Client-Facing Eval Scorecard Handoff (Handoff)

Client-Facing Eval Scorecard Handoff (Handoff)

1. 01

   Freeze the eval suite at the version the client signed off on

   Tag the dataset, judge prompts, and scoring rubric with a release number so the receiving team can reproduce any score the client saw during the pitch.

2. 02

   Package the scorecard as a one-page artifact the client can forward internally

   Include pass rate, p95 latency, cost per resolved conversation, and the three failure modes that remain open. Clients forward this to procurement; make it self-contained.

3. 03

   Transfer trace access with scoped credentials, not shared logins

   Issue read-only seats in the observability platform (Langfuse, Braintrust, or Arize all support role-scoped API keys) and document who holds admin.

4. 04

   Record the judge calibration session on video and hand over the recording

   When a human reviewer overrides an LLM judge, the reasoning is the transferable asset. A 20-minute recording prevents the next team from re-litigating scoring thresholds.

5. 05

   Write the drift alarm thresholds into the retainer scope document

   State the numeric trigger (for example, pass rate drops 8 points week over week) and the response time the agency commits to. Unwritten thresholds become disputes.

6. 06

   Schedule a 30-day joint review before the handoff meeting ends

   Book it on the calendar during the call. Handoffs without a scheduled follow-up decay into support tickets within six weeks.