localgate/conversion-audit

Conversion audit — rubric for annotators

This is the audit the paper reports. Three annotators, 248 items, independent, 2-of-3 majority, Fleiss' κ reported. It measures the error rate of the conversion filter, which is the instrument the whole open-ended corpus rests on.

There are two passes, and the order matters.

1. annotate_<yourname>.json — 248 items, the original question only. Do this first.
2. annotate2_<yourname>.json — 152 items, with the rewrite shown. Do not open this until pass one is completely finished, because it reveals which items the filter kept, and pass one depends on you not knowing that.

The annotate* files are read-only reading packages — they are gitignored, so anything typed into them can be lost. Type your answers into verdicts_<yourname>.json (pass one) and verdicts2_<yourname>.json (pass two) instead: they list the same items, by blind id, in your reading order, and they are tracked in git. Commit them as you go.

Do not open KEY.json, and do not discuss items with the other two until both verdicts files are done. Disagreement is the measurement; agreeing in advance destroys it.

Your rows are in a different order from the other annotators' — that is deliberate, so that tiredness late in the file doesn't hit the same items for all three of us. Items are identified by id, never by position.

---

Pass one — does the item convert?

You are shown the original MMLU-Pro question, its options, and which answer is correct — the same view for every item. You are not told what the filter decided, whether a rewrite exists, or what the other annotators thought. Every item looks identical, deliberately.

1. converts — would this question survive losing its options?

Imagine the options deleted and the question reworded to stand alone. Would a knowledgeable person give the reference answer, rather than a different, equally correct one?

• yes — the reference is the single correct answer.
• no — stripped of options it admits many correct answers, or stops meaning anything.
• borderline — you genuinely cannot decide. Use this sparingly.

Two kinds fail whatever else is true of them:

• It excludes rather than identifies — "which is NOT", "EXCEPT", "is false", "least likely". Everything in the world outside the answer set answers it correctly.
• The answer only means something beside the others — "all of the above", "both A and C". Exception: if the stem itself sets out the alternatives ("i) … ii) … iii) …"), an answer like "i and iii" names real things and is fine.

Common failures: asking for one member of a large category ("an example of X", "a true statement about X"); ranking the options ("best", "closest", "greatest"); an answer true among these ten choices but false as a claim about the world; leaning on a unit or convention only the options supplied.

A superlative is not automatically a ranking — ask what it compares. "Which is the most persuasive argument", with ten arguments supplied, ranks the list and fails. "Which best approximates the ratio of nonterminal to total nodes in a complete K-ary tree", answer "1/K", does not: the ratio has one value and "best approximates" hedges the rounding. If the stem itself fixes the quantity, the superlative has nothing left to rank.

Not failures: asking for several things at once ("calculate the efficiency and the reheat factor") — judge whether the answer is unique, not whether it is simple. Working the answer out from figures the stem supplies. Clumsy phrasing, which a rewrite fixes. And the reference answer's wording does not matter here — a later stage accepts any phrasing of the same fact.

2. answer_stands_alone — does the reference work as a grading target?

Judge this independently of question 1. An item can have a perfect question and an unusable answer; that combination is the reason this field exists.

A grader will see the question, this reference answer, and a model's response, and must decide whether the response is correct.

• yes — the reference is a usable target.
• no — it is not, for one of these reasons:
  • units or magnitude stripped — 1.12 where the question asks for a speed, and the real answer is 1.12 × 10⁴ m/s. Fine as one of ten numeric options; meaningless alone.
  • a fragment — greater and grander, which only completes the stem it came from.
  • it does not answer the question asked — the question says "describe the differences" and the reference is a single word like Alcoholism.
  • it names option labels — "I and II only", "b and c are Hermitian".
• borderline — as above, use sparingly.

MMLU-Pro option texts were written to be told apart from nine alternatives, not to stand alone. That property does not survive conversion, and nothing in the pipeline checks it — which is why we are asking you.

3. why — one short clause

What decided it. Half a line is enough: "many true statements about X", "units stripped, answer is 1.12e4 m/s", "unique: only one canonical answer". These are read when the three of us disagree, so write the reason you would give to the other two.

Worked examples

question	reference	converts	answer_stands_alone
"Which of the following is NOT a phase of matter?"	plasma	no — excludes	yes
"Calculate the minimum muzzle speed for a shell to escape Earth…"	1.12	yes — determinate	no — units stripped
"Which of the following statements about photosynthesis is true?"	(a true statement)	no — large category	yes
"The boiling point of water at sea level is"	100 degrees Celsius	yes	yes
"According to Paley, what is the key difference…"	greater and grander	yes	no — fragment
"Which of these are noble gases? i) helium ii) nitrogen iii) argon"	i and iii	yes — stem lists them	yes

---

Pass two — is the rewrite faithful?

Only after pass one is finished. You now see the original question, its options, the reference answer, and the rewritten open-ended question the filter produced.

Pass one asked whether the item could convert. This asks whether it actually did.

1. same_question — does the rewrite ask what the original asked?

• yes — a reader answering the rewrite would give the same answer they'd have given the original.
• no — it drifted. Two ways this happens and both matter:
  • it was written from the answer — the rewrite names a topic, property or scenario the stem never mentioned, narrowing a vague question until the reference is the obvious answer. "Which of the following is true?" becoming "Which statement about compact and complete spaces is true?" is the type case.
  • it makes the question easier or harder than the original.
• borderline.

2. self_contained — can it be answered with no options in view?

• yes — everything needed is present.
• no — it still points at a list that is gone ("which of the following", "the statements above"), or context was dropped: a passage, table, formula or case description the original supplied and the rewrite left out. Watch for items that began "This question refers to the following information" — if the passage is missing, the question is unanswerable however well it reads.
• borderline.

3. why — one short clause, as before.

---

Why the sample looks the way it does

248 items drawn from five strata — kept by the judge, kept at re-score, dropped at re-score, dropped at rewrite, and a borderline stratum of items whose re-score landed in bands 4–7. The strata differ by orders of magnitude in size, and the small ones are over-drawn on purpose, so the raw agreement rate over these 248 is not the filter's error rate; each stratum is reweighted by its share of the corpus afterwards.

The borderline stratum exists for a specific reason. The filter readmits a re-scored item at 5 or above, a threshold inherited from the method we adapted. Reading the run's own stated reasons, items scored 5 and 6 concede that the answer is not unique about as often as items scored 4 — which are rejected — while items scored 8 and above almost never do. That suggests the cut is in the wrong place, but it was measured by pattern-matching the model's prose, which is not evidence anyone should act on. Your verdicts on these 50 items decide where the threshold goes. They deliberately include items the filter both kept and dropped, so please don't try to infer which is which.

The sample is not a picture of the corpus and you should not expect your answers to look like one. Do not track your own yes/no ratio: judge each item on its own and let the total land wherever it lands.