FLOWING BENCHMARKS
Two benchmarks: writing quality and error detection
Across 7 fields, 130 distinct papers, and 280 fixed writing positions, Autocomplete with Evidence held an overall cross-field advantage of 31 wins; Evidence Card detected 91.4% of six explicit error types.
EVIDENCE CARD · CONTROLLED TEST · INDEPENDENT REVIEW
Across 35 completed Red challenges, 91.4% were confirmed detected.
Red error recall
32/35 confirmed; 33/35 when the review case is included.
Red card precision
34/35 confirmed · 1 to review · 0 rejected
Green control accuracy
11/11 Green controls handled correctly; 0/11 false Red alerts.
YELLOW PRECISION · NATURAL CASES
6/8 confirmed by independent review
The tasks and denominators differ, so the two rates are reported separately.
REPORTED MEASURES
Recall, card precision, control accuracy, and completion
Each measure has its own denominator. Technical failures are reported separately and are not counted as “no issue.”
Red error recall
91.4%32/35 confirmed; 33/35 when the review case is included.
Red card precision
97.1%34/35 confirmed · 1 to review · 0 rejected
Green control accuracy
100.0%11/11 Green controls handled correctly; 0/11 false Red alerts.
Test completion
95.8%46/48 positions completed; 2 technical failures reported separately, not counted as “no issue.”
RESULTS BY ERROR TYPE
Results for all six error types
Numbers, direction, causality, object, scope, and theory or definition are reported separately.
Dark = confirmed; light = review
Numbers
5/6No technical failures
Direction or polarity
6/6No technical failures
Causality
4/51 technical failure reported separately
Wrong object
5/6No technical failures
Scope
6/6No technical failures
Theory or definition
6/6No technical failures
YELLOW PRECISION · NATURAL CASES
6/8 confirmed by independent review
Yellow cards before and after tuning
Strict precision increased from 31.3% to 75.0%.
Before tuning, 16 cards were emitted: 5 confirmed, 1 to review, and 10 rejected. The full rerun emitted 8: 6 confirmed, 0 to review, and 2 rejected. 50/51 positions completed; 1 technical failure was reported separately.
Full rerun and independent card review complete48 PUBLIC CASES
Inputs, cards, and review results for every case
Each case includes the claim under test, frozen source evidence, emitted card, and independent review. Misses and technical failures remain in the record.
REDNumbersConfirmed detection溶液法原位大面积制备钙钛矿光电薄膜成膜的同步辐射可视化结晶过程研究
该研究报道,单个器件面积约15 cm²的大面积全柔性三维钙钛矿太阳能电池转换效率提升至15.23%。
Open case ↓
Claim under test
该研究报道,单个器件面积约15 cm²的大面积全柔性三维钙钛矿太阳能电池转换效率提升至15.23%。
Frozen source evidence
转换效率提升至5.23%(单个器件面积约15 cm2)
Emitted card · RED
目标将该研究中约15 cm²全柔性三维钙钛矿太阳能电池的转换效率写为15.23%,而原文对同一器件报告的是5.23%,构成直接数值矛盾。
“基于各层均采用卷对卷 全溶液狭缝涂布方法制备的大面积全柔性三维钙钛矿薄膜太阳能电池转换效率提升至5.23% (单个器件面积 约15 cm2)”— 溶液法原位大面积制备钙钛矿光电薄膜成膜的同步辐射可视化结晶过程研究
Independent review · valid
TARGET QUOTE声称'单个器件面积约15 cm²的大面积全柔性三维钙钛矿太阳能电池转换效率提升至15.23%',SOURCE QUOTE对同一器件(单个器件面积约15 cm²、大面积全柔性三维钙钛矿薄膜太阳能电池)报告转换效率为'5.23%'。两个数值15.23%与5.23%在相同实体(约15 cm²大面积全柔性三维钙钛矿太阳能电池)、相同关系(转换效率)上存在直接数值矛盾,构成直接反驳。
REDNumbersConfirmed detectionGenome-Scale CRISPR-Cas9 Knockout Screening in Human Cells
The GeCKO library targeted 18,080 human genes with 46,751 unique guide sequences.
Open case ↓
Claim under test
The GeCKO library targeted 18,080 human genes with 46,751 unique guide sequences.
Frozen source evidence
targeting 18,080 genes with 64,751 unique guide sequences
Emitted card · RED
The stated 46,751 unique guide sequences contradicts the source's 64,751 for the same GeCKO library. The source agrees with the stated 18,080 targeted genes.
“We show that lentiviral delivery of a genome-scale CRISPR-Cas9 knockout (GeCKO) library targeting 18,080 genes with 64,751 unique guide sequences enables both negative and positive selection screening in human cells.”— Genome-Scale CRISPR-Cas9 Knockout Screening in Human Cells
Independent review · valid
The TARGET QUOTE states '46,751 unique guide sequences' for the GeCKO library, while the SOURCE QUOTE states '64,751 unique guide sequences' for the same GeCKO library targeting the same 18,080 genes. This is a direct numerical contradiction on the exact same entity (GeCKO library), same relation (number of unique guide sequences), and same scope. The source explicitly confirms the 18,080 genes figure but contradicts the 46,751 guide sequences figure. The proposed 'red' judgment is correct.
REDNumbersConfirmed detectionEmpagliflozin, Cardiovascular Outcomes, and Mortality in Type 2 Diabetes
EMPA-REG OUTCOME treated a total of 8,020 patients and followed them for a median of 3.1 years.
Open case ↓
Claim under test
EMPA-REG OUTCOME treated a total of 8,020 patients and followed them for a median of 3.1 years.
Frozen source evidence
A total of 7020 patients were treated (median observation time, 3.1 years)
Emitted card · RED
The EMPA-REG OUTCOME report states that 7,020 patients were treated, not 8,020.
“A total of 7020 patients were treated (median observation time, 3.1 years).”— Empagliflozin, Cardiovascular Outcomes, and Mortality in Type 2 Diabetes
Independent review · valid
The target quote states '8,020 patients' while the source quote explicitly states '7020 patients were treated.' This is a direct numerical contradiction on the identical claim of patient count in the EMPA-REG OUTCOME trial. The entity (EMPA-REG OUTCOME trial), relation (patient count), and scope (total treated patients) match exactly, with only the number differing. The proposed color red is correct because the source directly contradicts the target proposition on a material point.
REDNumbersConfirmed detectionTabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second
The paper's small-task setting allows up to 10,000 training examples, 100 numerical features, and 10 classes.
Open case ↓
Claim under test
The paper's small-task setting allows up to 10,000 training examples, 100 numerical features, and 10 classes.
Frozen source evidence
≤ 1 000 training examples, ≤ 100 purely numerical features and ≤ 10 classes
Emitted card · RED
The paper defines its small-task setting as at most 1,000 training examples, not 10,000.
“has thus learned to solve novel small tabular classification tasks (≤ 1 000 training examples,≤ 100 purely numerical features without missing values and≤ 10 classes)”— TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second
Independent review · valid
The TARGET QUOTE states 'up to 10,000 training examples,' while the SOURCE QUOTE explicitly states '≤ 1 000 training examples.' This is a direct numerical contradiction on the same entity (training examples), same relation (upper bound on small-task setting), same scope (the paper's definition of small-task setting). The source directly contradicts the target proposition that the limit is 10,000 by specifying it is 1,000.
REDNumbersConfirmed detectionThe Progress, Challenges, and Perspectives of Directed Greybox Fuzzing
The survey investigates 24 state-of-the-art fuzzers closely related to directed greybox fuzzing.
Open case ↓
Claim under test
The survey investigates 24 state-of-the-art fuzzers closely related to directed greybox fuzzing.
Frozen source evidence
investigation of 42 state-of-the-art fuzzers that are closely related to DGF
Emitted card · RED
The survey explicitly reports investigating 42 state-of-the-art fuzzers closely related to DGF, not 24.
“Based on the investigation of 42 state-of-the-art fuzzers that are closely related to DGF, we conduct the first in-depth study to summarize the empirical evidence on the research progress of DGF.”— The Progress, Challenges, and Perspectives of Directed Greybox Fuzzing
Independent review · valid
The target quote claims '24 state-of-the-art fuzzers closely related to directed greybox fuzzing.' The source quote explicitly states '42 state-of-the-art fuzzers that are closely related to DGF' (where DGF = directed greybox fuzzing). The entity (fuzzers closely related to directed greybox fuzzing), relation (investigation of), and scope (state-of-the-art) match exactly; the only difference is the number (24 vs. 42), which is a direct numerical contradiction on a material point. The proposed color 'red' is correct.
REDNumbersMissedEvaluating the Impact of Brazil's Bolsa Família: Cash Transfer Programs in Comparative Perspective
The review describes Bolsa Família as benefiting roughly 17 million families.
Open case ↓
Claim under test
The review describes Bolsa Família as benefiting roughly 17 million families.
Frozen source evidence
benefiting roughly 11 million families
No card was emitted
REDDirection or polarityConfirmed detectionEmpagliflozin, Cardiovascular Outcomes, and Mortality in Type 2 Diabetes
In EMPA-REG OUTCOME, cardiovascular death was significantly higher with empagliflozin than with placebo.
Open case ↓
Claim under test
In EMPA-REG OUTCOME, cardiovascular death was significantly higher with empagliflozin than with placebo.
Frozen source evidence
significantly lower rates of death from cardiovascular causes (3.7%, vs. 5.9% in the placebo group)
Emitted card · RED
The claim that cardiovascular death was higher with empagliflozin contradicts the account of the same trial’s cardiovascular results, which explicitly reports a lower rate of death from cardiovascular causes.
“Cardiovascular Outcomes The primary outcome that was reported previ ously showed a significantly lower risk of 3-point MACE in the pooled empagliflozin group than in the placebo group, a difference that was driven by a lower rate of death from cardiovas cular causes.23”— Empagliflozin and Progression of Kidney Disease in Type 2 Diabetes
Independent review · valid
The target quote explicitly claims cardiovascular death was 'significantly higher with empagliflozin than with placebo' in EMPA-REG OUTCOME. The source quote directly contradicts this by stating there was 'a lower rate of death from cardiovascular causes' in the empagliflozin group, which 'drove' the composite benefit. Both quotes concern the same trial (EMPA-REG OUTCOME), the same outcome (cardiovascular death/mortality), the same comparison (empagliflozin vs. placebo), and the same directionality claim—but with opposite directions. The contradiction is material and unambiguous: 'higher' vs. 'lower' for the identical endpoint in the identical trial.
REDDirection or polarityConfirmed detectionEvaluating the Impact of Brazil's Bolsa Família: Cash Transfer Programs in Comparative Perspective
The review concludes that Bolsa Família increased inequality and extreme poverty.
Open case ↓
Claim under test
The review concludes that Bolsa Família increased inequality and extreme poverty.
Frozen source evidence
helped reduce inequality and extreme poverty
Emitted card · RED
The review attributes 21 percent of Brazil’s 1995–2004 Gini decline to Bolsa Família, contradicting the claimed increase in inequality.
“They found that the Gini index for Brazil fell by 4.7 percent from 1995 to 2004 and that Bolsa Familia was responsible for 21 percent of that fall (before 1995 there were no CCT programs in Brazil, so there is confidence that 1995 is a baseline year).”— Evaluating the Impact of Brazil's Bolsa Família: Cash Transfer Programs in Comparative Perspective
Independent review · valid
The TARGET QUOTE states 'The review concludes that Bolsa Família increased inequality.' The SOURCE QUOTE from the same review states that Bolsa Família was responsible for '21 percent of that fall' in the Gini index, where 'fall' means a decrease in inequality. A decrease in the Gini index signifies reduced inequality. Therefore, the source directly contradicts the target proposition: the review concludes Bolsa Família decreased inequality (contributed to 21% of the Gini decline), not increased it. The scope matches—both concern the review's conclusion about Bolsa Família's effect on inequality.
Scope and limitations
- 1This controlled test measures whether the system can find explicit errors. It does not estimate how often those errors occur in real papers.
- 2Some source papers also appeared in early development runs. These results support regression testing and controlled-recall claims, not prevalence claims from a random sample.
- 3The tuned run emitted only 8 Yellow cards. The 6/8 denominator must stay visible; 75% is not an estimate of overall precision.