Datasets:

rodriguescarson
/

polychart-shown-is-not-supported

+# The Deception Atlas
+21 real published charts verified against primary sources, spanning 11 deception mechanisms,
+including the AI industry's own. Every value cited, none reconstructed. Each carries a
+measure-then-claim reading and, where a single figure is defined, a computed Tufte Lie Factor.
+## truncated y axis  (4 charts)
+**"If the Bush Tax Cuts Expire" — top marginal income-tax rate** (Fox News, America's Newsroom, 2012) - Lie Factor 4.95
+  The top rate rises from 35% to 39.6%, an increase of 4.6 points (13%). [source](https://www.mediamatters.org/fox-business/dishonest-fox-chart-bush-tax-cut-edition)
+**ACA enrolment as of March 27 versus the CBO projection** (Fox News, America's Newsroom, 2014) - Lie Factor 2.55
+  6.0m enrolled is 85% of the 7.066m projection, not the one third the chart implies. [source](https://www.mediamatters.org/fox-news/dishonest-fox-charts-obamacare-enrollment-edition)
+**Number of Americans on Welfare vs. Number with Full-Time Jobs** (Fox News (Fox and Friends), 2013) - Lie Factor 4.69
+  108.6 million versus 101.7 million is a ratio of about 1.07 to 1, nearly equal; on a zero-baseline axis the bars would be almost the same height. [source](https://www.mediamatters.org/fox-friends/dishonest-fox-chart-overstates-comparison-welfare-full-time-work-500-percent)
+**Americans on Government Aid, 2009-2011 (truncated-axis growth chart)** (Fox News (America's Newsroom), 2012) - Lie Factor 6.22
+  The increase was under 12 percent over 2009 to 2011, a modest step on a zero baseline; from the 94m baseline the later bar is about 7 times the height of the earlier one, grossly exaggerating a recession-driven rise. [source](https://www.mediamatters.org/fox-news/today-dishonest-fox-charts-government-aid-edition)
+## inverted y axis  (1 charts)
+**Gun deaths in Florida, before and after Stand Your Ground** (Reuters / Business Insider (chart by C. Chan), 2014) - Lie Factor -1.0
+  Firearm deaths rose after the 2005 law, from 521 in 2005 to 825 in 2007, and stayed well above the pre-law level. [source](https://www.livescience.com/45083-misleading-gun-death-chart.html)
+## expanded y axis  (2 charts)
+**"The only #climatechange chart you need to see" — global temperature** (National Review (chart from Power Line), 2015) - Lie Factor 0.017
+  The global temperature anomaly rises about 1.1 c from 1880 to 2015, a real trend the axis flattens to near zero. [source](https://www.washingtonpost.com/news/the-fix/wp/2015/12/14/why-the-national-reviews-global-temperature-graph-is-so-misleading/)
+**Unemployment Rate Under Obama (Nov 2011 plotted as not-falling)** (Fox News, 2011) - named mechanism, severity banded
+  On a correct linear axis 8.6 is below 8.8 is below 9.0, so november at 8.6 percent is the lowest of the three points and the line clearly declines from october to november; unemployment was falling and the scale conceale. [source](https://www.mediamatters.org/laura-ingraham/today-dishonest-fox-news-graphics)
+## dual axis mismatch  (2 charts)
+**Abortions versus cancer screenings at Planned Parenthood** (Americans United for Life, shown at a US House hearing, 2015) - Lie Factor 2.86
+  In 2013 there were 935,573 cancer screening and prevention services and 327,000 abortions, so screenings outnumbered abortions about 2.9 to 1. [source](https://www.politifact.com/factchecks/2015/oct/01/jason-chaffetz/chart-shown-planned-parenthood-hearing-misleading-/)
+**Mask-mandate vs non-mandate COVID cases per 100k, Kansas (KDHE)** (Kansas Department of Health and Environment, presented by Dr. Lee Norman, 2020) - named mechanism, severity banded
+  On a single shared axis mask-mandate counties had higher absolute case rates than non-mandate counties throughout, and including the omitted jul 3 to 11 days shows mandate-county cases had been rising (9.1 to 25.3 per 10. [source](https://www.kwch.com/2020/08/11/factfinder-12-digs-into-controversy-concerning-kdhe-chart-on-masks/)
+## cherry picked window  (3 charts)
+**Rising Gas Prices Under Obama (three-point chart)** (Fox News (America's Newsroom), 2012) - named mechanism, severity banded
+  Prices peaked at $3.906 in may 2011, fell to $3.266 by december 2011, and only partly rebounded to $3.579 in feb 2012, still below the 2011 peak; the full eia series is a bumpy plateau, not a steady climb. [source](https://mediamatters.org/blog/2012/02/21/fox-still-struggling-with-basic-chart-concepts/185049)
+**'Satellite data show zero warming for 17-18 years' (series started at the 1998 El Nino)** (Sen. Ted Cruz, US Senate hearing and media appearances, 2015) - named mechanism, severity banded
+  The authoritative surface record rose from +0.46 c in 1997 to +0.90 c in 2015, about +0.44 c in 18 years, and 2014, 2015 and 2016 each set successive records; the satellite dataset was later revised upward, erasing the a. [source](https://www.politifact.com/factchecks/2015/mar/20/ted-cruz/ted-cruzs-worlds-fire-not-last-17-years/)
+**'Global warming stopped 16 years ago' (HadCRUT4, 1997-2012)** (The Mail on Sunday (David Rose), 2012) - named mechanism, severity banded
+  The met office's own analysis of the window gave a trend of about +0.03 c/decade, small but positive and not statistically distinguishable from long-term warming; warming resumed unmistakably after, with 2015 and 2016 se. [source](https://www.carbonbrief.org/the-mail-on-sunday-gives-david-rose-space-to-repeat-old-and-wrong-claims-that-global-warming-has-stopped)
+## reordered x  (1 charts)
+**Top 5 Counties with the Greatest Number of Confirmed COVID-19 Cases** (Georgia Department of Public Health, 2020) - named mechanism, severity banded
+  In correct chronological order the descending staircase disappears; cases were roughly flat to gently declining, and the apparent late drop was largely a reporting-lag artifact; dph apologized and attributed it to incorr. [source](https://www.route-fifty.com/management/2020/05/officials-apologize-retool-coronavirus-graph-after-backlash/165478/)
+## unequal bins  (1 charts)
+**New Cases Per Day (COVID-19) with an irregular unlabeled y-axis** (Fox 31 Denver (KDVR-TV), Colorado, 2020) - named mechanism, severity banded
+  On a uniform zero-based axis the same series climbs steeply, several-fold across the two-week window; because the gap from 90 to 100 (10 cases) is drawn the same size as 350 to 400 (50 cases), the design flattens and und. [source](https://saracdoan.com/index.php/misleading-covid-19-visualizations/)
+## pie 3d  (2 charts)
+**Steve Jobs' 3-D pie chart of US smartphone market share (Macworld 2008)** (Apple Macworld 2008 keynote slide (Steve Jobs, Jan 15 2008), 2008) - named mechanism, severity banded
+  Rim's 39.0% is almost exactly double apple's 19.5%, and the other catch-all at 21.2% was larger than apple's slice; on a flat correctly-scaled pie apple's wedge is clearly less than half of rim's. [source](https://gist.github.com/bljustice/d7af0a24c628a4ec992a350bdc137a46)
+**Major Categories of Federal Income and Outlays for Fiscal Year 2010** (US Internal Revenue Service (Form 1040 instruction booklet), 2011) - named mechanism, severity banded
+  On an undistorted 2d pie borrowing was the single largest income slice at 37 percent, about 1.4 times personal income taxes at 26 percent, meaning roughly 37 cents of every federal dollar in fy2010 was borrowed. [source](https://www.irs.gov/pub/irs-prior/i1040gi--2011.pdf)
+## percent absolute swap  (2 charts)
+**Lipitor 'reduces risk of heart attack by 36%' direct-to-consumer ad** (Pfizer DTC advertisement for Lipitor (atorvastatin), 2008) - named mechanism, severity banded
+  Over about 3.3 years heart attacks fell from 3.0% to 1.9%, an absolute reduction of 1.1 percentage points; about 91 patients must be treated for 3.3 years to prevent one heart attack, and the advertised 36% is just 1.1 d. [source](https://6xc.medium.com/the-marketing-of-lipitor-49a93089d8cb)
+**'Measles was already vanishing before the vaccine' (death rate shown as disease burden)** (Recurring anti-vaccination graphic genre, rebutted by CDC and Our World in Data, 2015) - named mechanism, severity banded
+  Case incidence, the measure of transmission, stayed high right up to the vaccine (257.9 per 100k in 1962) and only fell after mass vaccination (31.2 in 1967, 22.8 in 1970); better medicine cut deaths, only the vaccine cu. [source](https://ourworldindata.org/measles-vaccines-save-lives)
+## misscaled bars  (1 charts)
+**GPT-5 launch 'coding deception' bar chart (the 'mega chart screwup')** (OpenAI, GPT-5 launch livestream, 2025) - named mechanism, severity banded
+  As printed on the livestream, 50.0 is higher than 47.4, so the labeled gpt-5 deception rate was worse than o3's, the opposite of what the bar heights showed; the corrected blog figure for gpt-5 with thinking is 16.5. sam. [source](https://tech.co/news/openai-apologizes-mega-chart-screwup)
+## claim not supported  (2 charts)
+**Llama-4 Maverick on the LMArena leaderboard (experimental variant, not the release)** (Meta / LMArena leaderboard, 2025) - named mechanism, severity banded
+  The model developers could download was not the model on the leaderboard: the public maverick scored about 1370 versus the experimental variant's 1417, and researchers comparing outputs found the two models responded dif. [source](https://www.theregister.com/2025/04/08/meta_llama4_cheating/)
+**AutoScientist augmentation panel: '+16 win-rate points beyond 20,000 datapoints'** (Adaption AutoScientist platform, 2026) - named mechanism, severity banded
+  At matched base model and epochs, crossing 20,000 rows lowered the held-out category win rate from 76 to 72, so the observed effect is minus 4, not the advertised plus 16. [source](https://adaptionlabs.ai)

README.md CHANGED Viewed

@@ -1,78 +1,267 @@
 ---
-license: cc-by-4.0
 task_categories:
-- visual-question-answering
 tags:
 - chart-question-answering
 - data-visualization
-- claim-verification
-- autoscientist-challenge
 ---
-# PolyChart: Shown Is Not Supported
-Chart interpretation under a **measure-then-claim** contract. Every row pairs a
-rendered chart with the source table it was drawn from, and an answer that
-separates what the chart's encoding supports from what the data actually says.
-Built for the AutoScientist Challenge 2026 (Part 2, Data Visualization).
-Powered by **AutoScientist by Adaption**.
-## Why this exists
-A chart is a completion claim. It renders cleanly, states a confident finding,
-and the underlying data may not support it. Models reading charts inherit that
-gap: they answer from the visual impression because they never extracted the
-values. This dataset trains the opposite habit.
-## Composition
-| Slice | Rows | Share |
 |---|---|---|
-| supported | 1256 | 72.4% |
-| contradicted | 322 | 18.6% |
-| unanswerable | 157 | 9.0% |
-Split: **1558 train / 120 holdout**. The holdout contains only
-answerable rows and is split by chart, never by row, so it measures behaviour
-rather than memorisation.
-## How the labels are made
-Labels are **constructed, not judged**. Real data from Our World in Data is
-rendered twice: once faithfully, once with a distortion whose parameters are
-recorded. Those parameters are the ground truth, so every claim is derived
-arithmetically from the same table the chart was drawn from.
-| Distortion | Rows | Ground truth |
-|---|---|---|
-| faithful | 785 | zero-based axis, full window |
-| truncated_axis_honest | 628 | truncated axis, claim still supported |
-| truncated_y_axis | 50 | exaggeration factor, gated at >= 2.0x |
-| inverted_y_axis | 109 | fitted slope sign |
-| aspect_ratio_exaggeration | 105 | figure ratio, values unaltered |
-| cherry_picked_window | 58 | window and full-series trend both reverse |
-`truncated_axis_honest` is a control, not filler. Without charts that are
-truncated *and* correctly answerable, a model learns that a non-zero baseline
-means the chart is lying, and then objects to honest charts.
-## Verification
-Every row is re-derived from its own `source_table` by `polychart/verify.py`,
-which shares no code with the generator. A verifier built from the same code as
-the thing it checks is only re-running the generator.
-**Current status: 1735 / 1735 rows verified.**
-Five silent failures were caught this way during construction and are documented
-in the repository, including one in the verifier itself.
-## Provenance
-Data: Our World in Data (27 indicators, 30 countries), CC-BY.
-Per-row `source_citation` and `source_url` carry the original attribution.
-Charts are rendered locally; no chart image is scraped.
-Generation code is public and deterministic: the same catalog reproduces
-byte-identical charts.

 ---
+license: apache-2.0
 task_categories:
+- question-answering
+- text-generation
+language:
+- en
 tags:
 - chart-question-answering
 - data-visualization
+- misleading-charts
+- tufte-lie-factor
+- deception-detection
+- counterfactual-augmentation
+- benchmark
+pretty_name: 'PolyChart: Shown Is Not Supported'
+size_categories:
+- 1K<n<10K
 ---
+# Shown Is Not Supported
+A chart is a completion claim. It renders cleanly, states a confident finding,
+and the underlying data may not support it. A model reading that chart inherits
+the gap: it answers from the visual impression because it never extracted the
+values.
+This adapter reads the data first and corrects the chart when its encoding
+misleads. It ships with the first continuous score for *how much* a chart lies.
+Trained with AutoScientist by Adaption for the AutoScientist Challenge 2026,
+Part 2, Data Visualization.
+## Results
+| Evaluation | Base | Adapted |
 |---|---|---|
+| On the dataset | 15 | **85** |
+| Across the Data Visualization category (held-out) | 22 | **78** |
+Pairwise win rate, LLM judge, against the same untuned base (Llama-3.3-70B).
+78% on the held-out category clears the gate; the Legal category was won at 73.
+The number is the least interesting part. Here is what the work actually is.
+## The failure this probes: a shortcut, not a knowledge gap
+Geirhos et al. (2020) call a shortcut a decision rule that works on the benchmark
+and fails when the test gets harder. Their network learned an object's *location*
+instead of its *shape*, and broke the moment location stopped predicting the label.
+Chart models do the same thing. We ran a base VLM on 60 matched pairs, the same
+series drawn faithfully and drawn to deceive, and the result is not uniform:
+| Distortion | Wrong direction, faithful to distorted |
+|---|---|
+| Inverted axis | 17% to **93%** |
+| Truncated axis | 0% to 0% (magnitude error fell) |
+| Aspect ratio | 4% to 4% (no effect) |
+The model takes **direction from the geometry and magnitude from the axis labels,
+and never reconciles them**. Invert the axis and the geometry shortcut breaks:
+the trend reads backwards 93% of the time while the magnitude, read from the
+untouched labels, survives with identical error (52.1 both sides). Squash the
+aspect ratio, which changes steepness but not label values, and nothing happens,
+because steepness was never being read.
+That is a clean, isolated shortcut. The dataset is the intervention that closes it.
+## The intervention: counterfactual augmentation, built exact
+Kaushik, Hovy and Lipton (2020) fix shortcut learning by pairing each example with
+a counterfactual: the same content, minimally edited to flip the label, so a model
+cannot lean on the spurious pattern. Their edits were written by humans, which is
+slow, approximate, and does not scale.
+We build the same thing from parameters. Every chart is rendered twice from one
+OWID series: once honestly, once with a distortion whose parameters we set. The
+distortion *is* the counterfactual edit, and the label is derived from it, not
+annotated. Same data, altered encoding, exact label, no human in the loop.
+Composition, 2,129 rows, weighted to the failure cells (see next section):
+- **supported (54%)** accurate extraction and interpretation
+- **severity (27%)** the lie factor, see below
+- **contradicted (13%)** the encoding overstates or reverses the data; the answer
+  gives the true figure and names the mechanism
+- **unanswerable (5%)** the chart cannot answer the question; the model says so
+## Targeted, not uniform: hard-example mining
+Winning data-centric submissions (KAIST, Data-Centric AI Competition) do not add
+uniform coverage; they find where the model fails and concentrate data there. We
+ran the base model across every question-type by distortion cell and scored it
+against exact ground truth. The failures cluster in three places: severity
+estimation, precise value extraction under a distorted axis, and mechanism
+identification on inverted and cherry-picked charts. Comparison and abstention
+were already solved. We re-weighted the dataset to those failure cells and removed
+the solved ones.
+The effect is measurable. The uniform version of this dataset trained on a 109B
+base reached 76 on the held-out category. This failure-weighted version reached 78
+on a **smaller 70B base**. Targeted data compensated for less model capacity, and
+on-dataset win rate rose from 77 to 85. Curation and targeting beat both scale and
+capacity, the result the data-centric literature predicts.
+## The new part: deception has a magnitude, and it is computable
+Every misleading-chart dataset, including Misviz (ACL 2026), answers a yes/no
+question: is this chart misleading, by which of N mechanisms. None answers *how
+much*.
+Edward Tufte defined the Lie Factor in 1983: the size of the effect shown in the
+graphic over the size of the effect in the data. One is honest. Because we
+construct every distortion, we compute it exactly rather than estimate it. 471
+rows carry a lie factor and a severity band, from honest (1.0) through a
+truncation that triples the apparent slope (3.0) to an inverted axis that keeps
+the magnitude and reverses the sign (-1.0).
+This turns deception detection from a label into a graded, calibrated target. No
+released chart dataset provides it.
+## The Deception Atlas: the measure on charts that actually aired
+The distortions above are constructed, which is what makes their Lie Factors exact.
+The atlas shows the task transfers to the real world: **21 published charts, verified
+against primary sources, across 11 distinct mechanisms**, none reconstructed. This is
+the part of the dataset that is not a chart-QA benchmark. It is a quantified corpus of
+real chart lies, a thing no released dataset provides.
+| Chart | Mechanism | Severity | The real reading |
+|---|---|---|---|
+| Fox News, top tax rate (2012) | truncated bars | LF 4.95 | 35% to 39.6%, a 13% rise, drawn 5.6x taller |
+| Fox News, welfare vs work (2013) | truncated bars | LF 4.69 | 108.6M vs 101.7M, a 7% gap, drawn 5x taller |
+| Planned Parenthood, AUL chart (2015) | dual mismatched axes | LF 2.86 | screenings outnumber abortions 2.9 to 1, drawn as equal |
+| Reuters, Florida gun deaths (2014) | inverted axis | LF -1.0 | deaths rose 521 to 825 after 2005, drawn as a fall |
+| National Review, global temperature (2015) | expanded axis | LF 0.017 | a real 1.1 C rise flattened to under 2% of the frame |
+| Ted Cruz, satellite "pause" (2015) | cherry-picked window | severe | window anchored to the 1998 El Nino peak; +0.44 C over 1997-2015 |
+| Steve Jobs, smartphone share (2008) | 3D pie perspective | severe | Apple 19.5% tilted forward to rival RIM's 39% |
+| Georgia DPH, COVID counties (2020) | reordered x-axis | severe | dates out of order manufacture a decline that is not in the data |
+| Lipitor, "36% risk reduction" (2008) | percent-absolute swap | severe | a 1.1 point absolute drop (3.0 to 1.9) sold as 36% |
+| OpenAI, GPT-5 "coding deception" chart (2025) | mis-scaled bars | severe | a bar labeled 50.0 drawn shorter than one labeled 47.4; corrected to 16.5 |
+| Meta, Llama-4 on LMArena (2025) | claim vs data | severe | rank earned by an experimental variant; the released model scored ~1370, not 1417 |
+| Adaption, "+16 beyond 20k rows" | claim vs data | severe | this project's runs show 76 to 72, a minus 4 |
+The severities span the full signed range the measure defines: -1.0 where an inverted
+axis reverses a real rise, 0.017 where an expanded axis hides one, 4.95 where a truncated
+baseline invents one. A yes/no "is it misleading" label collapses all of these into the
+same bit. Where a single Lie Factor is not defined (a reordered axis, a 3D pie) the row
+names the mechanism and bands the severity rather than invent a number. Where a baseline
+was unlabelled it is marked inferred with the arithmetic shown.
+The atlas also holds the AI industry's own charts. The GPT-5 launch bar chart drew a bar
+labeled 50.0 shorter than one labeled 47.4, geometry against its own printed values, in
+Sam Altman's words a "mega chart screwup." The Llama-4 leaderboard rank was earned by a
+non-public variant the fine print called Experimental. And the sharpest case is the
+platform that trained this model: its augmentation panel advertises +16 win-rate points
+beyond 20,000 datapoints, and this project's controlled runs, at matched base and epochs,
+measured 76 falling to 72. The atlas audits the tools of its own field, on the dataset's
+own thesis: shown is not supported. Full corpus, sources, and computation are in
+`data/atlas/` and `data/real/`.
+## A controlled study, not a single run
+We ran the task through five base models AutoScientist selected and read the
+training curves out of every adapter.
+| Base | Params | On-data | Category | What the curve shows |
+|---|---|---|---|---|
+| Llama-3.3-70B (this model) | 70B | 15 to 85 | 22 to **78** | 3 epochs, hard-mined data, eval loss down |
+| Llama-4-Scout (uniform data) | 109B | 24 to 77 | 24 to 76 | same recipe, uniform coverage |
+| Mistral | 7B | 34 to 66 | 47 to 54 | held slightly |
+| gemma | 4B | 52 to 48 | n/a | eval loss 1.28 to 2.73; severe overfit |
+| Llama-3.2 | 3B | 49 to 51 | 54 to **47** | overfit; category fell below base |
+| Qwen3.5 | 0.8B | 41 to 59 | 56 to **45** | underfit and category below base |
+Two findings.
+**Capacity dominates for uniform data, but targeted data closes the gap.** With
+uniform coverage, only the large bases held their held-out performance: every base
+under 10B overfit at the platform default of 10 epochs, eval loss rising from the
+first checkpoint while train loss fell toward memorization. Yet the failure-weighted
+data let a 70B base reach 78, above the 109B's 76 on uniform data. Capacity sets the
+floor; targeting the data raises it. The Hardware Lottery (Hooker 2020) is real, and
+better data is the lever a competitor actually controls.
+**The headline number hides who gets hurt.** Every small model improved on its own
+data while degrading on the held-out category. This is the effect Hooker et al.
+(2019) documented for compression: aggregate accuracy barely moves while the
+underrepresented distribution is quietly forgotten. Our two-panel split, on-data
+against held-out category, is the instrument that exposes it, the same move as
+their Pruning Identified Exemplars. We report both panels so the concealment is
+visible.
+**More data made it worse.** We tested the platform's own claim that crossing
+20,000 datapoints yields +16 win-rate points. Expanding this dataset from 1,882 to
+21,882 rows with the platform's synthetic augmentation dropped the held-out
+category from 76 to 72, at matched epochs and base model, with eval loss still
+descending. It is dilution by unverified synthetic data, not overfitting: 1,882
+verified rows beat 21,882 augmented ones by four points. Curation beats scale.
+## The metric probes itself
+The eligibility number is LLM-judge pairwise win rate, so it inherits the biases
+Zheng et al. (2023) documented for that judge, verbosity among them: a preference
+for longer answers regardless of quality. We probe it directly with a terse-answer
+ablation, and we write answers to be correct-and-complete rather than terse, so
+substance and length are separated rather than confounded.
+## Nine silent failures, caught by an independent verifier
+The dataset is about models stating confident findings their input does not
+support. We built a verifier that shares no code with the generator and re-derives
+every claim from the source table. It caught nine such failures in our own
+pipeline before anything shipped, including:
+- an answer calling a 183% rise "far more modest than it looks" on a 1.28x
+  exaggeration
+- an ingest that returned unusable data for 22 of 28 sources while every request
+  returned 200 and the build reported success
+- three failures inside the verifier itself, which failed the standard it applies
+The last point is the thesis reproduced inside the tool built to study it: a
+grader that is confident, well-formed, and wrong until an independent check catches
+it. Every failure and its fix is in the repository.
+## Honest limitations
+- **Epoch count did not matter on this base.** At 3 epochs this model's eval loss
+  fell throughout (0.944 to 0.861, no overfit); an earlier 10-epoch run on related
+  data scored within noise (78 vs 76 category). The large bases are insensitive to
+  the epoch schedule, unlike the sub-10B bases in the study, which overfit badly at 10.
+- **Editor overrides did not persist.** AutoScientist re-selected the base model
+  regardless of the manual recipe, so the five-model comparison is confounded by
+  base as well as composition. We report the base per run rather than hide it.
+- **Text specifications for the headline model.** Charts are given as plotted
+  values plus axis parameters. A multimodal image variant is released separately.
+- **Single evaluation source**, platform-reported, with a documented plus or minus
+  8 to 10 point run-to-run noise we did not average out.
+## Related work
+Misviz and Misviz-synth (ACL 2026) release misleading visualizations and classify
+which of 12 design rules a chart breaks. Chart-QA benchmarks (ChartQA, CharXiv,
+ChartMuseum) test reading. Our contribution is orthogonal: a written correction
+grounded in the source table, and a continuous severity target, neither of which a
+released dataset provides.
+## Artifacts
+- Seed dataset (verified): `polychart-shown-is-not-supported`
+- Adapted dataset (trained this): `chart-qa-with-axis-tricks`
+- Real-world validation slice: `data/real/`, three published charts with cited data
+- Weights (this repo)
+- Generation and verification code: public, reproducible from one command
+- Source data: Our World in Data, CC-BY, per-row citation
+## References
+- Geirhos et al. Shortcut Learning in Deep Neural Networks. Nature Machine Intelligence, 2020.
+- Kaushik, Hovy, Lipton. Learning the Difference that Makes a Difference with Counterfactually-Augmented Data. ICLR, 2020.
+- Hooker et al. What Do Compressed Deep Neural Networks Forget? 2019.
+- Hooker. The Hardware Lottery. 2020.
+- Zheng et al. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. NeurIPS, 2023.
+- Mazumder et al. DataPerf: Benchmarks for Data-Centric AI Development. NeurIPS, 2023.
+- Tufte. The Visual Display of Quantitative Information. 1983.

atlas_charts_verified.json ADDED Viewed

	@@ -0,0 +1,486 @@

+[
+  {
+    "id": "lipitor_36pct_relative_risk_2008",
+    "title": "Lipitor 'reduces risk of heart attack by 36%' direct-to-consumer ad",
+    "publication": "Pfizer DTC advertisement for Lipitor (atorvastatin)",
+    "year": 2008,
+    "source_url": "https://6xc.medium.com/the-marketing-of-lipitor-49a93089d8cb",
+    "mechanism": "percent_absolute_swap",
+    "chart_type": "risk-reduction graphic",
+    "real_values": [
+      {
+        "label": "placebo heart-attack rate over ~3.3 yr (%)",
+        "value": 3.0
+      },
+      {
+        "label": "Lipitor heart-attack rate over ~3.3 yr (%)",
+        "value": 1.9
+      },
+      {
+        "label": "absolute risk reduction (percentage points)",
+        "value": 1.1
+      },
+      {
+        "label": "relative risk reduction as advertised (%)",
+        "value": 36
+      }
+    ],
+    "unit": "percent event rate over ~3.3 years",
+    "what_the_design_did": "headlined the 36% relative risk reduction while the absolute change, 3.0% to 1.9% or 1.1 percentage points, appeared only in fine print",
+    "true_reading": "over about 3.3 years heart attacks fell from 3.0% to 1.9%, an absolute reduction of 1.1 percentage points; about 91 patients must be treated for 3.3 years to prevent one heart attack, and the advertised 36% is just 1.1 divided by 3.0",
+    "data_source_citation": "Sever PS et al., ASCOT-LLA, Lancet 2003;361:1149-1158: atorvastatin 100 events (1.9%) vs placebo 154 events (3.0%), HR 0.64, N=10,305.",
+    "inferred_params": "NNT about 91 computed as 1/0.011; ad publication year approximate"
+  },
+  {
+    "id": "apple_3d_pie_smartphone_2008",
+    "title": "Steve Jobs' 3-D pie chart of US smartphone market share (Macworld 2008)",
+    "publication": "Apple Macworld 2008 keynote slide (Steve Jobs, Jan 15 2008)",
+    "year": 2008,
+    "source_url": "https://gist.github.com/bljustice/d7af0a24c628a4ec992a350bdc137a46",
+    "mechanism": "pie_3d",
+    "chart_type": "3D pie chart",
+    "real_values": [
+      {
+        "label": "RIM (BlackBerry)",
+        "value": 39.0
+      },
+      {
+        "label": "Other",
+        "value": 21.2
+      },
+      {
+        "label": "Apple (iPhone)",
+        "value": 19.5
+      },
+      {
+        "label": "Palm",
+        "value": 9.8
+      },
+      {
+        "label": "Motorola",
+        "value": 7.4
+      },
+      {
+        "label": "Nokia",
+        "value": 3.1
+      }
+    ],
+    "unit": "percent of US smartphone market share (2007)",
+    "what_the_design_did": "tilted the pie in 3-D so Apple's 19.5% slice sat at the front; perspective foreshortening enlarged the near Apple wedge and shrank the far RIM wedge, making 19.5% look nearly as large as RIM's 39%",
+    "true_reading": "RIM's 39.0% is almost exactly double Apple's 19.5%, and the Other catch-all at 21.2% was larger than Apple's slice; on a flat correctly-scaled pie Apple's wedge is clearly less than half of RIM's",
+    "data_source_citation": "Canalys US smartphone market-share data (2007) as shown on Apple's Macworld 2008 slide; values sum to 100.0%.",
+    "inferred_params": "data period cited as 2007, exact quarter not independently confirmed"
+  },
+  {
+    "id": "georgia_dph_covid_reordered_dates_2020",
+    "title": "Top 5 Counties with the Greatest Number of Confirmed COVID-19 Cases",
+    "publication": "Georgia Department of Public Health",
+    "year": 2020,
+    "source_url": "https://www.route-fifty.com/management/2020/05/officials-apologize-retool-coronavirus-graph-after-backlash/165478/",
+    "mechanism": "reordered_x",
+    "chart_type": "grouped vertical bar chart (5 counties per date, 14 dates)",
+    "real_values": [
+      {
+        "label": "x-axis dates as displayed",
+        "value": "Apr 28, Apr 27, Apr 29, May 1, Apr 30, May 4, May 6, May 5, May 2, May 7, Apr 26, May 3, May 8, May 9"
+      },
+      {
+        "label": "x-axis in true chronological order",
+        "value": "Apr 26, Apr 27, Apr 28, Apr 29, Apr 30, May 1, May 2, May 3, May 4, May 5, May 6, May 7, May 8, May 9"
+      },
+      {
+        "label": "counties re-sorted largest-to-smallest each day",
+        "value": "Fulton, Cobb, Gwinnett, DeKalb, Hall"
+      }
+    ],
+    "unit": "confirmed COVID-19 case counts, x-axis = calendar dates",
+    "what_the_design_did": "placed the 14 report-dates out of chronological order and re-sorted the county bars largest-to-smallest within each day, so the bars formed a smooth descending staircase and manufactured the impression of a steady daily decline just as the state pressed to reopen",
+    "true_reading": "in correct chronological order the descending staircase disappears; cases were roughly flat to gently declining, and the apparent late drop was largely a reporting-lag artifact; DPH apologized and attributed it to incorrect sorting logic",
+    "data_source_citation": "Georgia DPH bar chart posted ~May 10 2020; documented by Route Fifty, Georgia Public Broadcasting, and the Atlanta Journal-Constitution; displayed date order reproduced at covid-georgia.com.",
+    "inferred_params": "per-cell daily counts never published as a table; the documented manipulation is the ordering itself, not invented values"
+  },
+  {
+    "id": "fox_welfare_vs_fulltime_work_2013",
+    "title": "Number of Americans on Welfare vs. Number with Full-Time Jobs",
+    "publication": "Fox News (Fox and Friends)",
+    "year": 2013,
+    "source_url": "https://www.mediamatters.org/fox-friends/dishonest-fox-chart-overstates-comparison-welfare-full-time-work-500-percent",
+    "mechanism": "truncated_y_axis",
+    "chart_type": "two-bar vertical bar chart",
+    "axis_min": 99980000,
+    "real_values": [
+      {
+        "label": "people in households receiving federal means-tested benefits, Q4 2011",
+        "value": 108600000
+      },
+      {
+        "label": "people with full-time jobs, 2012",
+        "value": 101700000
+      }
+    ],
+    "unit": "number of people",
+    "what_the_design_did": "drew the two bars on a truncated non-zero axis so the welfare bar towered roughly five times taller than the full-time-work bar, even though the two figures differ by less than 7 percent",
+    "true_reading": "108.6 million versus 101.7 million is a ratio of about 1.07 to 1, nearly equal; on a zero-baseline axis the bars would be almost the same height",
+    "data_source_citation": "Fox and Friends, Oct 28 2013. Benefits: US Census Bureau SIPP Q4 2011, Table 2. Full-time workers: US Census Bureau CPS 2012. Reported by Media Matters.",
+    "inferred_params": "y-axis baseline not published; inferred ~99.98M by solving (108.6-b)/(101.7-b)=5 for the documented five-times-taller visual; the two data values are documented, not inferred"
+  },
+  {
+    "id": "fox_gas_prices_cherrypick_feb2012",
+    "title": "Rising Gas Prices Under Obama (three-point chart)",
+    "publication": "Fox News (America's Newsroom)",
+    "year": 2012,
+    "source_url": "https://mediamatters.org/blog/2012/02/21/fox-still-struggling-with-basic-chart-concepts/185049",
+    "mechanism": "cherry_picked_window",
+    "chart_type": "line chart plotting only ~3 points across ~1 year",
+    "real_values": [
+      {
+        "year": "Jan 2009 trough baseline",
+        "value": 1.788
+      },
+      {
+        "year": "Feb 2011",
+        "value": 3.211
+      },
+      {
+        "year": "May 2011 actual peak, omitted",
+        "value": 3.906
+      },
+      {
+        "year": "Dec 2011",
+        "value": 3.266
+      },
+      {
+        "year": "Jan 2012",
+        "value": 3.38
+      },
+      {
+        "year": "Feb 2012 framed as the high",
+        "value": 3.579
+      }
+    ],
+    "unit": "US dollars per gallon (national average, regular, EIA)",
+    "what_the_design_did": "plotted only about three points across a year to imply a steady climb to a new high, omitting the true May 2011 peak of $3.906, which was higher than the Feb 2012 value presented as the peak",
+    "true_reading": "prices peaked at $3.906 in May 2011, fell to $3.266 by December 2011, and only partly rebounded to $3.579 in Feb 2012, still below the 2011 peak; the full EIA series is a bumpy plateau, not a steady climb",
+    "data_source_citation": "Fox News, Feb 2012, per Media Matters. Real prices: US EIA monthly regular gasoline series EMM_EPMR_PTE_NUS_DPG.",
+    "inferred_params": "Fox's exact plotted values were not transcribed; the listed real_values are the true EIA data showing the omitted May 2011 peak exceeded the Feb 2012 value"
+  },
+  {
+    "id": "irs_3d_pie_federal_budget_fy2010",
+    "title": "Major Categories of Federal Income and Outlays for Fiscal Year 2010",
+    "publication": "US Internal Revenue Service (Form 1040 instruction booklet)",
+    "year": 2011,
+    "source_url": "https://www.irs.gov/pub/irs-prior/i1040gi--2011.pdf",
+    "mechanism": "pie_3d",
+    "chart_type": "3D-perspective pie chart",
+    "real_values": [
+      {
+        "label": "Income: borrowing to cover deficit",
+        "value": 37
+      },
+      {
+        "label": "Income: personal income taxes",
+        "value": 26
+      },
+      {
+        "label": "Income: social security/medicare/retirement taxes",
+        "value": 25
+      },
+      {
+        "label": "Income: corporate income taxes",
+        "value": 6
+      },
+      {
+        "label": "Income: excise/customs/estate/gift/misc",
+        "value": 6
+      }
+    ],
+    "unit": "percent of total federal income (FY2010: income $2,163B, deficit $1,293B)",
+    "what_the_design_did": "rendered the pie in 3D perspective, foreshortening the back slices and inflating the apparent area of the front slices so slice size no longer maps to true share",
+    "true_reading": "on an undistorted 2D pie borrowing was the single largest income slice at 37 percent, about 1.4 times personal income taxes at 26 percent, meaning roughly 37 cents of every federal dollar in FY2010 was borrowed",
+    "data_source_citation": "IRS 2011 Instructions for Form 1040, page 97; also critiqued by Naomi Robbins, Forbes, Apr 16 2012.",
+    "inferred_params": "none; all percentages printed by the IRS and read directly from page 97"
+  },
+  {
+    "id": "fox_unemployment_broken_scale_2011",
+    "title": "Unemployment Rate Under Obama (Nov 2011 plotted as not-falling)",
+    "publication": "Fox News",
+    "year": 2011,
+    "source_url": "https://www.mediamatters.org/laura-ingraham/today-dishonest-fox-news-graphics",
+    "mechanism": "expanded_y_axis",
+    "chart_type": "line chart of the monthly unemployment rate",
+    "real_values": [
+      {
+        "label": "Mar 2011 U-3 rate",
+        "value": 8.8
+      },
+      {
+        "label": "Oct 2011 U-3 rate",
+        "value": 9.0
+      },
+      {
+        "label": "Nov 2011 U-3 rate",
+        "value": 8.6
+      }
+    ],
+    "unit": "percent (U-3 unemployment rate, BLS)",
+    "what_the_design_did": "plotted the points on a broken non-linear vertical scale so November's 8.6 percent was drawn at the same height as October's 9.0 percent and above March's 8.8 percent, erasing the downward slope so a falling rate looked flat or rising",
+    "true_reading": "on a correct linear axis 8.6 is below 8.8 is below 9.0, so November at 8.6 percent is the lowest of the three points and the line clearly declines from October to November; unemployment was falling and the scale concealed it",
+    "data_source_citation": "Fox News, ~Dec 9-12 2011, per Media Matters (Dec 12 2011): the chart plotted Nov 8.6% level with Oct 9.0% and above Mar 8.8%. BLS U-3 rates. Also Business Insider and the Washington Post.",
+    "inferred_params": "mechanism mapped to expanded_y_axis as the closest fit; strictly a broken non-linear scale; the three percentages are real BLS figures, none inferred"
+  },
+  {
+    "id": "fox_govt_aid_recipients_truncated_2012",
+    "title": "Americans on Government Aid, 2009-2011 (truncated-axis growth chart)",
+    "publication": "Fox News (America's Newsroom)",
+    "year": 2012,
+    "source_url": "https://www.mediamatters.org/fox-news/today-dishonest-fox-charts-government-aid-edition",
+    "mechanism": "truncated_y_axis",
+    "chart_type": "column growth chart of people on government aid",
+    "axis_min": 94000000,
+    "real_values": [
+      {
+        "label": "people in a household receiving means-tested aid, ~2009",
+        "value": 96000000
+      },
+      {
+        "label": "people in a household receiving means-tested aid, ~2011",
+        "value": 108000000
+      }
+    ],
+    "unit": "number of people (chart showed no numeric labels; endpoints approximate)",
+    "what_the_design_did": "started the y-axis at 94,000,000 instead of 0, so a rise from roughly 96M to 108M recipients, under 12 percent, filled most of the chart height and looked like a near-vertical surge",
+    "true_reading": "the increase was under 12 percent over 2009 to 2011, a modest step on a zero baseline; from the 94M baseline the later bar is about 7 times the height of the earlier one, grossly exaggerating a recession-driven rise",
+    "data_source_citation": "Fox News, Aug 9 2012, per Media Matters; figures from Sen. Sessions' office based on US Census Bureau means-tested-aid data, 2009-2011.",
+    "inferred_params": "y-axis baseline 94,000,000 is documented by Media Matters, as is the under-12% rise; the ~96M and ~108M endpoints are approximate because the chart showed no numeric labels; the ~7x height ratio is arithmetic from the 94M baseline"
+  },
+  {
+    "id": "cruz_satellite_no_warming_2015",
+    "title": "'Satellite data show zero warming for 17-18 years' (series started at the 1998 El Nino)",
+    "publication": "Sen. Ted Cruz, US Senate hearing and media appearances",
+    "year": 2015,
+    "source_url": "https://www.politifact.com/factchecks/2015/mar/20/ted-cruz/ted-cruzs-worlds-fire-not-last-17-years/",
+    "mechanism": "cherry_picked_window",
+    "chart_type": "line",
+    "real_values": [
+      {
+        "year": 1997,
+        "value": 0.46
+      },
+      {
+        "year": 1998,
+        "value": 0.61
+      },
+      {
+        "year": 2005,
+        "value": 0.68
+      },
+      {
+        "year": 2010,
+        "value": 0.72
+      },
+      {
+        "year": 2014,
+        "value": 0.75
+      },
+      {
+        "year": 2015,
+        "value": 0.9
+      },
+      {
+        "year": 2016,
+        "value": 1.01
+      }
+    ],
+    "unit": "degrees C global surface temperature anomaly vs 1951-1980 (NASA GISTEMP v4)",
+    "what_the_design_did": "began the satellite series at the 1997-1998 El Nino, the strongest of the 20th century; anchoring a trend line to that anomalous peak flattens the apparent slope",
+    "true_reading": "the authoritative surface record rose from +0.46 C in 1997 to +0.90 C in 2015, about +0.44 C in 18 years, and 2014, 2015 and 2016 each set successive records; the satellite dataset was later revised upward, erasing the apparent pause",
+    "data_source_citation": "NASA GISS GISTEMP v4 annual means; PolitiFact 2015-03-20; Carbon Brief on RSS/UAH trends; Skeptical Science on the RSS v3.3 to v4 revision.",
+    "inferred_params": "real_values are the exact NASA GISTEMP surface record over the window the zero-warming claim contradicts; RSS v3.3 slope not stated as no exact citable figure was obtained"
+  },
+  {
+    "id": "rose_mail_global_warming_stopped_2012",
+    "title": "'Global warming stopped 16 years ago' (HadCRUT4, 1997-2012)",
+    "publication": "The Mail on Sunday (David Rose)",
+    "year": 2012,
+    "source_url": "https://www.carbonbrief.org/the-mail-on-sunday-gives-david-rose-space-to-repeat-old-and-wrong-claims-that-global-warming-has-stopped",
+    "mechanism": "cherry_picked_window",
+    "chart_type": "line",
+    "real_values": [
+      {
+        "year": 1996,
+        "value": 0.183
+      },
+      {
+        "year": 1997,
+        "value": 0.39
+      },
+      {
+        "year": 1998,
+        "value": 0.539
+      },
+      {
+        "year": 2005,
+        "value": 0.545
+      },
+      {
+        "year": 2010,
+        "value": 0.56
+      },
+      {
+        "year": 2012,
+        "value": 0.47
+      },
+      {
+        "year": 2013,
+        "value": 0.514
+      }
+    ],
+    "unit": "degrees C global temperature anomaly vs 1961-1990 (HadCRUT4)",
+    "what_the_design_did": "plotted HadCRUT4 from about January 1997, just before the 1998 El Nino spike of 0.539 C, so the fitted line looks nearly flat and appears to prove warming stopped",
+    "true_reading": "the Met Office's own analysis of the window gave a trend of about +0.03 C/decade, small but positive and not statistically distinguishable from long-term warming; warming resumed unmistakably after, with 2015 and 2016 setting records",
+    "data_source_citation": "Met Office HadCRUT4 global annual series; Carbon Brief rebuttal; Grantham Institute LSE; the Met Office publicly rebutted the piece the same weekend.",
+    "inferred_params": "real_values are exact HadCRUT4 annual means (same dataset the monthly chart used); +0.03 C/decade is the Met Office's stated regression value"
+  },
+  {
+    "id": "antivax_measles_mortality_not_incidence",
+    "title": "'Measles was already vanishing before the vaccine' (death rate shown as disease burden)",
+    "publication": "Recurring anti-vaccination graphic genre, rebutted by CDC and Our World in Data",
+    "year": 2015,
+    "source_url": "https://ourworldindata.org/measles-vaccines-save-lives",
+    "mechanism": "percent_absolute_swap",
+    "chart_type": "line",
+    "real_values": [
+      {
+        "year": 1938,
+        "case_rate_per_100k": 647.4,
+        "death_rate_per_100k": 2.5
+      },
+      {
+        "year": 1958,
+        "case_rate_per_100k": 438.0,
+        "death_rate_per_100k": 0.3
+      },
+      {
+        "year": 1962,
+        "case_rate_per_100k": 257.9,
+        "death_rate_per_100k": 0.2
+      },
+      {
+        "year": 1967,
+        "case_rate_per_100k": 31.2,
+        "death_rate_per_100k": 0.04
+      },
+      {
+        "year": 1970,
+        "case_rate_per_100k": 22.8,
+        "death_rate_per_100k": 0.04
+      }
+    ],
+    "unit": "US measles cases and deaths per 100,000 population",
+    "what_the_design_did": "plotted the measles death rate, which collapsed ~98% before the 1963 vaccine due to antibiotics and better care, implying measles was already nearly eliminated and the vaccine was superfluous",
+    "true_reading": "case incidence, the measure of transmission, stayed high right up to the vaccine (257.9 per 100k in 1962) and only fell after mass vaccination (31.2 in 1967, 22.8 in 1970); better medicine cut deaths, only the vaccine cut cases; plotting incidence instead of mortality reverses the conclusion",
+    "data_source_citation": "Our World in Data 'measles-cases-and-death-rate' (US) from US historical vital statistics/CDC; CDC 'History of Measles' gives ~3-4 million cases/yr pre-vaccine.",
+    "inferred_params": "the deception is metric substitution (mortality for incidence), mapped to the nearest mechanism; all per-100,000 rates are exact OWID/CDC values; vaccine licensed 1963"
+  },
+  {
+    "id": "fox31_denver_covid_unequal_yaxis_2020",
+    "title": "New Cases Per Day (COVID-19) with an irregular unlabeled y-axis",
+    "publication": "Fox 31 Denver (KDVR-TV), Colorado",
+    "year": 2020,
+    "source_url": "https://saracdoan.com/index.php/misleading-covid-19-visualizations/",
+    "mechanism": "unequal_bins",
+    "chart_type": "line chart of new cases per day (15 days)",
+    "real_values": [
+      {
+        "label": "y-axis gridline labels (drawn at equal spacing, numerically irregular)",
+        "value": "30, 60, 90, 100, 130, 160, 190, 240, 250, 300, 350, 400"
+      },
+      {
+        "label": "y-axis interval sequence between labels",
+        "value": "30, 30, 10, 30, 30, 30, 50, 10, 50, 50, 50"
+      },
+      {
+        "label": "date range",
+        "value": "Mar 18 to Apr 1 2020"
+      },
+      {
+        "label": "total cases printed on the chart",
+        "value": 3342
+      }
+    ],
+    "unit": "COVID-19 new cases per day on a non-linear unequal-interval y-axis",
+    "what_the_design_did": "labeled the y-axis at numerically irregular intervals (a 10-step from 90 to 100, then 50-steps, with a 10-step wedged between 240 and 250) while drawing every gridline at equal vertical spacing, so equal on-screen distance stands for very different case counts and the steep growth curve reads as gentle and near-linear",
+    "true_reading": "on a uniform zero-based axis the same series climbs steeply, several-fold across the two-week window; because the gap from 90 to 100 (10 cases) is drawn the same size as 350 to 400 (50 cases), the design flattens and understates how fast cases were rising",
+    "data_source_citation": "aired on Fox 31 Denver (KDVR); y-axis label set and equal spacing reconstructed and confirmed by freerangestats.info (2020-04-06), drewtyre.rbind.io, and Sara C. Doan, PhD (saracdoan.com); date range and 3,342 total printed on the chart.",
+    "inferred_params": "The verified artifact is the unequal y-axis label set and interval sequence, corroborated across three independent sources. The per-day plotted series (community reverse-engineering, c. 33 to 376) is NOT read from a table and sums to 3,159, which does not match the 3,342 total printed on the chart, so those daily values are treated as approximate and excluded from the ground-truth values above. Publisher Fox 31 Denver per saracdoan; freerangestats labels it 'Fox'."
+  },
+  {
+    "id": "kansas_kdhe_mask_mandate_dual_axis_2020",
+    "title": "Mask-mandate vs non-mandate COVID cases per 100k, Kansas (KDHE)",
+    "publication": "Kansas Department of Health and Environment, presented by Dr. Lee Norman",
+    "year": 2020,
+    "source_url": "https://www.kwch.com/2020/08/11/factfinder-12-digs-into-controversy-concerning-kdhe-chart-on-masks/",
+    "mechanism": "dual_axis_mismatch",
+    "chart_type": "dual-axis line chart",
+    "real_values": [
+      {
+        "label": "mask-mandate counties, change in cases per 100k (Jul 12 to Aug 6)",
+        "value": "-34.4%, plotted on the left y-axis"
+      },
+      {
+        "label": "non-mandate counties, change over the same window",
+        "value": "+0.7%, plotted on the right y-axis at a different scale"
+      },
+      {
+        "label": "hidden window Jul 3 to Jul 11 (excluded)",
+        "value": "mandate counties rose from 9.1 to 25.3 daily cases per 100k"
+      },
+      {
+        "label": "mandate effective date vs chart start",
+        "value": "mandate Jul 3; chart starts Jul 12, the local peak"
+      }
+    ],
+    "unit": "COVID-19 cases per 100,000 population",
+    "what_the_design_did": "plotted the two county groups on two different y-axis scales, so lines in the same unit were rescaled to look comparable and dramatize the mandate group's relative drop, and started the x-axis on Jul 12, the case peak, rather than Jul 3 when the mandate took effect, hiding the Jul 3 to 11 rise",
+    "true_reading": "on a single shared axis mask-mandate counties had higher absolute case rates than non-mandate counties throughout, and including the omitted Jul 3 to 11 days shows mandate-county cases had been rising (9.1 to 25.3 per 100k) before the cherry-picked start; KDHE Secretary Norman conceded 'when you have two different axes working at the same time, our eyes attract to that'",
+    "data_source_citation": "dual differently-scaled y-axes and Norman's on-camera admission reported by KWCH FactFinder 12 (2020-08-11) and WIBW (2020-08); cherry-picked Jul 12 start and hidden Jul 3-11 rise from a Kansas Policy Institute open-records request via The Sentinel.",
+    "inferred_params": "exact numeric tick ranges of the two axes not independently re-verified; the load-bearing fact, two separate y-axes on different scales for the same unit, is confirmed by mainstream local news and by Norman's own admission"
+  },
+  {
+    "id": "openai_gpt5_deception_chart_2025",
+    "title": "GPT-5 launch 'coding deception' bar chart (the 'mega chart screwup')",
+    "publication": "OpenAI, GPT-5 launch livestream",
+    "year": 2025,
+    "source_url": "https://tech.co/news/openai-apologizes-mega-chart-screwup",
+    "mechanism": "misscaled_bars",
+    "chart_type": "bar chart of deception-eval rates",
+    "real_values": [
+      {"label": "GPT-5 (with thinking), as labeled on the livestream bar", "value": 50.0},
+      {"label": "o3, as labeled", "value": 47.4},
+      {"label": "GPT-5 (with thinking), corrected value in the blog post", "value": 16.5}
+    ],
+    "unit": "percent deception rate (lower is better)",
+    "what_the_design_did": "drew the GPT-5 bar labeled 50.0 noticeably shorter than the o3 bar labeled 47.4, so the bar geometry contradicted the printed values and made the new model look better at avoiding deception than the labels said; a printed number on another slide was itself wrong",
+    "true_reading": "as printed on the livestream, 50.0 is higher than 47.4, so the labeled GPT-5 deception rate was worse than o3's, the opposite of what the bar heights showed; the corrected blog figure for GPT-5 with thinking is 16.5. Sam Altman: 'The numbers here were accurate but we screwed up the bar charts in the livestream overnight; on another slide we screwed up numbers'",
+    "data_source_citation": "OpenAI GPT-5 launch livestream (Aug 2025) and corrected launch blog post; Altman quote via tech.co; values via Engadget/Yahoo coverage of the livestream chart and the corrected blog. OpenAI's official account called it an 'unintentional chart crime'.",
+    "inferred_params": "none; all three values and both quotes are documented in the cited coverage, and OpenAI itself published the correction"
+  },
+  {
+    "id": "llama4_maverick_lmarena_2025",
+    "title": "Llama-4 Maverick on the LMArena leaderboard (experimental variant, not the release)",
+    "publication": "Meta / LMArena leaderboard",
+    "year": 2025,
+    "source_url": "https://www.theregister.com/2025/04/08/meta_llama4_cheating/",
+    "mechanism": "claim_not_supported",
+    "chart_type": "leaderboard ranking chart",
+    "real_values": [
+      {"label": "Llama-4-Maverick-03-26-Experimental, ELO at launch", "value": 1417},
+      {"label": "publicly released Maverick, ELO (approx, per The Register)", "value": 1370}
+    ],
+    "unit": "LMArena ELO (human pairwise preference)",
+    "what_the_design_did": "the leaderboard position was earned by a non-public chat-optimized variant, disclosed only in fine print as Llama-4-Maverick-03-26-Experimental, whose style (verbosity, enthusiastic tone, emojis) was tailored to human-preference voting; the launch messaging presented the rank as the released model's",
+    "true_reading": "the model developers could download was not the model on the leaderboard: the public Maverick scored about 1370 versus the experimental variant's 1417, and researchers comparing outputs found the two models responded differently; LMArena stated Meta should have made the customization clearer and updated its leaderboard policies",
+    "data_source_citation": "The Register, 2025-04-08 (Meta accused of Llama 4 bait-n-switch to juice LMArena rank), Meta-confirmed variant status; LMArena public statement and policy update; Hugging Face discussion on meta-llama/Llama-4-Maverick-17B-128E-Instruct.",
+    "inferred_params": "released-model ELO reported as approximately 1370 by The Register; the load-bearing facts, a non-public variant on the leaderboard and the released model scoring materially lower, are multiply documented"
+  }
+]

deception_atlas_rows.jsonl ADDED Viewed

	@@ -0,0 +1,46 @@

+{"chart_id": "fox_tax_2012", "kind": "real", "question_type": "value_lookup", "label": "supported", "mechanism": "truncated_y_axis", "prompt": "Chart: \"If the Bush Tax Cuts Expire\" \u2014 top marginal income-tax rate (Fox News, America's Newsroom, 2012)\nType: bar chart\nY axis: %, displayed range 34 to 42.\nThe axis does not start at zero.\nPlotted values:\n  Now: 35\n  If Bush tax cuts expire: 39.6\n\nQuestion: According to this chart, what is the value for \"Now\"?", "response": "35 %. Read from the plotted value, not the bar height, which the truncated axis distorts.", "source_citation": "US federal top marginal income-tax rate: 35.0% (2012) versus the 39.6% Clinton-era rate that would return on expiration.", "source_url": "https://www.mediamatters.org/fox-business/dishonest-fox-chart-bush-tax-cut-edition"}
+{"chart_id": "fox_tax_2012", "kind": "real", "question_type": "value_lookup", "label": "supported", "mechanism": "truncated_y_axis", "prompt": "Chart: \"If the Bush Tax Cuts Expire\" \u2014 top marginal income-tax rate (Fox News, America's Newsroom, 2012)\nType: bar chart\nY axis: %, displayed range 34 to 42.\nThe axis does not start at zero.\nPlotted values:\n  Now: 35\n  If Bush tax cuts expire: 39.6\n\nQuestion: According to this chart, what is the value for \"If Bush tax cuts expire\"?", "response": "39.6 %. Read from the plotted value, not the bar height, which the truncated axis distorts.", "source_citation": "US federal top marginal income-tax rate: 35.0% (2012) versus the 39.6% Clinton-era rate that would return on expiration.", "source_url": "https://www.mediamatters.org/fox-business/dishonest-fox-chart-bush-tax-cut-edition"}
+{"chart_id": "fox_tax_2012", "kind": "real", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "truncated_y_axis", "prompt": "Chart: \"If the Bush Tax Cuts Expire\" \u2014 top marginal income-tax rate (Fox News, America's Newsroom, 2012)\nType: bar chart\nY axis: %, displayed range 34 to 42.\nThe axis does not start at zero.\nPlotted values:\n  Now: 35\n  If Bush tax cuts expire: 39.6\n\nQuestion: The taller bar looks about 5.6x the shorter one. Is that the real difference?", "response": "No. The bars differ 5.6x in height because the axis starts at 34, not zero. The values differ only 1.131x: the top rate rises from 35% to 39.6%, an increase of 4.6 points (13%). The non-zero baseline exaggerates the difference.", "source_citation": "US federal top marginal income-tax rate: 35.0% (2012) versus the 39.6% Clinton-era rate that would return on expiration.", "source_url": "https://www.mediamatters.org/fox-business/dishonest-fox-chart-bush-tax-cut-edition"}
+{"chart_id": "fox_tax_2012", "kind": "real", "question_type": "lie_factor", "label": "severity", "mechanism": "truncated_y_axis", "prompt": "Chart: \"If the Bush Tax Cuts Expire\" \u2014 top marginal income-tax rate (Fox News, America's Newsroom, 2012)\nType: bar chart\nY axis: %, displayed range 34 to 42.\nThe axis does not start at zero.\nPlotted values:\n  Now: 35\n  If Bush tax cuts expire: 39.6\n\nQuestion: On Tufte's Lie Factor (1.0 = honest, higher = exaggerated), how severely does this chart's design distort the data?", "response": "Lie Factor about 4.95, severe. The bars are drawn 5.6x apart in height while the values are only 1.131x apart, because the axis is truncated at 34. The top rate rises from 35% to 39.6%, an increase of 4.6 points (13%).", "check": {"lie_factor": 4.95, "severity_band": "severe", "height_ratio": 5.6, "value_ratio": 1.131}, "source_citation": "US federal top marginal income-tax rate: 35.0% (2012) versus the 39.6% Clinton-era rate that would return on expiration.", "source_url": "https://www.mediamatters.org/fox-business/dishonest-fox-chart-bush-tax-cut-edition"}
+{"chart_id": "fox_aca_2014", "kind": "real", "question_type": "value_lookup", "label": "supported", "mechanism": "truncated_y_axis", "prompt": "Chart: ACA enrolment as of March 27 versus the CBO projection (Fox News, America's Newsroom, 2014)\nType: bar chart\nY axis: million, displayed range 5.467 to 7.2. (baseline inferred from the aired chart)\nThe axis does not start at zero.\nPlotted values:\n  March 27 enrolment: 6\n  CBO projection: 7.066\n\nQuestion: According to this chart, what is the value for \"March 27 enrolment\"?", "response": "6 million. Read from the plotted value, not the bar height, which the truncated axis distorts.", "source_citation": "HHS: ~6.0M enrolled by 27 Mar 2014; CBO original projection 7.066M by 31 Mar 2014.", "source_url": "https://www.mediamatters.org/fox-news/dishonest-fox-charts-obamacare-enrollment-edition"}
+{"chart_id": "fox_aca_2014", "kind": "real", "question_type": "value_lookup", "label": "supported", "mechanism": "truncated_y_axis", "prompt": "Chart: ACA enrolment as of March 27 versus the CBO projection (Fox News, America's Newsroom, 2014)\nType: bar chart\nY axis: million, displayed range 5.467 to 7.2. (baseline inferred from the aired chart)\nThe axis does not start at zero.\nPlotted values:\n  March 27 enrolment: 6\n  CBO projection: 7.066\n\nQuestion: According to this chart, what is the value for \"CBO projection\"?", "response": "7.066 million. Read from the plotted value, not the bar height, which the truncated axis distorts.", "source_citation": "HHS: ~6.0M enrolled by 27 Mar 2014; CBO original projection 7.066M by 31 Mar 2014.", "source_url": "https://www.mediamatters.org/fox-news/dishonest-fox-charts-obamacare-enrollment-edition"}
+{"chart_id": "fox_aca_2014", "kind": "real", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "truncated_y_axis", "prompt": "Chart: ACA enrolment as of March 27 versus the CBO projection (Fox News, America's Newsroom, 2014)\nType: bar chart\nY axis: million, displayed range 5.467 to 7.2. (baseline inferred from the aired chart)\nThe axis does not start at zero.\nPlotted values:\n  March 27 enrolment: 6\n  CBO projection: 7.066\n\nQuestion: The taller bar looks about 3.0x the shorter one. Is that the real difference?", "response": "No. The bars differ 3.0x in height because the axis starts at 5.467, not zero. The values differ only 1.178x: 6.0M enrolled is 85% of the 7.066M projection, not the one third the chart implies. The non-zero baseline exaggerates the difference.", "source_citation": "HHS: ~6.0M enrolled by 27 Mar 2014; CBO original projection 7.066M by 31 Mar 2014.", "source_url": "https://www.mediamatters.org/fox-news/dishonest-fox-charts-obamacare-enrollment-edition"}
+{"chart_id": "fox_aca_2014", "kind": "real", "question_type": "lie_factor", "label": "severity", "mechanism": "truncated_y_axis", "prompt": "Chart: ACA enrolment as of March 27 versus the CBO projection (Fox News, America's Newsroom, 2014)\nType: bar chart\nY axis: million, displayed range 5.467 to 7.2. (baseline inferred from the aired chart)\nThe axis does not start at zero.\nPlotted values:\n  March 27 enrolment: 6\n  CBO projection: 7.066\n\nQuestion: On Tufte's Lie Factor (1.0 = honest, higher = exaggerated), how severely does this chart's design distort the data?", "response": "Lie Factor about 2.55, moderate. The bars are drawn 3.0x apart in height while the values are only 1.178x apart, because the axis is truncated at 5.467. 6.0M enrolled is 85% of the 7.066M projection, not the one third the chart implies.", "check": {"lie_factor": 2.55, "severity_band": "moderate", "height_ratio": 3.0, "value_ratio": 1.178}, "source_citation": "HHS: ~6.0M enrolled by 27 Mar 2014; CBO original projection 7.066M by 31 Mar 2014.", "source_url": "https://www.mediamatters.org/fox-news/dishonest-fox-charts-obamacare-enrollment-edition"}
+{"chart_id": "national_review_climate_2015", "kind": "real", "question_type": "lie_factor", "label": "severity", "mechanism": "expanded_y_axis", "prompt": "Chart: \"The only #climatechange chart you need to see\" \u2014 global temperature (National Review (chart from Power Line), 2015)\nType: line chart\nX axis: Year, 1880 to 2015\nY axis: degrees C anomaly, displayed range -23.3 to 43.3 (a span of 67).\n\nPlotted values (every ~20 years):\n  1880: 0.0608\n  1899: 0.0197\n  1918: -0.0621\n  1937: 0.3479\n  1956: 0.099\n  1975: 0.258\n  1994: 0.5914\n  2013: 0.9826\n\nQuestion: On Tufte's Lie Factor (1.0 = honest, below 1 = the design under-states the effect), how does this chart's axis affect the trend?", "response": "Lie Factor about 0.017, severe understatement. The series rises 1.121 degrees C anomaly from 1880 to 2015, a real trend, but the axis span of 67 compresses it into under 2% of the frame, so it reads as flat. Expanding an axis hides a change as surely as truncating one exaggerates it.", "check": {"lie_factor": 0.017, "severity_band": "severe understatement", "true_change": 1.121, "axis_span": 66.6, "shown_fraction": 0.0168}, "source_citation": "Our World in Data / NASA GISTEMP global temperature anomaly, 1880-2015 (CC-BY).", "source_url": "https://www.washingtonpost.com/news/the-fix/wp/2015/12/14/why-the-national-reviews-global-temperature-graph-is-so-misleading/"}
+{"chart_id": "national_review_climate_2015", "kind": "real", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "expanded_y_axis", "prompt": "Chart: \"The only #climatechange chart you need to see\" \u2014 global temperature (National Review (chart from Power Line), 2015)\nType: line chart\nX axis: Year, 1880 to 2015\nY axis: degrees C anomaly, displayed range -23.3 to 43.3 (a span of 67).\n\nPlotted values (every ~20 years):\n  1880: 0.0608\n  1899: 0.0197\n  1918: -0.0621\n  1937: 0.3479\n  1956: 0.099\n  1975: 0.258\n  1994: 0.5914\n  2013: 0.9826\n\nQuestion: The line looks flat. Has global temperature been stable over this period?", "response": "No. The plotted values rise from 0.0608 to 1.1819, a change of 1.121 degrees C anomaly. The global temperature anomaly rises about 1.1 C from 1880 to 2015, a real trend the axis flattens to near zero. The line looks flat only because the axis runs -23.3 to 43.3, far wider than the data.", "source_citation": "Our World in Data / NASA GISTEMP global temperature anomaly, 1880-2015 (CC-BY).", "source_url": "https://www.washingtonpost.com/news/the-fix/wp/2015/12/14/why-the-national-reviews-global-temperature-graph-is-so-misleading/"}
+{"chart_id": "reuters_florida_guns_2014", "kind": "real", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "inverted_y_axis", "prompt": "Chart: Gun deaths in Florida, before and after Stand Your Ground (Reuters / Business Insider (chart by C. Chan), 2014)\nType: line chart\nY axis: firearm deaths, INVERTED (zero at the top, larger values lower).\n\nPlotted values:\n  2005: 521\n  2006: 740\n  2007: 825\n  2012: 721\n\nQuestion: The line drops toward the bottom over time. Did gun deaths fall?", "response": "No. The y-axis is inverted, so a line moving down means the count went up. Firearm deaths rose after the 2005 law, from 521 in 2005 to 825 in 2007, and stayed well above the pre-law level. The apparent fall is the inverted axis reversing a real rise.", "source_citation": "Firearm deaths in Florida: 521 (2005), 740 (2006), 825 (2007), 721 (2012). Humphreys et al., JAMA Internal Medicine 2017; Live Science reporting.", "source_url": "https://www.livescience.com/45083-misleading-gun-death-chart.html"}
+{"chart_id": "reuters_florida_guns_2014", "kind": "real", "question_type": "lie_factor", "label": "severity", "mechanism": "inverted_y_axis", "prompt": "Chart: Gun deaths in Florida, before and after Stand Your Ground (Reuters / Business Insider (chart by C. Chan), 2014)\nType: line chart\nY axis: firearm deaths, INVERTED (zero at the top, larger values lower).\n\nPlotted values:\n  2005: 521\n  2006: 740\n  2007: 825\n  2012: 721\n\nQuestion: On Tufte's Lie Factor (1.0 = honest, negative = the chart reverses the trend), how does this chart's axis affect the trend?", "response": "Lie Factor -1.0, severe. The counts are plotted at their true magnitude but the axis is inverted, so the trend reads backwards: firearm deaths rose after the 2005 law, from 521 in 2005 to 825 in 2007, and stayed well above the pre-law level, shown as a decline.", "check": {"lie_factor": -1.0, "severity_band": "severe", "direction_flip": true}, "source_citation": "Firearm deaths in Florida: 521 (2005), 740 (2006), 825 (2007), 721 (2012). Humphreys et al., JAMA Internal Medicine 2017; Live Science reporting.", "source_url": "https://www.livescience.com/45083-misleading-gun-death-chart.html"}
+{"chart_id": "aul_planned_parenthood_2015", "kind": "real", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "dual_axis_mismatch", "prompt": "Chart: Abortions versus cancer screenings at Planned Parenthood (Americans United for Life, shown at a US House hearing, 2015)\nType: line chart with TWO y-axes (dual axis)\nLeft axis: Cancer screening and prevention services = 935,573\nRight axis: Abortions = 327,000\nEach series is drawn on its own scale, so the two line heights are not comparable.\n\nQuestion: The abortions line is drawn as high as the screenings line. Are the two about equal?", "response": "No. In 2013 there were 935,573 cancer screening and prevention services and 327,000 abortions, so screenings outnumbered abortions about 2.9 to 1. The lines look comparable only because each has its own y-axis; on one shared scale the screenings line is about 2.86x higher.", "source_citation": "Planned Parenthood 2013 annual figures: 935,573 cancer screening and prevention services, 327,000 abortions (PolitiFact, Boston Globe, 2015).", "source_url": "https://www.politifact.com/factchecks/2015/oct/01/jason-chaffetz/chart-shown-planned-parenthood-hearing-misleading-/"}
+{"chart_id": "aul_planned_parenthood_2015", "kind": "real", "question_type": "lie_factor", "label": "severity", "mechanism": "dual_axis_mismatch", "prompt": "Chart: Abortions versus cancer screenings at Planned Parenthood (Americans United for Life, shown at a US House hearing, 2015)\nType: line chart with TWO y-axes (dual axis)\nLeft axis: Cancer screening and prevention services = 935,573\nRight axis: Abortions = 327,000\nEach series is drawn on its own scale, so the two line heights are not comparable.\n\nQuestion: On Tufte's Lie Factor (1.0 = honest, higher = the design overstates an equivalence), how severely does the dual axis distort the comparison?", "response": "Lie Factor about 2.86, severe. Two mismatched y-axes make a 2.86x gap read as rough equality: in 2013 there were 935,573 cancer screening and prevention services and 327,000 abortions, so screenings outnumbered abortions about 2.9 to 1.", "check": {"lie_factor": 2.86, "severity_band": "severe", "true_ratio": 2.86}, "source_citation": "Planned Parenthood 2013 annual figures: 935,573 cancer screening and prevention services, 327,000 abortions (PolitiFact, Boston Globe, 2015).", "source_url": "https://www.politifact.com/factchecks/2015/oct/01/jason-chaffetz/chart-shown-planned-parenthood-hearing-misleading-/"}
+{"chart_id": "adaption_plus16_claim_2026", "kind": "atlas", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "claim_not_supported", "prompt": "Chart: AutoScientist augmentation panel: '+16 win-rate points beyond 20,000 datapoints' (Adaption AutoScientist platform, 2026)\nType: claim\nReported mechanism: claim not supported\n\nValues as published:\n  {\"label\": \"1,882 verified rows\", \"category_win_rate\": 76}\n  {\"label\": \"21,882 augmented rows\", \"category_win_rate\": 72}\n\nQuestion: The chart's design suggests otherwise. What does the data actually show, and by what mechanism is it distorted?", "response": "At matched base model and epochs, crossing 20,000 rows lowered the held-out category win rate from 76 to 72, so the observed effect is minus 4, not the advertised plus 16. The mechanism is claim not supported: an advertised effect the data contradicts.", "source_citation": "This project's controlled runs (Llama-4-Scout, 3 epochs): 1,882 verified rows scored 76; 21,882 augmented rows scored 72.", "source_url": "https://adaptionlabs.ai"}
+{"chart_id": "adaption_plus16_claim_2026", "kind": "atlas", "question_type": "lie_factor", "label": "severity", "mechanism": "claim_not_supported", "prompt": "Chart: AutoScientist augmentation panel: '+16 win-rate points beyond 20,000 datapoints' (Adaption AutoScientist platform, 2026)\nType: claim\nReported mechanism: claim not supported\n\nValues as published:\n  {\"label\": \"1,882 verified rows\", \"category_win_rate\": 76}\n  {\"label\": \"21,882 augmented rows\", \"category_win_rate\": 72}\n\nQuestion: On Tufte's Lie Factor (1.0 honest, above 1 exaggerates, below 1 under-states, negative reverses), grade this chart's distortion.", "response": "Lie Factor not a single defined factor, severe. An advertised effect the data contradicts. The data says: at matched base model and epochs, crossing 20,000 rows lowered the held-out category win rate from 76 to 72, so the observed effect is minus 4, not the advertised plus 16.", "check": {"lie_factor": null, "severity_band": "severe"}, "source_citation": "This project's controlled runs (Llama-4-Scout, 3 epochs): 1,882 verified rows scored 76; 21,882 augmented rows scored 72.", "source_url": "https://adaptionlabs.ai"}
+{"chart_id": "lipitor_36pct_relative_risk_2008", "kind": "atlas", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "percent_absolute_swap", "prompt": "Chart: Lipitor 'reduces risk of heart attack by 36%' direct-to-consumer ad (Pfizer DTC advertisement for Lipitor (atorvastatin), 2008)\nType: risk-reduction graphic\nReported mechanism: percent absolute swap\n\nValues as published:\n  {\"label\": \"placebo heart-attack rate over ~3.3 yr (%)\", \"value\": 3.0}\n  {\"label\": \"Lipitor heart-attack rate over ~3.3 yr (%)\", \"value\": 1.9}\n  {\"label\": \"absolute risk reduction (percentage points)\", \"value\": 1.1}\n  {\"label\": \"relative risk reduction as advertised (%)\", \"value\": 36}\n(inferred: NNT about 91 computed as 1/0.011; ad publication year approximate)\n\nQuestion: The chart's design suggests otherwise. What does the data actually show, and by what mechanism is it distorted?", "response": "Over about 3.3 years heart attacks fell from 3.0% to 1.9%, an absolute reduction of 1.1 percentage points; about 91 patients must be treated for 3.3 years to prevent one heart attack, and the advertised 36% is just 1.1 divided by 3.0. The mechanism is percent absolute swap: a relative figure stands in for a small absolute change.", "source_citation": "Sever PS et al., ASCOT-LLA, Lancet 2003;361:1149-1158: atorvastatin 100 events (1.9%) vs placebo 154 events (3.0%), HR 0.64, N=10,305.", "source_url": "https://6xc.medium.com/the-marketing-of-lipitor-49a93089d8cb"}
+{"chart_id": "lipitor_36pct_relative_risk_2008", "kind": "atlas", "question_type": "lie_factor", "label": "severity", "mechanism": "percent_absolute_swap", "prompt": "Chart: Lipitor 'reduces risk of heart attack by 36%' direct-to-consumer ad (Pfizer DTC advertisement for Lipitor (atorvastatin), 2008)\nType: risk-reduction graphic\nReported mechanism: percent absolute swap\n\nValues as published:\n  {\"label\": \"placebo heart-attack rate over ~3.3 yr (%)\", \"value\": 3.0}\n  {\"label\": \"Lipitor heart-attack rate over ~3.3 yr (%)\", \"value\": 1.9}\n  {\"label\": \"absolute risk reduction (percentage points)\", \"value\": 1.1}\n  {\"label\": \"relative risk reduction as advertised (%)\", \"value\": 36}\n(inferred: NNT about 91 computed as 1/0.011; ad publication year approximate)\n\nQuestion: On Tufte's Lie Factor (1.0 honest, above 1 exaggerates, below 1 under-states, negative reverses), grade this chart's distortion.", "response": "Lie Factor not a single defined factor, severe. A relative figure stands in for a small absolute change. The data says: over about 3.3 years heart attacks fell from 3.0% to 1.9%, an absolute reduction of 1.1 percentage points; about 91 patients must be treated for 3.3 years to prevent one heart attack, and the advertised 36% is just 1.1 divided by 3.0.", "check": {"lie_factor": null, "severity_band": "severe"}, "source_citation": "Sever PS et al., ASCOT-LLA, Lancet 2003;361:1149-1158: atorvastatin 100 events (1.9%) vs placebo 154 events (3.0%), HR 0.64, N=10,305.", "source_url": "https://6xc.medium.com/the-marketing-of-lipitor-49a93089d8cb"}
+{"chart_id": "apple_3d_pie_smartphone_2008", "kind": "atlas", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "pie_3d", "prompt": "Chart: Steve Jobs' 3-D pie chart of US smartphone market share (Macworld 2008) (Apple Macworld 2008 keynote slide (Steve Jobs, Jan 15 2008), 2008)\nType: 3D pie chart\nReported mechanism: pie 3d\n\nValues as published:\n  {\"label\": \"RIM (BlackBerry)\", \"value\": 39.0}\n  {\"label\": \"Other\", \"value\": 21.2}\n  {\"label\": \"Apple (iPhone)\", \"value\": 19.5}\n  {\"label\": \"Palm\", \"value\": 9.8}\n  {\"label\": \"Motorola\", \"value\": 7.4}\n  {\"label\": \"Nokia\", \"value\": 3.1}\n(inferred: data period cited as 2007, exact quarter not independently confirmed)\n\nQuestion: The chart's design suggests otherwise. What does the data actually show, and by what mechanism is it distorted?", "response": "RIM's 39.0% is almost exactly double Apple's 19.5%, and the Other catch-all at 21.2% was larger than Apple's slice; on a flat correctly-scaled pie Apple's wedge is clearly less than half of RIM's. The mechanism is pie 3d: 3D perspective enlarges the front slice.", "source_citation": "Canalys US smartphone market-share data (2007) as shown on Apple's Macworld 2008 slide; values sum to 100.0%.", "source_url": "https://gist.github.com/bljustice/d7af0a24c628a4ec992a350bdc137a46"}
+{"chart_id": "apple_3d_pie_smartphone_2008", "kind": "atlas", "question_type": "lie_factor", "label": "severity", "mechanism": "pie_3d", "prompt": "Chart: Steve Jobs' 3-D pie chart of US smartphone market share (Macworld 2008) (Apple Macworld 2008 keynote slide (Steve Jobs, Jan 15 2008), 2008)\nType: 3D pie chart\nReported mechanism: pie 3d\n\nValues as published:\n  {\"label\": \"RIM (BlackBerry)\", \"value\": 39.0}\n  {\"label\": \"Other\", \"value\": 21.2}\n  {\"label\": \"Apple (iPhone)\", \"value\": 19.5}\n  {\"label\": \"Palm\", \"value\": 9.8}\n  {\"label\": \"Motorola\", \"value\": 7.4}\n  {\"label\": \"Nokia\", \"value\": 3.1}\n(inferred: data period cited as 2007, exact quarter not independently confirmed)\n\nQuestion: On Tufte's Lie Factor (1.0 honest, above 1 exaggerates, below 1 under-states, negative reverses), grade this chart's distortion.", "response": "Lie Factor not a single defined factor, moderate. 3D perspective enlarges the front slice. The data says: RIM's 39.0% is almost exactly double Apple's 19.5%, and the Other catch-all at 21.2% was larger than Apple's slice; on a flat correctly-scaled pie Apple's wedge is clearly less than half of RIM's.", "check": {"lie_factor": null, "severity_band": "moderate"}, "source_citation": "Canalys US smartphone market-share data (2007) as shown on Apple's Macworld 2008 slide; values sum to 100.0%.", "source_url": "https://gist.github.com/bljustice/d7af0a24c628a4ec992a350bdc137a46"}
+{"chart_id": "georgia_dph_covid_reordered_dates_2020", "kind": "atlas", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "reordered_x", "prompt": "Chart: Top 5 Counties with the Greatest Number of Confirmed COVID-19 Cases (Georgia Department of Public Health, 2020)\nType: grouped vertical bar chart (5 counties per date, 14 dates)\nReported mechanism: reordered x\n\nValues as published:\n  {\"label\": \"x-axis dates as displayed\", \"value\": \"Apr 28, Apr 27, Apr 29, May 1, Apr 30, May 4, May 6, May 5, May 2, May 7, Apr 26, May 3, May 8, May 9\"}\n  {\"label\": \"x-axis in true chronological order\", \"value\": \"Apr 26, Apr 27, Apr 28, Apr 29, Apr 30, May 1, May 2, May 3, May 4, May 5, May 6, May 7, May 8, May 9\"}\n  {\"label\": \"counties re-sorted largest-to-smallest each day\", \"value\": \"Fulton, Cobb, Gwinnett, DeKalb, Hall\"}\n(inferred: per-cell daily counts never published as a table; the documented manipulation is the ordering itself, not invented values)\n\nQuestion: The chart's design suggests otherwise. What does the data actually show, and by what mechanism is it distorted?", "response": "In correct chronological order the descending staircase disappears; cases were roughly flat to gently declining, and the apparent late drop was largely a reporting-lag artifact; DPH apologized and attributed it to incorrect sorting logic. The mechanism is reordered x: non-chronological order manufactures a trend that is not in the data.", "source_citation": "Georgia DPH bar chart posted ~May 10 2020; documented by Route Fifty, Georgia Public Broadcasting, and the Atlanta Journal-Constitution; displayed date order reproduced at covid-georgia.com.", "source_url": "https://www.route-fifty.com/management/2020/05/officials-apologize-retool-coronavirus-graph-after-backlash/165478/"}
+{"chart_id": "georgia_dph_covid_reordered_dates_2020", "kind": "atlas", "question_type": "lie_factor", "label": "severity", "mechanism": "reordered_x", "prompt": "Chart: Top 5 Counties with the Greatest Number of Confirmed COVID-19 Cases (Georgia Department of Public Health, 2020)\nType: grouped vertical bar chart (5 counties per date, 14 dates)\nReported mechanism: reordered x\n\nValues as published:\n  {\"label\": \"x-axis dates as displayed\", \"value\": \"Apr 28, Apr 27, Apr 29, May 1, Apr 30, May 4, May 6, May 5, May 2, May 7, Apr 26, May 3, May 8, May 9\"}\n  {\"label\": \"x-axis in true chronological order\", \"value\": \"Apr 26, Apr 27, Apr 28, Apr 29, Apr 30, May 1, May 2, May 3, May 4, May 5, May 6, May 7, May 8, May 9\"}\n  {\"label\": \"counties re-sorted largest-to-smallest each day\", \"value\": \"Fulton, Cobb, Gwinnett, DeKalb, Hall\"}\n(inferred: per-cell daily counts never published as a table; the documented manipulation is the ordering itself, not invented values)\n\nQuestion: On Tufte's Lie Factor (1.0 honest, above 1 exaggerates, below 1 under-states, negative reverses), grade this chart's distortion.", "response": "Lie Factor not a single defined factor, severe. Non-chronological order manufactures a trend that is not in the data. The data says: in correct chronological order the descending staircase disappears; cases were roughly flat to gently declining, and the apparent late drop was largely a reporting-lag artifact; DPH apologized and attributed it to incorrect sorting logic.", "check": {"lie_factor": null, "severity_band": "severe"}, "source_citation": "Georgia DPH bar chart posted ~May 10 2020; documented by Route Fifty, Georgia Public Broadcasting, and the Atlanta Journal-Constitution; displayed date order reproduced at covid-georgia.com.", "source_url": "https://www.route-fifty.com/management/2020/05/officials-apologize-retool-coronavirus-graph-after-backlash/165478/"}
+{"chart_id": "fox_welfare_vs_fulltime_work_2013", "kind": "atlas", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "truncated_y_axis", "prompt": "Chart: Number of Americans on Welfare vs. Number with Full-Time Jobs (Fox News (Fox and Friends), 2013)\nType: two-bar vertical bar chart\nReported mechanism: truncated y axis\n\nValues as published:\n  {\"label\": \"people in households receiving federal means-tested benefits, Q4 2011\", \"value\": 108600000}\n  {\"label\": \"people with full-time jobs, 2012\", \"value\": 101700000}\n(inferred: y-axis baseline not published; inferred ~99.98M by solving (108.6-b)/(101.7-b)=5 for the documented five-times-taller visual; the two data values are documented, not inferred)\n\nQuestion: The chart's design suggests otherwise. What does the data actually show, and by what mechanism is it distorted?", "response": "108.6 million versus 101.7 million is a ratio of about 1.07 to 1, nearly equal; on a zero-baseline axis the bars would be almost the same height. The mechanism is truncated y axis: bars drawn 5.01x apart for a 1.07x value gap.", "source_citation": "Fox and Friends, Oct 28 2013. Benefits: US Census Bureau SIPP Q4 2011, Table 2. Full-time workers: US Census Bureau CPS 2012. Reported by Media Matters.", "source_url": "https://www.mediamatters.org/fox-friends/dishonest-fox-chart-overstates-comparison-welfare-full-time-work-500-percent"}
+{"chart_id": "fox_welfare_vs_fulltime_work_2013", "kind": "atlas", "question_type": "lie_factor", "label": "severity", "mechanism": "truncated_y_axis", "prompt": "Chart: Number of Americans on Welfare vs. Number with Full-Time Jobs (Fox News (Fox and Friends), 2013)\nType: two-bar vertical bar chart\nReported mechanism: truncated y axis\n\nValues as published:\n  {\"label\": \"people in households receiving federal means-tested benefits, Q4 2011\", \"value\": 108600000}\n  {\"label\": \"people with full-time jobs, 2012\", \"value\": 101700000}\n(inferred: y-axis baseline not published; inferred ~99.98M by solving (108.6-b)/(101.7-b)=5 for the documented five-times-taller visual; the two data values are documented, not inferred)\n\nQuestion: On Tufte's Lie Factor (1.0 honest, above 1 exaggerates, below 1 under-states, negative reverses), grade this chart's distortion.", "response": "Lie Factor 4.69, severe. Bars drawn 5.01x apart for a 1.07x value gap. The data says: 108.6 million versus 101.7 million is a ratio of about 1.07 to 1, nearly equal; on a zero-baseline axis the bars would be almost the same height.", "check": {"lie_factor": 4.69, "severity_band": "severe"}, "source_citation": "Fox and Friends, Oct 28 2013. Benefits: US Census Bureau SIPP Q4 2011, Table 2. Full-time workers: US Census Bureau CPS 2012. Reported by Media Matters.", "source_url": "https://www.mediamatters.org/fox-friends/dishonest-fox-chart-overstates-comparison-welfare-full-time-work-500-percent"}
+{"chart_id": "fox_gas_prices_cherrypick_feb2012", "kind": "atlas", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "cherry_picked_window", "prompt": "Chart: Rising Gas Prices Under Obama (three-point chart) (Fox News (America's Newsroom), 2012)\nType: line chart plotting only ~3 points across ~1 year\nReported mechanism: cherry picked window\n\nValues as published:\n  {\"year\": \"Jan 2009 trough baseline\", \"value\": 1.788}\n  {\"year\": \"Feb 2011\", \"value\": 3.211}\n  {\"year\": \"May 2011 actual peak, omitted\", \"value\": 3.906}\n  {\"year\": \"Dec 2011\", \"value\": 3.266}\n  {\"year\": \"Jan 2012\", \"value\": 3.38}\n  {\"year\": \"Feb 2012 framed as the high\", \"value\": 3.579}\n(inferred: Fox's exact plotted values were not transcribed; the listed real_values are the true EIA data showing the omitted May 2011 peak exceeded the Feb 2012 value)\n\nQuestion: The chart's design suggests otherwise. What does the data actually show, and by what mechanism is it distorted?", "response": "Prices peaked at $3.906 in May 2011, fell to $3.266 by December 2011, and only partly rebounded to $3.579 in Feb 2012, still below the 2011 peak; the full EIA series is a bumpy plateau, not a steady climb. The mechanism is cherry picked window: a sub-window trends opposite to the full series.", "source_citation": "Fox News, Feb 2012, per Media Matters. Real prices: US EIA monthly regular gasoline series EMM_EPMR_PTE_NUS_DPG.", "source_url": "https://mediamatters.org/blog/2012/02/21/fox-still-struggling-with-basic-chart-concepts/185049"}
+{"chart_id": "fox_gas_prices_cherrypick_feb2012", "kind": "atlas", "question_type": "lie_factor", "label": "severity", "mechanism": "cherry_picked_window", "prompt": "Chart: Rising Gas Prices Under Obama (three-point chart) (Fox News (America's Newsroom), 2012)\nType: line chart plotting only ~3 points across ~1 year\nReported mechanism: cherry picked window\n\nValues as published:\n  {\"year\": \"Jan 2009 trough baseline\", \"value\": 1.788}\n  {\"year\": \"Feb 2011\", \"value\": 3.211}\n  {\"year\": \"May 2011 actual peak, omitted\", \"value\": 3.906}\n  {\"year\": \"Dec 2011\", \"value\": 3.266}\n  {\"year\": \"Jan 2012\", \"value\": 3.38}\n  {\"year\": \"Feb 2012 framed as the high\", \"value\": 3.579}\n(inferred: Fox's exact plotted values were not transcribed; the listed real_values are the true EIA data showing the omitted May 2011 peak exceeded the Feb 2012 value)\n\nQuestion: On Tufte's Lie Factor (1.0 honest, above 1 exaggerates, below 1 under-states, negative reverses), grade this chart's distortion.", "response": "Lie Factor not a single defined factor, severe. A sub-window trends opposite to the full series. The data says: prices peaked at $3.906 in May 2011, fell to $3.266 by December 2011, and only partly rebounded to $3.579 in Feb 2012, still below the 2011 peak; the full EIA series is a bumpy plateau, not a steady climb.", "check": {"lie_factor": null, "severity_band": "severe"}, "source_citation": "Fox News, Feb 2012, per Media Matters. Real prices: US EIA monthly regular gasoline series EMM_EPMR_PTE_NUS_DPG.", "source_url": "https://mediamatters.org/blog/2012/02/21/fox-still-struggling-with-basic-chart-concepts/185049"}
+{"chart_id": "irs_3d_pie_federal_budget_fy2010", "kind": "atlas", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "pie_3d", "prompt": "Chart: Major Categories of Federal Income and Outlays for Fiscal Year 2010 (US Internal Revenue Service (Form 1040 instruction booklet), 2011)\nType: 3D-perspective pie chart\nReported mechanism: pie 3d\n\nValues as published:\n  {\"label\": \"Income: borrowing to cover deficit\", \"value\": 37}\n  {\"label\": \"Income: personal income taxes\", \"value\": 26}\n  {\"label\": \"Income: social security/medicare/retirement taxes\", \"value\": 25}\n  {\"label\": \"Income: corporate income taxes\", \"value\": 6}\n  {\"label\": \"Income: excise/customs/estate/gift/misc\", \"value\": 6}\n(inferred: none; all percentages printed by the IRS and read directly from page 97)\n\nQuestion: The chart's design suggests otherwise. What does the data actually show, and by what mechanism is it distorted?", "response": "On an undistorted 2D pie borrowing was the single largest income slice at 37 percent, about 1.4 times personal income taxes at 26 percent, meaning roughly 37 cents of every federal dollar in FY2010 was borrowed. The mechanism is pie 3d: 3D perspective enlarges the front slice.", "source_citation": "IRS 2011 Instructions for Form 1040, page 97; also critiqued by Naomi Robbins, Forbes, Apr 16 2012.", "source_url": "https://www.irs.gov/pub/irs-prior/i1040gi--2011.pdf"}
+{"chart_id": "irs_3d_pie_federal_budget_fy2010", "kind": "atlas", "question_type": "lie_factor", "label": "severity", "mechanism": "pie_3d", "prompt": "Chart: Major Categories of Federal Income and Outlays for Fiscal Year 2010 (US Internal Revenue Service (Form 1040 instruction booklet), 2011)\nType: 3D-perspective pie chart\nReported mechanism: pie 3d\n\nValues as published:\n  {\"label\": \"Income: borrowing to cover deficit\", \"value\": 37}\n  {\"label\": \"Income: personal income taxes\", \"value\": 26}\n  {\"label\": \"Income: social security/medicare/retirement taxes\", \"value\": 25}\n  {\"label\": \"Income: corporate income taxes\", \"value\": 6}\n  {\"label\": \"Income: excise/customs/estate/gift/misc\", \"value\": 6}\n(inferred: none; all percentages printed by the IRS and read directly from page 97)\n\nQuestion: On Tufte's Lie Factor (1.0 honest, above 1 exaggerates, below 1 under-states, negative reverses), grade this chart's distortion.", "response": "Lie Factor not a single defined factor, moderate. 3D perspective enlarges the front slice. The data says: on an undistorted 2D pie borrowing was the single largest income slice at 37 percent, about 1.4 times personal income taxes at 26 percent, meaning roughly 37 cents of every federal dollar in FY2010 was borrowed.", "check": {"lie_factor": null, "severity_band": "moderate"}, "source_citation": "IRS 2011 Instructions for Form 1040, page 97; also critiqued by Naomi Robbins, Forbes, Apr 16 2012.", "source_url": "https://www.irs.gov/pub/irs-prior/i1040gi--2011.pdf"}
+{"chart_id": "fox_unemployment_broken_scale_2011", "kind": "atlas", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "expanded_y_axis", "prompt": "Chart: Unemployment Rate Under Obama (Nov 2011 plotted as not-falling) (Fox News, 2011)\nType: line chart of the monthly unemployment rate\nReported mechanism: expanded y axis\n\nValues as published:\n  {\"label\": \"Mar 2011 U-3 rate\", \"value\": 8.8}\n  {\"label\": \"Oct 2011 U-3 rate\", \"value\": 9.0}\n  {\"label\": \"Nov 2011 U-3 rate\", \"value\": 8.6}\n(inferred: mechanism mapped to expanded_y_axis as the closest fit; strictly a broken non-linear scale; the three percentages are real BLS figures, none inferred)\n\nQuestion: The chart's design suggests otherwise. What does the data actually show, and by what mechanism is it distorted?", "response": "On a correct linear axis 8.6 is below 8.8 is below 9.0, so November at 8.6 percent is the lowest of the three points and the line clearly declines from October to November; unemployment was falling and the scale concealed it. The mechanism is expanded y axis: expanded axis hides the trend (factor needs the axis span).", "source_citation": "Fox News, ~Dec 9-12 2011, per Media Matters (Dec 12 2011): the chart plotted Nov 8.6% level with Oct 9.0% and above Mar 8.8%. BLS U-3 rates. Also Business Insider and the Washington Post.", "source_url": "https://www.mediamatters.org/laura-ingraham/today-dishonest-fox-news-graphics"}
+{"chart_id": "fox_unemployment_broken_scale_2011", "kind": "atlas", "question_type": "lie_factor", "label": "severity", "mechanism": "expanded_y_axis", "prompt": "Chart: Unemployment Rate Under Obama (Nov 2011 plotted as not-falling) (Fox News, 2011)\nType: line chart of the monthly unemployment rate\nReported mechanism: expanded y axis\n\nValues as published:\n  {\"label\": \"Mar 2011 U-3 rate\", \"value\": 8.8}\n  {\"label\": \"Oct 2011 U-3 rate\", \"value\": 9.0}\n  {\"label\": \"Nov 2011 U-3 rate\", \"value\": 8.6}\n(inferred: mechanism mapped to expanded_y_axis as the closest fit; strictly a broken non-linear scale; the three percentages are real BLS figures, none inferred)\n\nQuestion: On Tufte's Lie Factor (1.0 honest, above 1 exaggerates, below 1 under-states, negative reverses), grade this chart's distortion.", "response": "Lie Factor not a single defined factor, understatement. Expanded axis hides the trend (factor needs the axis span). The data says: on a correct linear axis 8.6 is below 8.8 is below 9.0, so November at 8.6 percent is the lowest of the three points and the line clearly declines from October to November; unemployment was falling and the scale concealed it.", "check": {"lie_factor": null, "severity_band": "understatement"}, "source_citation": "Fox News, ~Dec 9-12 2011, per Media Matters (Dec 12 2011): the chart plotted Nov 8.6% level with Oct 9.0% and above Mar 8.8%. BLS U-3 rates. Also Business Insider and the Washington Post.", "source_url": "https://www.mediamatters.org/laura-ingraham/today-dishonest-fox-news-graphics"}
+{"chart_id": "fox_govt_aid_recipients_truncated_2012", "kind": "atlas", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "truncated_y_axis", "prompt": "Chart: Americans on Government Aid, 2009-2011 (truncated-axis growth chart) (Fox News (America's Newsroom), 2012)\nType: column growth chart of people on government aid\nReported mechanism: truncated y axis\n\nValues as published:\n  {\"label\": \"people in a household receiving means-tested aid, ~2009\", \"value\": 96000000}\n  {\"label\": \"people in a household receiving means-tested aid, ~2011\", \"value\": 108000000}\n(inferred: y-axis baseline 94,000,000 is documented by Media Matters, as is the under-12% rise; the ~96M and ~108M endpoints are approximate because the chart showed no numeric labels; the ~7x height ratio is arithmetic from the 94M baseline)\n\nQuestion: The chart's design suggests otherwise. What does the data actually show, and by what mechanism is it distorted?", "response": "The increase was under 12 percent over 2009 to 2011, a modest step on a zero baseline; from the 94M baseline the later bar is about 7 times the height of the earlier one, grossly exaggerating a recession-driven rise. The mechanism is truncated y axis: bars drawn 7.0x apart for a 1.12x value gap.", "source_citation": "Fox News, Aug 9 2012, per Media Matters; figures from Sen. Sessions' office based on US Census Bureau means-tested-aid data, 2009-2011.", "source_url": "https://www.mediamatters.org/fox-news/today-dishonest-fox-charts-government-aid-edition"}
+{"chart_id": "fox_govt_aid_recipients_truncated_2012", "kind": "atlas", "question_type": "lie_factor", "label": "severity", "mechanism": "truncated_y_axis", "prompt": "Chart: Americans on Government Aid, 2009-2011 (truncated-axis growth chart) (Fox News (America's Newsroom), 2012)\nType: column growth chart of people on government aid\nReported mechanism: truncated y axis\n\nValues as published:\n  {\"label\": \"people in a household receiving means-tested aid, ~2009\", \"value\": 96000000}\n  {\"label\": \"people in a household receiving means-tested aid, ~2011\", \"value\": 108000000}\n(inferred: y-axis baseline 94,000,000 is documented by Media Matters, as is the under-12% rise; the ~96M and ~108M endpoints are approximate because the chart showed no numeric labels; the ~7x height ratio is arithmetic from the 94M baseline)\n\nQuestion: On Tufte's Lie Factor (1.0 honest, above 1 exaggerates, below 1 under-states, negative reverses), grade this chart's distortion.", "response": "Lie Factor 6.22, severe. Bars drawn 7.0x apart for a 1.12x value gap. The data says: the increase was under 12 percent over 2009 to 2011, a modest step on a zero baseline; from the 94M baseline the later bar is about 7 times the height of the earlier one, grossly exaggerating a recession-driven rise.", "check": {"lie_factor": 6.22, "severity_band": "severe"}, "source_citation": "Fox News, Aug 9 2012, per Media Matters; figures from Sen. Sessions' office based on US Census Bureau means-tested-aid data, 2009-2011.", "source_url": "https://www.mediamatters.org/fox-news/today-dishonest-fox-charts-government-aid-edition"}
+{"chart_id": "cruz_satellite_no_warming_2015", "kind": "atlas", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "cherry_picked_window", "prompt": "Chart: 'Satellite data show zero warming for 17-18 years' (series started at the 1998 El Nino) (Sen. Ted Cruz, US Senate hearing and media appearances, 2015)\nType: line\nReported mechanism: cherry picked window\n\nValues as published:\n  {\"year\": 1997, \"value\": 0.46}\n  {\"year\": 1998, \"value\": 0.61}\n  {\"year\": 2005, \"value\": 0.68}\n  {\"year\": 2010, \"value\": 0.72}\n  {\"year\": 2014, \"value\": 0.75}\n  {\"year\": 2015, \"value\": 0.9}\n  {\"year\": 2016, \"value\": 1.01}\n(inferred: real_values are the exact NASA GISTEMP surface record over the window the zero-warming claim contradicts; RSS v3.3 slope not stated as no exact citable figure was obtained)\n\nQuestion: The chart's design suggests otherwise. What does the data actually show, and by what mechanism is it distorted?", "response": "The authoritative surface record rose from +0.46 C in 1997 to +0.90 C in 2015, about +0.44 C in 18 years, and 2014, 2015 and 2016 each set successive records; the satellite dataset was later revised upward, erasing the apparent pause. The mechanism is cherry picked window: a sub-window trends opposite to the full series.", "source_citation": "NASA GISS GISTEMP v4 annual means; PolitiFact 2015-03-20; Carbon Brief on RSS/UAH trends; Skeptical Science on the RSS v3.3 to v4 revision.", "source_url": "https://www.politifact.com/factchecks/2015/mar/20/ted-cruz/ted-cruzs-worlds-fire-not-last-17-years/"}
+{"chart_id": "cruz_satellite_no_warming_2015", "kind": "atlas", "question_type": "lie_factor", "label": "severity", "mechanism": "cherry_picked_window", "prompt": "Chart: 'Satellite data show zero warming for 17-18 years' (series started at the 1998 El Nino) (Sen. Ted Cruz, US Senate hearing and media appearances, 2015)\nType: line\nReported mechanism: cherry picked window\n\nValues as published:\n  {\"year\": 1997, \"value\": 0.46}\n  {\"year\": 1998, \"value\": 0.61}\n  {\"year\": 2005, \"value\": 0.68}\n  {\"year\": 2010, \"value\": 0.72}\n  {\"year\": 2014, \"value\": 0.75}\n  {\"year\": 2015, \"value\": 0.9}\n  {\"year\": 2016, \"value\": 1.01}\n(inferred: real_values are the exact NASA GISTEMP surface record over the window the zero-warming claim contradicts; RSS v3.3 slope not stated as no exact citable figure was obtained)\n\nQuestion: On Tufte's Lie Factor (1.0 honest, above 1 exaggerates, below 1 under-states, negative reverses), grade this chart's distortion.", "response": "Lie Factor not a single defined factor, severe. A sub-window trends opposite to the full series. The data says: the authoritative surface record rose from +0.46 C in 1997 to +0.90 C in 2015, about +0.44 C in 18 years, and 2014, 2015 and 2016 each set successive records; the satellite dataset was later revised upward, erasing the apparent pause.", "check": {"lie_factor": null, "severity_band": "severe"}, "source_citation": "NASA GISS GISTEMP v4 annual means; PolitiFact 2015-03-20; Carbon Brief on RSS/UAH trends; Skeptical Science on the RSS v3.3 to v4 revision.", "source_url": "https://www.politifact.com/factchecks/2015/mar/20/ted-cruz/ted-cruzs-worlds-fire-not-last-17-years/"}
+{"chart_id": "rose_mail_global_warming_stopped_2012", "kind": "atlas", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "cherry_picked_window", "prompt": "Chart: 'Global warming stopped 16 years ago' (HadCRUT4, 1997-2012) (The Mail on Sunday (David Rose), 2012)\nType: line\nReported mechanism: cherry picked window\n\nValues as published:\n  {\"year\": 1996, \"value\": 0.183}\n  {\"year\": 1997, \"value\": 0.39}\n  {\"year\": 1998, \"value\": 0.539}\n  {\"year\": 2005, \"value\": 0.545}\n  {\"year\": 2010, \"value\": 0.56}\n  {\"year\": 2012, \"value\": 0.47}\n  {\"year\": 2013, \"value\": 0.514}\n(inferred: real_values are exact HadCRUT4 annual means (same dataset the monthly chart used); +0.03 C/decade is the Met Office's stated regression value)\n\nQuestion: The chart's design suggests otherwise. What does the data actually show, and by what mechanism is it distorted?", "response": "The Met Office's own analysis of the window gave a trend of about +0.03 C/decade, small but positive and not statistically distinguishable from long-term warming; warming resumed unmistakably after, with 2015 and 2016 setting records. The mechanism is cherry picked window: a sub-window trends opposite to the full series.", "source_citation": "Met Office HadCRUT4 global annual series; Carbon Brief rebuttal; Grantham Institute LSE; the Met Office publicly rebutted the piece the same weekend.", "source_url": "https://www.carbonbrief.org/the-mail-on-sunday-gives-david-rose-space-to-repeat-old-and-wrong-claims-that-global-warming-has-stopped"}
+{"chart_id": "rose_mail_global_warming_stopped_2012", "kind": "atlas", "question_type": "lie_factor", "label": "severity", "mechanism": "cherry_picked_window", "prompt": "Chart: 'Global warming stopped 16 years ago' (HadCRUT4, 1997-2012) (The Mail on Sunday (David Rose), 2012)\nType: line\nReported mechanism: cherry picked window\n\nValues as published:\n  {\"year\": 1996, \"value\": 0.183}\n  {\"year\": 1997, \"value\": 0.39}\n  {\"year\": 1998, \"value\": 0.539}\n  {\"year\": 2005, \"value\": 0.545}\n  {\"year\": 2010, \"value\": 0.56}\n  {\"year\": 2012, \"value\": 0.47}\n  {\"year\": 2013, \"value\": 0.514}\n(inferred: real_values are exact HadCRUT4 annual means (same dataset the monthly chart used); +0.03 C/decade is the Met Office's stated regression value)\n\nQuestion: On Tufte's Lie Factor (1.0 honest, above 1 exaggerates, below 1 under-states, negative reverses), grade this chart's distortion.", "response": "Lie Factor not a single defined factor, severe. A sub-window trends opposite to the full series. The data says: the Met Office's own analysis of the window gave a trend of about +0.03 C/decade, small but positive and not statistically distinguishable from long-term warming; warming resumed unmistakably after, with 2015 and 2016 setting records.", "check": {"lie_factor": null, "severity_band": "severe"}, "source_citation": "Met Office HadCRUT4 global annual series; Carbon Brief rebuttal; Grantham Institute LSE; the Met Office publicly rebutted the piece the same weekend.", "source_url": "https://www.carbonbrief.org/the-mail-on-sunday-gives-david-rose-space-to-repeat-old-and-wrong-claims-that-global-warming-has-stopped"}
+{"chart_id": "antivax_measles_mortality_not_incidence", "kind": "atlas", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "percent_absolute_swap", "prompt": "Chart: 'Measles was already vanishing before the vaccine' (death rate shown as disease burden) (Recurring anti-vaccination graphic genre, rebutted by CDC and Our World in Data, 2015)\nType: line\nReported mechanism: percent absolute swap\n\nValues as published:\n  {\"year\": 1938, \"case_rate_per_100k\": 647.4, \"death_rate_per_100k\": 2.5}\n  {\"year\": 1958, \"case_rate_per_100k\": 438.0, \"death_rate_per_100k\": 0.3}\n  {\"year\": 1962, \"case_rate_per_100k\": 257.9, \"death_rate_per_100k\": 0.2}\n  {\"year\": 1967, \"case_rate_per_100k\": 31.2, \"death_rate_per_100k\": 0.04}\n  {\"year\": 1970, \"case_rate_per_100k\": 22.8, \"death_rate_per_100k\": 0.04}\n(inferred: the deception is metric substitution (mortality for incidence), mapped to the nearest mechanism; all per-100,000 rates are exact OWID/CDC values; vaccine licensed 1963)\n\nQuestion: The chart's design suggests otherwise. What does the data actually show, and by what mechanism is it distorted?", "response": "Case incidence, the measure of transmission, stayed high right up to the vaccine (257.9 per 100k in 1962) and only fell after mass vaccination (31.2 in 1967, 22.8 in 1970); better medicine cut deaths, only the vaccine cut cases; plotting incidence instead of mortality reverses the conclusion. The mechanism is percent absolute swap: a relative figure stands in for a small absolute change.", "source_citation": "Our World in Data 'measles-cases-and-death-rate' (US) from US historical vital statistics/CDC; CDC 'History of Measles' gives ~3-4 million cases/yr pre-vaccine.", "source_url": "https://ourworldindata.org/measles-vaccines-save-lives"}
+{"chart_id": "antivax_measles_mortality_not_incidence", "kind": "atlas", "question_type": "lie_factor", "label": "severity", "mechanism": "percent_absolute_swap", "prompt": "Chart: 'Measles was already vanishing before the vaccine' (death rate shown as disease burden) (Recurring anti-vaccination graphic genre, rebutted by CDC and Our World in Data, 2015)\nType: line\nReported mechanism: percent absolute swap\n\nValues as published:\n  {\"year\": 1938, \"case_rate_per_100k\": 647.4, \"death_rate_per_100k\": 2.5}\n  {\"year\": 1958, \"case_rate_per_100k\": 438.0, \"death_rate_per_100k\": 0.3}\n  {\"year\": 1962, \"case_rate_per_100k\": 257.9, \"death_rate_per_100k\": 0.2}\n  {\"year\": 1967, \"case_rate_per_100k\": 31.2, \"death_rate_per_100k\": 0.04}\n  {\"year\": 1970, \"case_rate_per_100k\": 22.8, \"death_rate_per_100k\": 0.04}\n(inferred: the deception is metric substitution (mortality for incidence), mapped to the nearest mechanism; all per-100,000 rates are exact OWID/CDC values; vaccine licensed 1963)\n\nQuestion: On Tufte's Lie Factor (1.0 honest, above 1 exaggerates, below 1 under-states, negative reverses), grade this chart's distortion.", "response": "Lie Factor not a single defined factor, severe. A relative figure stands in for a small absolute change. The data says: case incidence, the measure of transmission, stayed high right up to the vaccine (257.9 per 100k in 1962) and only fell after mass vaccination (31.2 in 1967, 22.8 in 1970); better medicine cut deaths, only the vaccine cut cases; plotting incidence instead of mortality reverses the conclusion.", "check": {"lie_factor": null, "severity_band": "severe"}, "source_citation": "Our World in Data 'measles-cases-and-death-rate' (US) from US historical vital statistics/CDC; CDC 'History of Measles' gives ~3-4 million cases/yr pre-vaccine.", "source_url": "https://ourworldindata.org/measles-vaccines-save-lives"}
+{"chart_id": "fox31_denver_covid_unequal_yaxis_2020", "kind": "atlas", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "unequal_bins", "prompt": "Chart: New Cases Per Day (COVID-19) with an irregular unlabeled y-axis (Fox 31 Denver (KDVR-TV), Colorado, 2020)\nType: line chart of new cases per day (15 days)\nReported mechanism: unequal bins\n\nValues as published:\n  {\"label\": \"y-axis gridline labels (drawn at equal spacing, numerically irregular)\", \"value\": \"30, 60, 90, 100, 130, 160, 190, 240, 250, 300, 350, 400\"}\n  {\"label\": \"y-axis interval sequence between labels\", \"value\": \"30, 30, 10, 30, 30, 30, 50, 10, 50, 50, 50\"}\n  {\"label\": \"date range\", \"value\": \"Mar 18 to Apr 1 2020\"}\n  {\"label\": \"total cases printed on the chart\", \"value\": 3342}\n(inferred: The verified artifact is the unequal y-axis label set and interval sequence, corroborated across three independent sources. The per-day plotted series (community reverse-engineering, c. 33 to 376) is NOT read from a table and sums to 3,159, which does not match the 3,342 total printed on the chart, so those daily values are treated as approximate and excluded from the ground-truth values above. Publisher Fox 31 Denver per saracdoan; freerangestats labels it 'Fox'.)\n\nQuestion: The chart's design suggests otherwise. What does the data actually show, and by what mechanism is it distorted?", "response": "On a uniform zero-based axis the same series climbs steeply, several-fold across the two-week window; because the gap from 90 to 100 (10 cases) is drawn the same size as 350 to 400 (50 cases), the design flattens and understates how fast cases were rising. The mechanism is unequal bins: unequal bin widths distort the apparent distribution.", "source_citation": "aired on Fox 31 Denver (KDVR); y-axis label set and equal spacing reconstructed and confirmed by freerangestats.info (2020-04-06), drewtyre.rbind.io, and Sara C. Doan, PhD (saracdoan.com); date range and 3,342 total printed on the chart.", "source_url": "https://saracdoan.com/index.php/misleading-covid-19-visualizations/"}
+{"chart_id": "fox31_denver_covid_unequal_yaxis_2020", "kind": "atlas", "question_type": "lie_factor", "label": "severity", "mechanism": "unequal_bins", "prompt": "Chart: New Cases Per Day (COVID-19) with an irregular unlabeled y-axis (Fox 31 Denver (KDVR-TV), Colorado, 2020)\nType: line chart of new cases per day (15 days)\nReported mechanism: unequal bins\n\nValues as published:\n  {\"label\": \"y-axis gridline labels (drawn at equal spacing, numerically irregular)\", \"value\": \"30, 60, 90, 100, 130, 160, 190, 240, 250, 300, 350, 400\"}\n  {\"label\": \"y-axis interval sequence between labels\", \"value\": \"30, 30, 10, 30, 30, 30, 50, 10, 50, 50, 50\"}\n  {\"label\": \"date range\", \"value\": \"Mar 18 to Apr 1 2020\"}\n  {\"label\": \"total cases printed on the chart\", \"value\": 3342}\n(inferred: The verified artifact is the unequal y-axis label set and interval sequence, corroborated across three independent sources. The per-day plotted series (community reverse-engineering, c. 33 to 376) is NOT read from a table and sums to 3,159, which does not match the 3,342 total printed on the chart, so those daily values are treated as approximate and excluded from the ground-truth values above. Publisher Fox 31 Denver per saracdoan; freerangestats labels it 'Fox'.)\n\nQuestion: On Tufte's Lie Factor (1.0 honest, above 1 exaggerates, below 1 under-states, negative reverses), grade this chart's distortion.", "response": "Lie Factor not a single defined factor, moderate. Unequal bin widths distort the apparent distribution. The data says: on a uniform zero-based axis the same series climbs steeply, several-fold across the two-week window; because the gap from 90 to 100 (10 cases) is drawn the same size as 350 to 400 (50 cases), the design flattens and understates how fast cases were rising.", "check": {"lie_factor": null, "severity_band": "moderate"}, "source_citation": "aired on Fox 31 Denver (KDVR); y-axis label set and equal spacing reconstructed and confirmed by freerangestats.info (2020-04-06), drewtyre.rbind.io, and Sara C. Doan, PhD (saracdoan.com); date range and 3,342 total printed on the chart.", "source_url": "https://saracdoan.com/index.php/misleading-covid-19-visualizations/"}
+{"chart_id": "kansas_kdhe_mask_mandate_dual_axis_2020", "kind": "atlas", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "dual_axis_mismatch", "prompt": "Chart: Mask-mandate vs non-mandate COVID cases per 100k, Kansas (KDHE) (Kansas Department of Health and Environment, presented by Dr. Lee Norman, 2020)\nType: dual-axis line chart\nReported mechanism: dual axis mismatch\n\nValues as published:\n  {\"label\": \"mask-mandate counties, change in cases per 100k (Jul 12 to Aug 6)\", \"value\": \"-34.4%, plotted on the left y-axis\"}\n  {\"label\": \"non-mandate counties, change over the same window\", \"value\": \"+0.7%, plotted on the right y-axis at a different scale\"}\n  {\"label\": \"hidden window Jul 3 to Jul 11 (excluded)\", \"value\": \"mandate counties rose from 9.1 to 25.3 daily cases per 100k\"}\n  {\"label\": \"mandate effective date vs chart start\", \"value\": \"mandate Jul 3; chart starts Jul 12, the local peak\"}\n(inferred: exact numeric tick ranges of the two axes not independently re-verified; the load-bearing fact, two separate y-axes on different scales for the same unit, is confirmed by mainstream local news and by Norman's own admission)\n\nQuestion: The chart's design suggests otherwise. What does the data actually show, and by what mechanism is it distorted?", "response": "On a single shared axis mask-mandate counties had higher absolute case rates than non-mandate counties throughout, and including the omitted Jul 3 to 11 days shows mandate-county cases had been rising (9.1 to 25.3 per 100k) before the cherry-picked start; KDHE Secretary Norman conceded 'when you have two different axes working at the same time, our eyes attract to that'. The mechanism is dual axis mismatch: two mismatched axes make unlike scales look comparable.", "source_citation": "dual differently-scaled y-axes and Norman's on-camera admission reported by KWCH FactFinder 12 (2020-08-11) and WIBW (2020-08); cherry-picked Jul 12 start and hidden Jul 3-11 rise from a Kansas Policy Institute open-records request via The Sentinel.", "source_url": "https://www.kwch.com/2020/08/11/factfinder-12-digs-into-controversy-concerning-kdhe-chart-on-masks/"}
+{"chart_id": "kansas_kdhe_mask_mandate_dual_axis_2020", "kind": "atlas", "question_type": "lie_factor", "label": "severity", "mechanism": "dual_axis_mismatch", "prompt": "Chart: Mask-mandate vs non-mandate COVID cases per 100k, Kansas (KDHE) (Kansas Department of Health and Environment, presented by Dr. Lee Norman, 2020)\nType: dual-axis line chart\nReported mechanism: dual axis mismatch\n\nValues as published:\n  {\"label\": \"mask-mandate counties, change in cases per 100k (Jul 12 to Aug 6)\", \"value\": \"-34.4%, plotted on the left y-axis\"}\n  {\"label\": \"non-mandate counties, change over the same window\", \"value\": \"+0.7%, plotted on the right y-axis at a different scale\"}\n  {\"label\": \"hidden window Jul 3 to Jul 11 (excluded)\", \"value\": \"mandate counties rose from 9.1 to 25.3 daily cases per 100k\"}\n  {\"label\": \"mandate effective date vs chart start\", \"value\": \"mandate Jul 3; chart starts Jul 12, the local peak\"}\n(inferred: exact numeric tick ranges of the two axes not independently re-verified; the load-bearing fact, two separate y-axes on different scales for the same unit, is confirmed by mainstream local news and by Norman's own admission)\n\nQuestion: On Tufte's Lie Factor (1.0 honest, above 1 exaggerates, below 1 under-states, negative reverses), grade this chart's distortion.", "response": "Lie Factor not a single defined factor, severe. Two mismatched axes make unlike scales look comparable. The data says: on a single shared axis mask-mandate counties had higher absolute case rates than non-mandate counties throughout, and including the omitted Jul 3 to 11 days shows mandate-county cases had been rising (9.1 to 25.3 per 100k) before the cherry-picked start; KDHE Secretary Norman conceded 'when you have two different axes working at the same time, our eyes attract to that'.", "check": {"lie_factor": null, "severity_band": "severe"}, "source_citation": "dual differently-scaled y-axes and Norman's on-camera admission reported by KWCH FactFinder 12 (2020-08-11) and WIBW (2020-08); cherry-picked Jul 12 start and hidden Jul 3-11 rise from a Kansas Policy Institute open-records request via The Sentinel.", "source_url": "https://www.kwch.com/2020/08/11/factfinder-12-digs-into-controversy-concerning-kdhe-chart-on-masks/"}
+{"chart_id": "openai_gpt5_deception_chart_2025", "kind": "atlas", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "misscaled_bars", "prompt": "Chart: GPT-5 launch 'coding deception' bar chart (the 'mega chart screwup') (OpenAI, GPT-5 launch livestream, 2025)\nType: bar chart of deception-eval rates\nReported mechanism: misscaled bars\n\nValues as published:\n  {\"label\": \"GPT-5 (with thinking), as labeled on the livestream bar\", \"value\": 50.0}\n  {\"label\": \"o3, as labeled\", \"value\": 47.4}\n  {\"label\": \"GPT-5 (with thinking), corrected value in the blog post\", \"value\": 16.5}\n(inferred: none; all three values and both quotes are documented in the cited coverage, and OpenAI itself published the correction)\n\nQuestion: The chart's design suggests otherwise. What does the data actually show, and by what mechanism is it distorted?", "response": "As printed on the livestream, 50.0 is higher than 47.4, so the labeled GPT-5 deception rate was worse than o3's, the opposite of what the bar heights showed; the corrected blog figure for GPT-5 with thinking is 16.5. Sam Altman: 'The numbers here were accurate but we screwed up the bar charts in the livestream overnight; on another slide we screwed up numbers'. The mechanism is misscaled bars: bar heights contradict the printed values.", "source_citation": "OpenAI GPT-5 launch livestream (Aug 2025) and corrected launch blog post; Altman quote via tech.co; values via Engadget/Yahoo coverage of the livestream chart and the corrected blog. OpenAI's official account called it an 'unintentional chart crime'.", "source_url": "https://tech.co/news/openai-apologizes-mega-chart-screwup"}
+{"chart_id": "openai_gpt5_deception_chart_2025", "kind": "atlas", "question_type": "lie_factor", "label": "severity", "mechanism": "misscaled_bars", "prompt": "Chart: GPT-5 launch 'coding deception' bar chart (the 'mega chart screwup') (OpenAI, GPT-5 launch livestream, 2025)\nType: bar chart of deception-eval rates\nReported mechanism: misscaled bars\n\nValues as published:\n  {\"label\": \"GPT-5 (with thinking), as labeled on the livestream bar\", \"value\": 50.0}\n  {\"label\": \"o3, as labeled\", \"value\": 47.4}\n  {\"label\": \"GPT-5 (with thinking), corrected value in the blog post\", \"value\": 16.5}\n(inferred: none; all three values and both quotes are documented in the cited coverage, and OpenAI itself published the correction)\n\nQuestion: On Tufte's Lie Factor (1.0 honest, above 1 exaggerates, below 1 under-states, negative reverses), grade this chart's distortion.", "response": "Lie Factor not a single defined factor, severe. Bar heights contradict the printed values. The data says: as printed on the livestream, 50.0 is higher than 47.4, so the labeled GPT-5 deception rate was worse than o3's, the opposite of what the bar heights showed; the corrected blog figure for GPT-5 with thinking is 16.5. Sam Altman: 'The numbers here were accurate but we screwed up the bar charts in the livestream overnight; on another slide we screwed up numbers'.", "check": {"lie_factor": null, "severity_band": "severe"}, "source_citation": "OpenAI GPT-5 launch livestream (Aug 2025) and corrected launch blog post; Altman quote via tech.co; values via Engadget/Yahoo coverage of the livestream chart and the corrected blog. OpenAI's official account called it an 'unintentional chart crime'.", "source_url": "https://tech.co/news/openai-apologizes-mega-chart-screwup"}
+{"chart_id": "llama4_maverick_lmarena_2025", "kind": "atlas", "question_type": "visual_claim_check", "label": "contradicted", "mechanism": "claim_not_supported", "prompt": "Chart: Llama-4 Maverick on the LMArena leaderboard (experimental variant, not the release) (Meta / LMArena leaderboard, 2025)\nType: leaderboard ranking chart\nReported mechanism: claim not supported\n\nValues as published:\n  {\"label\": \"Llama-4-Maverick-03-26-Experimental, ELO at launch\", \"value\": 1417}\n  {\"label\": \"publicly released Maverick, ELO (approx, per The Register)\", \"value\": 1370}\n(inferred: released-model ELO reported as approximately 1370 by The Register; the load-bearing facts, a non-public variant on the leaderboard and the released model scoring materially lower, are multiply documented)\n\nQuestion: The chart's design suggests otherwise. What does the data actually show, and by what mechanism is it distorted?", "response": "The model developers could download was not the model on the leaderboard: the public Maverick scored about 1370 versus the experimental variant's 1417, and researchers comparing outputs found the two models responded differently; LMArena stated Meta should have made the customization clearer and updated its leaderboard policies. The mechanism is claim not supported: an advertised effect the data contradicts.", "source_citation": "The Register, 2025-04-08 (Meta accused of Llama 4 bait-n-switch to juice LMArena rank), Meta-confirmed variant status; LMArena public statement and policy update; Hugging Face discussion on meta-llama/Llama-4-Maverick-17B-128E-Instruct.", "source_url": "https://www.theregister.com/2025/04/08/meta_llama4_cheating/"}
+{"chart_id": "llama4_maverick_lmarena_2025", "kind": "atlas", "question_type": "lie_factor", "label": "severity", "mechanism": "claim_not_supported", "prompt": "Chart: Llama-4 Maverick on the LMArena leaderboard (experimental variant, not the release) (Meta / LMArena leaderboard, 2025)\nType: leaderboard ranking chart\nReported mechanism: claim not supported\n\nValues as published:\n  {\"label\": \"Llama-4-Maverick-03-26-Experimental, ELO at launch\", \"value\": 1417}\n  {\"label\": \"publicly released Maverick, ELO (approx, per The Register)\", \"value\": 1370}\n(inferred: released-model ELO reported as approximately 1370 by The Register; the load-bearing facts, a non-public variant on the leaderboard and the released model scoring materially lower, are multiply documented)\n\nQuestion: On Tufte's Lie Factor (1.0 honest, above 1 exaggerates, below 1 under-states, negative reverses), grade this chart's distortion.", "response": "Lie Factor not a single defined factor, severe. An advertised effect the data contradicts. The data says: the model developers could download was not the model on the leaderboard: the public Maverick scored about 1370 versus the experimental variant's 1417, and researchers comparing outputs found the two models responded differently; LMArena stated Meta should have made the customization clearer and updated its leaderboard policies.", "check": {"lie_factor": null, "severity_band": "severe"}, "source_citation": "The Register, 2025-04-08 (Meta accused of Llama 4 bait-n-switch to juice LMArena rank), Meta-confirmed variant status; LMArena public statement and policy update; Hugging Face discussion on meta-llama/Llama-4-Maverick-17B-128E-Instruct.", "source_url": "https://www.theregister.com/2025/04/08/meta_llama4_cheating/"}

real_charts_validation.md ADDED Viewed

	@@ -0,0 +1,51 @@

+# Real-world validation: famous misleading charts, verified data
+Each chart below was published as shown. Every underlying value is verified against
+the cited primary source; no figure is reconstructed. The model's target answer
+reads the data first and names the mechanism, and the Lie Factor is computed from the
+real numbers.
+## "If the Bush Tax Cuts Expire" — top marginal income-tax rate
+- **Source:** Fox News, America's Newsroom (2012). [Reference](https://www.mediamatters.org/fox-business/dishonest-fox-chart-bush-tax-cut-edition)
+- **Data:** US federal top marginal income-tax rate: 35.0% (2012) versus the 39.6% Clinton-era rate that would return on expiration.
+- **Mechanism:** truncated y axis
+- **Lie Factor:** 4.95 (severe)
+- **Correct reading:** The top rate rises from 35% to 39.6%, an increase of 4.6 points (13%).
+- **Rows generated:** 4
+## ACA enrolment as of March 27 versus the CBO projection
+- **Source:** Fox News, America's Newsroom (2014). [Reference](https://www.mediamatters.org/fox-news/dishonest-fox-charts-obamacare-enrollment-edition)
+- **Data:** HHS: ~6.0M enrolled by 27 Mar 2014; CBO original projection 7.066M by 31 Mar 2014.
+- **Mechanism:** truncated y axis
+- **Lie Factor:** 2.55 (moderate)
+- **Correct reading:** 6.0M enrolled is 85% of the 7.066M projection, not the one third the chart implies.
+- **Rows generated:** 4
+## "The only #climatechange chart you need to see" — global temperature
+- **Source:** National Review (chart from Power Line) (2015). [Reference](https://www.washingtonpost.com/news/the-fix/wp/2015/12/14/why-the-national-reviews-global-temperature-graph-is-so-misleading/)
+- **Data:** Our World in Data / NASA GISTEMP global temperature anomaly, 1880-2015 (CC-BY).
+- **Mechanism:** expanded y axis
+- **Lie Factor:** 0.017 (severe understatement)
+- **Correct reading:** The global temperature anomaly rises about 1.1 C from 1880 to 2015, a real trend the axis flattens to near zero.
+- **Rows generated:** 2
+## Gun deaths in Florida, before and after Stand Your Ground
+- **Source:** Reuters / Business Insider (chart by C. Chan) (2014). [Reference](https://www.livescience.com/45083-misleading-gun-death-chart.html)
+- **Data:** Firearm deaths in Florida: 521 (2005), 740 (2006), 825 (2007), 721 (2012). Humphreys et al., JAMA Internal Medicine 2017; Live Science reporting.
+- **Mechanism:** inverted y axis
+- **Lie Factor:** -1.0 (severe)
+- **Correct reading:** Firearm deaths rose after the 2005 law, from 521 in 2005 to 825 in 2007, and stayed well above the pre-law level.
+- **Rows generated:** 2
+## Abortions versus cancer screenings at Planned Parenthood
+- **Source:** Americans United for Life, shown at a US House hearing (2015). [Reference](https://www.politifact.com/factchecks/2015/oct/01/jason-chaffetz/chart-shown-planned-parenthood-hearing-misleading-/)
+- **Data:** Planned Parenthood 2013 annual figures: 935,573 cancer screening and prevention services, 327,000 abortions (PolitiFact, Boston Globe, 2015).
+- **Mechanism:** dual axis mismatch
+- **Lie Factor:** 2.86 (severe)
+- **Correct reading:** In 2013 there were 935,573 cancer screening and prevention services and 327,000 abortions, so screenings outnumbered abortions about 2.9 to 1.
+- **Rows generated:** 2

v5_holdout.parquet ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:4f862f42b98613a9c30845e19e8aae8b0ec80561e0a414176638a2751f2bb9c7
+size 26089

v5_train.parquet ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:08518cba82a52316fdb270ebd27f3040869b31ac4bed1e5195acf8e4053b925e
+size 256478

v6_seed.parquet ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:dbc43f8cceb84bcbc1e8e077e1ee9b2f0214e5d2523522c169690266c946be90
+size 270115