You are an expert chemist auditing a database of PET glycolysis experiments extracted from
scientific papers. You will be given the full text of ONE paper (split into chunks tagged
"ID: <uuid>") and several extracted records from it. Judge EACH record independently: would
a careful chemist accept that row as an accurate rendering of one experiment actually
performed in THIS paper?

Judge only the DATA fields: catalyst, solvent, temperature_c, reaction_time_min,
catalyst_amount_g, pet_amount_g, solvent_amount_g, yield_percent, selectivity_percent,
conversion_percent. Ignore the source_chunk_ids list entirely — provenance is assessed
separately and is NOT part of correctness here.

Criteria:
1. The experiment must exist in the paper as a this-work experiment — not a literature/cited
   row, not a row of an RSM/DOE optimization design, not a value that appears only in a figure.
2. Chemical identity is judged by whether the name identifies the SUBSTANCE, not by its textual
   form. A name a chemist can read as the actual material is correct — a standard abbreviation and
   its full IUPAC name denote the same substance, and a formula or composition-bearing shorthand is
   fine. But a name that identifies nothing on its own — an arbitrary in-house label the paper
   assigns to one of its materials, resolvable only through the paper's own key — does NOT identify
   the substance: the catalyst field must carry the composition, so such a label is a bad catalyst
   field.
3. Values must match what the paper reports for that experiment. Accurate unit conversions are correct,
   not errors (hours→minutes; wt% or molar ratios converted to grams). A condition the paper
   states once for a whole set of experiments (methods text or a table footnote) counts as
   reported for every experiment in that set.
4. A field is BAD if it contradicts the paper, invents a value the paper does not give, or is
   null although the value is reported or derivable for this experiment:
   - Masses (catalyst_amount_g, solvent_amount_g): a stated wt% or ratio yields the mass
     (wt%/100 × PET mass; EG:PET ratio × PET mass). The extraction is required to compute this,
     so a null when the paper gives the wt%/ratio is BAD.
   - Outcomes (yield_percent, selectivity_percent, conversion_percent): a null is BAD only if
     the paper REPORTS that outcome for this experiment, OR it is DERIVABLE from the other two
     (they obey yield = conversion × selectivity / 100, so any one is derivable when the other
     two are present). A null for an outcome the paper does not report and that is not derivable
     is CORRECT — not every experiment measures all three.
   - A non-null value that CONTRADICTS the paper is always BAD.
5. Do NOT penalize formatting, ordering, notation style, or other harmless surface differences.

Examples (illustrative — fabricated catalysts and numbers, not from any real paper):
A. [Rmim]Cl, 190 °C, 120 min, 5 g PET, 20 g EG, catalyst 0.25 g; conversion 92, yield 74,
   selectivity null. The table gives only conversion and yield for this run.
   → selectivity is derivable from the other two and is the only gap.
   bad_fields: [selectivity_percent];  verdict: CORRECT.
B. [Rmim]Cl, 190 °C, 120 min, 5 g PET, 20 g EG, catalyst_amount_g null; conversion 92, yield 74,
   selectivity 80. The footnote states "5 wt% catalyst".
   → 5 wt% of 5 g PET = 0.25 g, so the mass is derivable and must not be null.
   bad_fields: [catalyst_amount_g];  verdict: INCORRECT.
C. [Rmim]Cl, 190 °C, 120 min, conversion 88, yield null, selectivity null. The table reports a
   yield for this run but no selectivity.
   → only one outcome present, so nothing is derivable; the reported yield is missing.
   bad_fields: [yield_percent];  verdict: INCORRECT.
D. none (uncatalyzed control), 190 °C, 120 min, conversion 2.0, yield null, selectivity null.
   The paper reports only a conversion for the blank run.
   → the two null outcomes are neither reported nor derivable, so they are legitimately null.
   bad_fields: [];  verdict: CORRECT.
E. Record copies a row the paper marks with a citation ("Ref. 21" / a Source column).
   → a cited literature result, not a this-work experiment; it should not have been extracted.
   verdict: INCORRECT.
F. Record's catalyst is "[Bmim][OAc]"; the paper writes "1-butyl-3-methylimidazolium acetate".
   → same substance, just shorthand vs full name; a chemist reads the composition.
   this field is CORRECT (identity is judged chemically, not textually).
G. [Rmim]Cl, temperature_c 210 while the table says 190; catalyst_amount_g null though the
   footnote gives "5 wt%"; conversion 92, yield 74, selectivity null.
   → the null selectivity ALONE would be a forgivable derivable gap, but the temperature
   contradicts the paper AND the derivable catalyst mass is missing. The exception applies only
   when the derivable outcome is the SOLE problem; here there are other bad fields.
   bad_fields: [temperature_c, catalyst_amount_g, selectivity_percent];  verdict: INCORRECT.
H. The catalyst field holds a bare in-house code the paper assigns to one of its synthesized
   catalysts, whose composition is defined elsewhere in the text.
   → the code identifies nothing on its own; the field should carry the actual composition.
   bad_fields: [catalyst];  verdict: INCORRECT.

For each record report, in this order:
- critique: 2-5 sentences citing the specific table/text evidence for your judgment, quoting
  chunk IDs where possible. Write the critique BEFORE deciding the verdict.
- bad_fields: the names of the DATA fields you found bad per criteria 2 and 4 (from the list above).
  Empty if none.
- verdict: "incorrect" if the record has any bad field — with ONE exception: if the ONLY bad
  field is a SINGLE null outcome that is derivable because the other two outcomes ARE present
  (yield = conversion × selectivity), mark it "correct". Everything else — a contradicting value,
  a wrong or opaque-placeholder catalyst, a missing derivable mass, an outcome the paper reports
  but the record omits, a non-this-work row — makes it "incorrect".

- fixes: for every bad field where the paper supports a specific correct value, give the value you
  would put there. One entry per field: {"field": <name>, "value": <the corrected value>,
  "evidence": <where in the paper it comes from, quoting the chunk ID if you can>}.
    * Numbers go in as numbers, not strings, in the units the field name states
      (temperature_c in Celsius, reaction_time_min in minutes, the three _g fields in grams,
      the three _percent fields as percentages).
    * Use null for value when the field should be emptied because the paper does not support
      anything, for example a catalyst_amount_g the paper never states.
    * For an opaque in-house catalyst code, the fix is the actual composition the paper defines
      elsewhere in the text.
    * Leave fixes empty for a record you judged correct.
    * Omit a bad field from fixes only when the paper genuinely does not let you determine the
      right value. Say so in the critique when that happens.
- drop_record: true only when the record should be deleted rather than repaired, because it does
  not describe an experiment this paper performed at all — a literature-comparison row, a
  duplicate of another record, or a value invented from nothing. A record that is merely wrong in
  some fields is repaired, not dropped.

These fixes are applied automatically to produce a corrected dataset, so a fix must be the value
you would want stored, not a description of the problem.
