GDP.pdf Review Report

Abstract

GDP.pdf asks models to read PDFs and answer questions about business, engineering, research and other professional subjects. We reviewed the 100 publicly available tasks: 40 Accepted, 13 Doubtful and 47 Refused.

Most confirmed issues concern reference answers that conflict with the source or grading requirements absent from the prompt. Some criteria turn a rule against wrong answers into a ban on accurate context or explanations of excluded options, causing correct work to lose credit.

Scope and version

The report covers 100 tasks in the openly available copy. Fixed dataset version: 7f76578

The scoring implementation uses a fixed version of surge-ai/gdp-pdf: 7a72a51

Report updated on November 4, 2026.

Distribution of findings

Each task is counted once, under its primary issue category.

Representative cases

Task-by-task findings

Refused: 47 tasks

• gdp-002

  C2Scoring requirements not stated in the task

  The prompt asks only for the pages containing fee information. The rubric additionally requires all four fee categories to be restated. Correct page references with genuine examples can still fail this extra requirement.
• gdp-004

  C2Scoring requirements not stated in the task

  The transformer and part number 60536 were correct. An accurate additional reference to a 120V diagram on another page nevertheless caused the only failed check, leaving a mean score of 0.75.
• gdp-006

  C1Incorrect answers or scoring logic

  The reference assigns a PESI risk score of 138 to Class IV. Both the attached table and the original study place it in Class V. A correct classification conflicts with the grading target.
• gdp-009

  C2Scoring requirements not stated in the task

  The public prompt contains only an editing note, and the attachment supplies no question. The rubric nevertheless requires a hail-damage, storm-attribution and maintenance analysis for each of ten photographs. The core task was never communicated to the solver.
• gdp-013

  C2Scoring requirements not stated in the task

  The answer placed people with unknown fitness in a separate Unknown category rather than mixing them into the fitness gradient. Their presence in the table still caused its only failed check, although the prompt did not prohibit this accurate clarification.
• gdp-015

  C1Incorrect answers or scoring logic

  The rubric requires the insurer to have a duty to defend, overlooking the attached exclusion for assault- and battery-related suits, including negligence theories. A conditional analysis that follows the policy conflicts with this fixed conclusion.
• gdp-018

  C1Incorrect answers or scoring logic

  The rubric combines water immersion and accelerated aging into one condition. The source gives hardness changes of ±10 for immersion and +15/−0 for aging. This mismatch caused credit to be withheld for a correctly answered immersion requirement.
• gdp-021

  C2Scoring requirements not stated in the task

  The prompt permits ranking suppliers by dates and/or problems. The rubric forbids ranking by dates alone and fixes the worst supplier, excluding a different date-based ranking supported by the source table.
• gdp-022

  C1Incorrect answers or scoring logic

  The rubric reverses the load-bearing relationship, requiring recognition that an extractor hood can support a wall. The source instead makes the hood an exception to the restriction on wall cabinets on a weak wall; it does not support the reversed claim.
• gdp-023

  C1Incorrect answers or scoring logic

  The task asks which project category has the largest increase. The reference accepts Power & Regional, but the chart shows a larger increase for Neighborhood & Community between the end of 2023 and the end of 2025.
• gdp-024

  C2Scoring requirements not stated in the task

  The rubric treats eight months at the current agency as only eight months of qualifying federal service in total. The prompt does not rule out previous service, and the handbook allows the required 12 months to accumulate across agencies.
• gdp-025

  C1Incorrect answers or scoring logic

  The rubric requires reporting that the drawings lack a roof pitch. Page 2 of the attachment already labels both roof slopes 6:12, so a correct drawing review conflicts with the expected missing-item claim.
• gdp-026

  C2Scoring requirements not stated in the task

  The report does not explain how pending registrations enter the valuation model, yet the rubric requires predicting that the valuation will stay unchanged. A reasoned answer that the direction cannot be determined fails this fixed conclusion.
• gdp-028

  C2Scoring requirements not stated in the task

  The normal answer quoted a real heading containing Mr. Henley in an accurate page reference. It did not identify him as an executive vice president, but the mere appearance of his name still caused a failed check.
• gdp-031

  C2Scoring requirements not stated in the task

  The rubric silently restricts HFCS to two finished products with that name, excluding an interpretation based on the high-fructose corn syrup content of recipes. The two readings change production-line eligibility for 28 products, a core scheduling constraint.
• gdp-036

  C2Scoring requirements not stated in the task

  A complete, correct four-item list can violate the rubric merely by adding an accurate drawing-version reference. The prompt does not ban such source information. Neither recorded run triggered this rejection path.
• gdp-040

  C1Incorrect answers or scoring logic

  The reference part number for a ceramic bearing's outer ring is 316L616. The manufacturer's main drawing and both outer-ring symbols show 361L616. Copying the correct number conflicts with three rubric requirements.
• gdp-041

  C2Scoring requirements not stated in the task

  The normal answer distinguished tenure before and after childbirth and supplied the correct medical costs. It still lost credit for including the 0–4-year tenure bracket, which the public instructions did not prohibit.
• gdp-042

  C1Incorrect answers or scoring logic

  The rubric turns a slower rate of BMI increase in the 18–20 age group into no increase or change in BMI. The source table shows that BMI was still rising, just more slowly, confusing a change in level with a change in growth rate.
• gdp-043

  C2Scoring requirements not stated in the task

  The rubric requires identifying and moving or avoiding a recessed ceiling speaker that the public materials do not establish. An answer faithfully identifying the equipment actually shown can still miss this required fact.
• gdp-045

  C2Scoring requirements not stated in the task

  The word type can reasonably refer to the PDF's explicit TYPE column, yielding 13 first-floor double-hung windows. The rubric instead silently interprets it as a MARK number and accepts only zero.
• gdp-048

  C2Scoring requirements not stated in the task

  An open preliminary electrical-planning request is restricted by the rubric to 12–16 circuits and at least 12–14 board ways. It excludes a reasoned alternative with eight current circuits and two spare ways.
• gdp-052

  C2Scoring requirements not stated in the task

  The rubric requires advice about a statutory demand letter at least 30 days before filing suit. That procedural rule is absent from the attachment, while the prompt prohibits outside legal material.
• gdp-053

  C2Scoring requirements not stated in the task

  The rubric unconditionally bans inverse-square-law calculations. This excludes a reasonable adjustment from the study's 50-foot far-field data to 15 metres, even when the estimate meets the numerical requirement. The recorded runs did not test this rejection path.
• gdp-056

  C1Incorrect answers or scoring logic

  The rubric treats Kokkonen as the only shared citation, rejecting an accurately identified lower-court decision, 103 F.4th 294, that both documents also cite. The normal answer lost credit for this addition and failed the all-pass result.
• gdp-057

  C1Incorrect answers or scoring logic

  The rubric restricts the diseases associated with four intermediate-frequency models to PTSD, omitting four other diseases in Table 5. It consequently requires two drugs that are not shared by all the associated diseases.
• gdp-060

  C1Incorrect answers or scoring logic

  The financial report's displayed cost-to-revenue ratio falls from 17% to 16%, a decrease of 100 basis points. The reference requires 1,000 basis points, overstating this line item's change by a factor of ten.
• gdp-061

  C2Scoring requirements not stated in the task

  The normal answer supplied a complete field list but added ordinary grouping headings for readability. An undisclosed ban on these headings caused a failed check, making formatting an extra condition for accepting correct content.
• gdp-062

  C1Incorrect answers or scoring logic

  The rubric treats Chloe's absence from the named-insured list as sufficient to rule out all basic first-party medical coverage. It overlooks the policy's conformity-to-law clause and relevant resident-relative coverage rules.
• gdp-064

  C2Scoring requirements not stated in the task

  The answer selected the correct two age groups and accurately explained why three others did not qualify. The rubric expanded 'do not select other groups' into 'do not mention them,' causing the only failed check and a mean score of 0.875.
• gdp-067

  C1Incorrect answers or scoring logic

  The reference dates the air-conditioner A-coil replacement to 2020. Page 4 of the PDF explicitly records September 2022, so correctly extracting the date conflicts with the grading target.
• gdp-070

  C1Incorrect answers or scoring logic

  The reference accepts a 2019 mean sale price of $236,227.50, instead of the correct $229,241.67. The erroneous value is reproduced by including a 2018 sale in the 2019 group.
• gdp-071

  C1Incorrect answers or scoring logic

  The rubric requires describing the downslope stockpile silt fence as unclear or only partially provided. Drawing C-3 already shows and labels it SF, so correctly identifying the existing fence conflicts with the reference assessment.
• gdp-073

  C1Incorrect answers or scoring logic

  The rubric merges five four-base CTGC repeats into the three-base TGC group, requiring counts of CTGC=0 and TGC=12. Reading the underlined source sequences correctly gives five and seven respectively.
• gdp-076

  C2Scoring requirements not stated in the task

  The rubric prohibits a percentage increase in male relative mortality, excluding a normal comparison supported by the chart and explicitly labelled approximate. The prompt does not impose this ban, and the recorded runs did not trigger it.
• gdp-079

  C1Incorrect answers or scoring logic

  The task asks about the ECG in acute ST-elevation myocardial infarction, but the rubric requires depression in several leads. It excludes the corresponding elevation pattern supported by primary cases and cohort evidence. The normal run did not demonstrate rejection on these four checks.
• gdp-080

  C1Incorrect answers or scoring logic

  The rubric applies a leave accrual rate of 1.75 days per month, available only after 15 years of service, to the entire preceding two years. It requires 42 days, whereas applying the rates to the stated timeline normally gives about 36.
• gdp-082

  C1Incorrect answers or scoring logic

  The rubric requires saying that the paper gives no explicit methodological advantages or disadvantages. The first-page abstract states advantages of deep learning and both strengths and limitations of the two-layer analytical fit.
• gdp-087

  C2Scoring requirements not stated in the task

  The prompt asks for weak recommendations from all tables. The rubric silently limits the scope to Tables 2A–2D and bans three items in other tables that also carry the Category II label. Complete extraction therefore conflicts with grading.
• gdp-089

  C2Scoring requirements not stated in the task

  Claim 5 omits the first actual-cash-value payment date and the date repair costs were incurred, so the policy's 180-day period cannot be calculated. The rubric nevertheless fixes a denial of the hold-back payment, substituting two different dates.
• gdp-091

  C1Incorrect answers or scoring logic

  For a second-quarter profitability comparison, the reference uses the industry's first-half share of unprofitable institutions, 6.30%. The same-quarter figure is 6.59%, so the gap from community banks' 6.75% is overstated.
• gdp-093

  C2Scoring requirements not stated in the task

  Table 16 lists improvement opportunities, not a ranking of current standards' strengths. The rubric nevertheless requires naming Mfg Pyramid as currently strongest, treating opportunity coverage as existing capability.
• gdp-094

  C3Missed critical errors

  The task asks for the consequences of nutrient deficiencies during pregnancy, but the rubric checks table structure and nutrient names without checking the consequences. Replacing the whole column with blindness from an adjacent childhood example still satisfies the relevant written conditions. This answer was not run through the judge.
• gdp-097

  C2Scoring requirements not stated in the task

  The requested template calls for hourly production forecasts. The normal answer provided independently reproducible forecasts of 1,595.8 and 1,665.5 boxes per hour, but grading also required weekly totals and rejected these two items.
• gdp-098

  C1Incorrect answers or scoring logic

  The rubric treats records involving public funds as categorically ineligible for the attorney-communication exemption. It omits the source's non-privileged-content condition and the protection retained for confidential legal billing material.
• gdp-099

  C2Scoring requirements not stated in the task

  The rubric limits machining-parameter answers to general formulas, excluding qualitative speed and feed advice permitted by the PDF. The normal answer gave no specific rate values but still lost credit for such advice.
• gdp-100

  C2Scoring requirements not stated in the task

  The normal answer used and explained the paper's own R/I/E mechanism names. A rubric condition banned these terms and rejected the answer, although the public task did not prohibit source-supported terminology.

Doubtful: 13 tasks

• gdp-019

  C2Scoring requirements not stated in the task

  A computed value of 19.275 rounds to 19.28 at two decimal places, while the reference says 19.27. Likewise, 94.53% may exceed a literal upper bound of 94.5%. The recorded answers did not correctly submit these values, so the acceptance range remains unconfirmed.
• gdp-029

  C2Scoring requirements not stated in the task

  The rubric may prohibit supporting monetary data from another chart in the paragraph about headcount, even when the support is clearly labelled and the headcount analysis is correct. Whether the public task requires this strict separation remains unresolved.
• gdp-030

  C2Scoring requirements not stated in the task

  The instrument's error bound is ±50 ppm, while some CO₂ readings are barely above 50 ppm. The rubric fixes a conclusion about six-hour pH control, but different underlying conditions fit that error bound, so the evidence may not determine a unique conclusion.
• gdp-033

  C2Scoring requirements not stated in the task

  The prompt omits prior periodontal-treatment history, but the rubric fixes a 70% patient share for one service. A qualified answer received credit, weakening the false-rejection concern; whether the percentage depends on missing facts remains unresolved.
• gdp-047

  C2Scoring requirements not stated in the task

  The prompt asks for the numerical difference without specifying an absolute difference or subtraction order. Both 17/22 and −17/−22 have plausible interpretations, but the rubric fixes the negative values and expressly bans the positive ones.
• gdp-058

  C2Scoring requirements not stated in the task

  Compensation changes around an executive transition can be calculated by office or by individual. The normal answer explicitly used year-end officeholders and added same-person comparisons, but lost credit for not using the rubric's entry/exit convention. Its necessity remains unestablished.
• gdp-063

  C2Scoring requirements not stated in the task

  The answer correctly reported the target years and labelled two additional rows as baseline context. This caused its only failed check. The prompt does not expressly ban accurate background, but whether the extra rows exceed its requested scope remains disputed.
• gdp-066

  C1Incorrect answers or scoring logic

  The rubric describes cruise operating days' 5% contribution to revenue growth as a 5% increase in operating days themselves. Other correct operating metrics could legitimately satisfy the check, so this confusion's independent effect on credit has not been isolated.
• gdp-074

  C2Scoring requirements not stated in the task

  The prompt specifies an expanded steam-reformer boundary, while the rubric fixes whether streams 401, 428 and 430 are inputs. The recorded answer contains other mismatches and omissions, so its lost credit cannot yet be attributed to the boundary interpretation.
• gdp-083

  C1Incorrect answers or scoring logic

  It remains unclear how using an RV as a permanent residence affects uninsured-motorist bodily-injury coverage under this policy. The normal answer's conditional analysis failed three checks, while another accepted the qualification.
• gdp-085

  C1Incorrect answers or scoring logic

  Stable overall cell expansion and increased EdU proliferation-assay results can coexist, but the rubric does not clearly distinguish them. Two local failures strengthen the concern, though a complete correct answer may still satisfy the condition.
• gdp-088

  C1Incorrect answers or scoring logic

  The rubric gives competitiveness as an example of a bad explanation, potentially penalizing a source-supported goal of competitiveness in a specific field. The full context permits a compatible reading, and the recorded answers did not test this wording.
• gdp-092

  C2Scoring requirements not stated in the task

  The public question asks for three promising mitigation techniques, while the rubric fixes a five-item candidate list. The recorded runs did not test whether a reasonable third choice outside that list would be rejected.

Accepted: 40 tasks

• gdp-001
• gdp-003
• gdp-005
• gdp-007
• gdp-008
• gdp-010
• gdp-011
• gdp-012
• gdp-014
• gdp-016
• gdp-017
• gdp-020
• gdp-027
• gdp-032
• gdp-034
• gdp-035
• gdp-037
• gdp-038
• gdp-039
• gdp-044
• gdp-046
• gdp-049
• gdp-050
• gdp-051
• gdp-054
• gdp-055
• gdp-059
• gdp-065
• gdp-068
• gdp-069
• gdp-072
• gdp-075
• gdp-077
• gdp-078
• gdp-081
• gdp-084
• gdp-086
• gdp-090
• gdp-095
• gdp-096

Effects on evaluation and follow-up

Document-QA criteria need to follow the source and accept reasonable answers permitted by the prompt. Wrong numbers, dates or entities can reverse the relationship between correct extraction and credit. Hidden word bans and exclusions can turn accurate explanations into reasons for rejection.

The domain-specific judgments still require further verification by experts. Each issue’s effect on actual scoring depends on its trigger conditions. We will update the relevant conclusions when verifiable counterevidence, author clarifications or revisions become available.