels.

When to use this​

  • You have baseline evaluation data and want consistent promotion decisions.
  • You want to reduce “it seems good” subjectivity.

Prerequisites​

  • A channel model (experimental, candidate, stable)
  • Evaluation metrics (deterministic + rubric + stability)
  • A decision process (human approval now; automation later)

Steps​

  1. Choose the metrics that matter (pass rate, stability, policy adherence).
  2. Set minimum thresholds per channel (start conservative).
  3. Record the thresholds as a decision record and apply consistently.

Verify​

  • Two reviewers using the same data reach the same promotion decision.
  • Promotions are reversible (rollback strategy exists).

Troubleshooting​

  • Thresholds too strict → start with candidate and tighten gradually.
  • Thresholds too loose → require higher k or stricter policy gates.
  • Mixed workloads → define thresholds per capability class (read-only vs write/irreversible).

Cleanup / Rollback​

  • Revise thresholds only via a versioned decision record.

Next steps​