Evaluations and datasets · Lesson 8 of 18 · 5 min

Compare evaluation runs

Controlled comparisons require the same cases and scoring rules. Separate an actual model/prompt improvement from a change in the evaluation itself.

Open the lab Course map

Mastery 0/3Run itExplain itApply it

learner@clairos:~ $

↑ history · Tab complete · Ctrl+C cancel · Esc, Tab to leaveExit 0

Your objectives

Hints are always available. Wrong turns never cost XP.

1
Read the case evidence
Read /opt/ai-lab/cases/aiops-evals-datasets-008.json. Separate the recorded facts from your assumptions.
Hint & worked example
Controlled comparisons require the same cases and scoring rules. Separate an actual model/prompt improvement from a change in the evaluation itself.
The complete case JSON is read successfully.
```
cat /opt/ai-lab/cases/aiops-evals-datasets-008.json
```
Common mistake to avoid: Different scoring rules produce directly comparable scores by default.
2
Isolate the deciding signal
Select the candidate_scorer field from the case JSON. What does it establish—and what does it not establish?
Hint & worked example
Align the scorer versions before attributing the score difference to the candidate.
A successful JSON query isolates candidate_scorer.
```
jq .candidate_scorer /opt/ai-lab/cases/aiops-evals-datasets-008.json
```
Common mistake to avoid: Using a different case or an unrelated signal as the justification.
3
Record your operational decision
Choose the justified code: align-scorer, change-without-evidence, or ignore-the-signal. Save that one code followed by a newline in /home/learner/lab/aiops-evals-datasets-008-decision.txt.
Hint & worked example
Align the scorer versions before attributing the score difference to the candidate.
The decision artifact contains the correct code, not just a matching command in history.
```
printf '%s\n' 'align-scorer' > /home/learner/lab/aiops-evals-datasets-008-decision.txt
```
Common mistake to avoid: Different scoring rules produce directly comparable scores by default.

Understand the pattern

Controlled comparisons require the same cases and scoring rules. Separate an actual model/prompt improvement from a change in the evaluation itself.

Interpret the evidence

Align the scorer versions before attributing the score difference to the candidate.

A tempting mistake

Different scoring rules produce directly comparable scores by default.

All names, measurements, prices, versions, and thresholds in this exercise are fictional training data. The evidence supports this bounded scenario, not a claim about a live provider.

Worked example and expected result

$ cat /opt/ai-lab/cases/aiops-evals-datasets-008.json

{
  "fixture": "Fictional training data; not a live measurement.",
  "baseline_dataset": "gold-v4",
  "candidate_dataset": "gold-v4",
  "baseline_scorer": "v2",
  "candidate_scorer": "v3"
}

Exit status: 0. The command completed successfully in the training fixture.

$ jq .candidate_scorer /opt/ai-lab/cases/aiops-evals-datasets-008.json

"v3"

Exit status: 0. The command completed successfully in the training fixture.

$ printf '%s\n' 'align-scorer' > /home/learner/lab/aiops-evals-datasets-008-decision.txt

(no output; exit 0)

Exit status: 0. The command completed successfully in the training fixture.

Try a variation

Name the assumption that would have to change for “align-scorer” to stop being the justified action. Do not silently replace missing evidence with a guess.

Explain it

What does the evidence in “Compare evaluation runs” justify?Align the scorer versions before attributing the score difference to the candidate.Different scoring rules produce directly comparable scores by default.Make the change now and gather the missing evidence afterward.

Apply it elsewhere

A colleague concludes: “Different scoring rules produce directly comparable scores by default.” What should you do?Accept it because the training command completed.Treat this fixture as a live measurement.Reject that conclusion; return to the case evidence and its stated limits.

Keep the momentum.

Complete the lab, then explain and apply the pattern.

Next: Set a release threshold Try an incident drill

Anonymous lesson and mastery progress lasts for this session. Sign in to keep synced lesson XP between sessions; mastery stays device-local. No speed bonus, lost lives, or hint penalties.

Keep learning

Case methodology: evidence, limits, decision