Actuary

Reference note

CI evals green, production worse: the missing quality floor

Your pull request gate was green. Weeks later your users tell you the agent has got worse. Both things can be true at once, because offline evals and production monitoring are measuring different things — and this note explains what each of them actually tells you.

The short version: keep the gates you already have, whether that is Braintrust, LangSmith or something you wrote yourself. Then add one instrument that keeps watching after the code has merged.

CI evals

What CI evals are for

  • Candidate cleared a known bar
  • Block merge
  • Compare candidates

Weak at:

  • Served quality after live
  • Provider/aggregator changes without a PR
  • Gradual drift
  • Continuous peeking over weeks

Production

What production monitoring is for

Live route above pre-committed floor on frozen panel (or POSTed scores).

  1. Freeze
  2. Score
  3. E-process alarm at E ≥ 1/α
  4. Slack/email package

Public version: Watchtower. Quiet cells published on purpose.

Both true

Why both can be true

Scenario · which instrument sees it

Scenario CI Production monitor
Bad prompt merged Should fail if covered May alarm after ship
Upstream model revision, same API name Often invisible Route-level panel drop
Tool schema drift in prod only Missed Caught if scores hit it
Eval set overfit in CI False green Live panel can disagree

The floor

The quality floor is the contract

Floor + α + panel/suite version hash-committed.

FAR caveats including independence and the 56.38% block-5 dependence stress: methodology and LLM false alarm rate under continuous peeking.

Handover

Where eval studios stop and monitoring starts

  1. Experiment/compare
  2. PR gate
  3. Traces
  4. Gateway
  5. Quality floor (Actuary)

Complement, not replace.

If pain is missing spans, buy a tracer. If pain is CI green + users found regression, you are missing the floor.

Probes

Hosted probes vs your own scores

Hosted probes

Five families, deterministic checkers, no judge model.

Your scores

POST /v1/scores in [0, 1]; descriptive unless registered.

Neither removes the need to choose a floor.

Operating split

Minimal operating split

  1. Keep CI suite on critical path
  2. Freeze one production monitor on the failure you fear
  3. Wire Slack to latched alarms only
  4. Review quiet weekly
  5. New epoch when changing panel/floor/α

Study 001 = one trajectory, not a customer logo story.

Boundary

What this is not

Not claimed

Not “evals are dead.” Not “replace Braintrust.” Not SOC 2. Not a guarantee every silent failure becomes a public incident.

Next step

Next step

Start Watch free

How it works · Watchtower.