Production Recap 5 · Proving it: golden sets, judge bias and trajectory eval

[akash-coded]

Exercise 5: Proving it, validation and QA

Language: Python, concept, diagrams Topics: golden set and scoring, LLM-as-judge bias, trajectory eval, red-team, cost and TRIM, the go/no-go gate Level: applied (predict, trace, debug)

Fifth foundation. You predict eval scores, trace judgement, and fix broken checks. Predict-output answers are the exact printed result.

Q1. Predict the exact output.

def score(ans, checks):
    r = {}
    if "grounded" in checks:    r["grounded"]    = 1 if "[source:" in ans else 0
    if "in_scope" in checks:    r["in_scope"]    = 0 if "ceo" in ans.lower() else 1
    if "has_options" in checks: r["has_options"] = 1 if "6E-" in ans else 0
    return round(sum(r.values()) / len(r), 2)

print(score("Free rebooking. Options: 6E-114. [source: fare-rules]", ["grounded","in_scope","has_options"]))
print(score("The CEO is fine. Options: 6E-114.", ["grounded","in_scope","has_options"]))

• A) 1.0 then 0.33
• B) 1.0 then 0.67
• C) 0.67 then 0.33
• D) 1.0 then 0.5

Show answer

A) First passes all three. Second fails grounded (no [source:) and in_scope (contains ceo), passes has_options, so 1 of 3 is 0.33.

Q2. Predict the pass rate.

GOLDEN = [
    {"ans": "[source:x] options 6E-114", "checks": ["grounded","has_options"]},
    {"ans": "options 6E-114",            "checks": ["grounded","has_options"]},
    {"ans": "[source:x] no options",     "checks": ["grounded","has_options"]},
]
def s(ans, checks):
    r = []
    if "grounded" in checks:    r.append(1 if "[source:" in ans else 0)
    if "has_options" in checks: r.append(1 if "6E-" in ans else 0)
    return sum(r) / len(r)

print(round(sum(s(c["ans"], c["checks"]) for c in GOLDEN) / len(GOLDEN), 3))

• A) 0.5
• B) 0.667
• C) 0.833
• D) 1.0

Show answer

B) Case scores are 1.0, 0.5, 0.5. Mean is 2 of 3, which prints 0.667.

Q3. A judge model keeps scoring longer answers higher regardless of correctness. This bias, and its guard, are:

• A) self-preference, use a judge from a different model family
• B) rubric drift, anchor each score to a concrete example
• C) verbosity bias, make the rubric reward grounding, not length
• D) ungrounded judging, give the judge the source text to score against

Show answer

C) The symptom is length. The other three are real judge traps, but not what is happening here.

Q4. This trajectory checker passes a wrong-order path. Predict its output, then read the fix.

def check(actual, expected):
    return set(actual) == set(expected)
print(check(["lookup_booking", "get_rebooking_options", "get_disruption_reason"],
            ["lookup_booking", "get_disruption_reason", "get_rebooking_options"]))

• A) it prints False, and the fix is to compare lengths as well
• B) it prints True, and the intended fix is to sort both sequences before comparing them
• C) it prints False, and the code is already correct
• D) it prints True, and the fix is actual == expected to respect order

Show answer

D) set() throws away order, so a scrambled path passes. Comparing the ordered lists (actual == expected) is the fix. Sorting would also lose the order, so it is not the fix.

Q5. Haiku scores below the acceptance bar; Sonnet clears it. The program concludes, and the switch is cheap because:

• A) ship Sonnet, the eval decided it, and LiteLLM makes the swap a one-string change
• B) ship Haiku with extra guardrails, since eval scores are only a guide
• C) run both and route by question difficulty to save on cost
• D) re-run the eval until Haiku finally passes, since nine cases is too few to trust anyway

Show answer

A) The eval, not preference, picks the model, and the model is a single swappable string. Re-running until it passes is gaming the test.

Q6. An agent gives the right rebooking answer, but the trace shows it never called lookup_booking. Why does this fail a trajectory check?

• A) a skipped tool means the final answer must be wrong
• B) it assumed the tier, so it will fail on a different passenger
• C) trajectory checks require every tool to be called in strict order
• D) the skipped call means the answer was not grounded in policy

Show answer

B) It got lucky by assuming the tier. Change the passenger and the same path returns the wrong answer.

Q7. A red-team prompt defeats a guardrail during testing. Which are the right follow-ups? (select all that apply)

• A) patch the guardrail so the attack no longer gets through
• B) add the failing case to the golden set so it is regression-tested forever
• C) remove the targeted feature until the model improves
• D) lower the guardrail threshold so it triggers earlier on everything

Show answer

A and B. Patch it, then make the failure a permanent case. Removing the feature or blanket-lowering the threshold trades one problem for another.

Q8. Complete the validate pipeline. Bank: a) all bars cleared b) a bar failed c) in progress d) timed out

flowchart LR
  B["New build or prompt change"] --> E["Golden set + trajectory"]
  E --> R["Red-team"]
  R --> C["Cost and latency"]
  C --> G{"Gate"}
  G -->|"blank 1"| S["Sign-off, ship"]
  G -->|"blank 2"| H["Hold, log failures as cases, fix"]
  H --> B

Loading

Show answer

blank 1 = a (all bars cleared, ship), blank 2 = b (a bar failed, hold and loop back).

Q9. Match each LLM-as-judge trap to its guard. Bank: a) use a judge from a different model family b) reward grounding, not length c) anchor each score with a concrete example d) give the judge the source and score only agreement

1. Self-preference
2. Verbosity bias
3. Rubric drift
4. Ungrounded judge

Show answer

1 = a, 2 = b, 3 = c, 4 = d.

Q10. This scope scorer lets an off-scope answer pass. Which line is wrong?

1  def in_scope(ans):
2      bad = ["ceo", "raw record", "every pnr"]
3      return 1 if any(w in ans.lower() for w in bad) else 0

• A) line 2, the bad list is missing entries
• B) line 1, the function should also take the full list of checks as a second argument
• C) line 3, the 1 and 0 are inverted; an off-scope answer should score 0
• D) line 3, any should be all

Show answer

C) As written, containing a banned phrase returns 1 (pass). Flip it: return 0 when a banned phrase appears.

Q11. True or False: an answer that is correct but reached by the wrong path still passes a trajectory check, because trajectory only cares about the final answer.

• A) True
• B) False

Show answer

B) False. Trajectory asserts on the path. A right answer by the wrong path is a latent failure that fails the check.

Q12. Match each scorer line to its effect. Bank: a) rewards a cited source b) fails an off-scope answer c) turns the checks into a partial-credit fraction

1  r["grounded"] = 1 if "[source:" in ans else 0
2  r["in_scope"] = 0 if "ceo" in ans.lower() else 1
3  return sum(r.values()) / len(r)

Show answer

1 = a, 2 = b, 3 = c.

Q13. The golden set passes and guardrails hold, but cost is over budget. Everything else is green. The verdict is:

• A) NO-GO, a bar failed
• B) GO, cost is a soft concern
• C) re-run the whole suite before deciding
• D) CONDITIONAL, ship with a stated limit

Show answer

D) A cost-only miss with everything else green is CONDITIONAL: ship with a named limit, for example scope reduced or approval kept on.

Q14. Which set of levers is TRIM, the program's cost checklist?

• A) Tier the model, Reuse context via cache, move Idle work to batch, Minimize context
• B) Tune, Retry, Index, Merge
• C) Trace, Rank, Interrupt, Monitor
• D) Trim the prompt tokens, Rotate the access keys, Isolate every tool, and Mock all the inputs

Show answer

A) Tier, Reuse, Idle-to-batch, Minimize. Sonnet where the eval demands it, Haiku on the cheap steps, cache the policy that repeats.

Q15. True or False: LLM-as-judge scores should be treated as evidence and spot-checked against human labels, not accepted as final verdicts.

• A) True
• B) False

Show answer

A) True. A judge is a force multiplier and a liability in one tool. Treat its scores as evidence, and audit a sample.