DavidYor06/llm-disagreement

Lenz Frontier-LLM Disagreement — v1.1

Five frontier language models each rated the same 997 real fact-check claims submitted by users of Lenz. This dataset is the per-claim record of where they agreed and where they did not.

In 63% of real-world fact-checks, top AI models don't agree on the answer — at least one model dissents from the majority, or no majority forms at all (95% CI 60–66%).

At a glance

Claims	997 complete (of 1,000 harvested)
Models	5
claims config	997 rows — one per claim, wide, the analysis table
responses config	5,000 rows — one per claim x model, with each model's reasoning and sources
Verdict scale	True -> Mostly True -> Mixed -> Mostly False -> False
Panel reliability	Krippendorff's ordinal α = 0.77
Data as of	2026-07-18T11:09:40.272025+00:00
Snapshot	v1_1
License	Creative Commons Attribution 4.0 International

Quick start

from datasets import load_dataset

ds = load_dataset("DavidYor06/llm-disagreement", split="train")
print(ds[0]["atomic_claim"], ds[0]["max_pairwise_bucket_distance"])

That is the claims config — one row per claim, five verdict columns, the disagreement measure. For each model's actual argument and citations, load the responses config:

import json

resp = load_dataset("DavidYor06/llm-disagreement", "responses", split="train")
r = resp[0]
print(r["model"], r["verdict"], r["confidence"])
print(r["reasoning"])
print(json.loads(r["sources"]))   # sources is a JSON array in a string column

main tracks the newest snapshot, and the verdict scale has changed between versions before — so for anything you intend to reproduce, pin the revision rather than tracking main:

ds = load_dataset("DavidYor06/llm-disagreement", revision="v1.1", split="train")

Version tags are created per snapshot; v1.1 is the first. If a tag you expect is missing, the unpinned call above still resolves, and the Snapshot field in Provenance below tells you which version main currently holds.

Columns — responses config

One row per (claim, model): 5,000 rows over all 1,000 harvested claims. Join to claims on claim_id.

Column	Type	Description
claim_id	string	Join key. Same value as claims.claim_id.
claim	string	Claim text, repeated on each of the model's five rows so this config reads standalone.
date	string	As-of date given to the model.
category	string	Same vocabulary as claims.domain.
model	string	One of the five model ids below.
verdict	string	Parsed verdict, or "" when parsing failed (see error).
confidence	int	The model's own 1-10 confidence.
reasoning	string	The model's argument for its verdict, verbatim.
sources	string	A JSON array of URLs — json.loads it. Always valid JSON, [] rather than blank when nothing was cited. It is a string and not a list because datasets resolves one file format per dataset and claims is CSV (see below). Empty means "cited nothing on this answer", not "retrieved nothing ever".
fallback_used	bool	True when this answer came from the fallback model. Present on every row.
fallback_model	string	Which model actually answered when fallback_used; "" otherwise.
error	string	Non-empty when the response could not be parsed into a verdict.

Cost and latency are deliberately not published. The harvest records per-call spend and wall time; neither is here, in either config. Any cost figure in the paper is the operator's own and cannot be recomputed from this dataset.

This config covers all 1,000 harvested claims, not just the 997 in claims — the 3 excluded claims are excluded because a model's output would not parse, so their rows are the evidence for that exclusion. Restrict to the analysis cohort by joining on claims.claim_id (4,985 rows).

Columns — claims config

Column	Type	Description
claim_id	string	8-char hex id of the user-submitted claim.
claim_url	string	Claim page on Lenz (https://lenz.io/c/<claim_id>). May 404 — see Limitations.
atomic_claim	string	The exact text the models were asked to rate.
claim_created_at	string	ISO-8601 timestamp of first submission.
domain	string	One of: Finance, General, Health, History, Legal, Politics, Science, Tech.
<model>_verdict (x5)	string	Parsed verdict label on the scale below.
<model>_confidence (x5)	int	The model's own 1-10 confidence in its verdict.
max_pairwise_bucket_distance	int	0-4. Ordinal distance between the two most-disagreeing models.

Models

Column names use the model id; the paper uses the display name.

Column prefix (id)	Display name	Provider	Tier
claude-fable-5	Claude Fable 5	Anthropic	parametric
gpt-5.6-search	GPT-5.6-Sol	OpenAI	parametric
gemini-3-retrieval	Gemini 3.1 Pro + Search	Google	retrieval
sonar-deep-research	Sonar Deep Research	Perplexity	retrieval
grok-4.5-search	Grok 4.5	xAI	parametric

Verdict scale

Verdicts are ordinal. max_pairwise_bucket_distance is the index distance between the two furthest-apart models, so 0 is unanimity and 4 is one model at True while another is at False.

0. True
1. Mostly True
2. Mixed
3. Mostly False
4. False

Cohort

1,000 claims were harvested; 997 are in this file. 3 were dropped because at least one model produced output that could not be parsed into a verdict, and every pairwise measure here needs all five.

Per-model unparseable output:

Model	Cells	Unparseable
Claude Fable 5	1,000	3
GPT-5.6-Sol	1,000	0
Gemini 3.1 Pro + Search	1,000	0
Sonar Deep Research	1,000	0
Grok 4.5	1,000	0

The three break down as 1 refusal and 2 unparseable responses, and the error column in the responses config says which: 6b59c840 carries finish_reason=refusal (the claude-opus-4-8 retry refused as well), while 4af188ca and 9de4c899 carry parse_error. A refusal is a model declining the task; a parse error is output that arrived but could not be read as a verdict. Both are excluded on the same ground — no verdict — but they are not the same event.

The claude-fable-5 column contains fallback answers

A failed Claude Fable 5 call was retried once against claude-opus-4-8, and the retry's answer was kept under the original model id. 142 of the 997 rows here (14.2%) carry a claude-fable-5_verdict that was actually produced by claude-opus-4-8. No other model has any fallbacks — the whole effect is in this one column.

The wide claims table has no column for this, but the responses config does — filter it before treating any row as a genuine Fable 5 answer:

resp = load_dataset("DavidYor06/llm-disagreement", "responses", split="train")
genuine = resp.filter(lambda r: r["model"] == "claude-fable-5" and not r["fallback_used"])

Across all 1,000 harvested claims the count is 143; 142 of those fall in the 997-claim claims cohort. This matters for anything characterising Fable 5 specifically rather than "one frontier panel member".

Note this is separate from the 3 parse failures above: a fallback that returned a usable verdict is present in these 997 rows, not dropped from them.

What is deliberately not here

Lenz's own verdict on each claim. This dataset measures frontier-model disagreement, not Lenz against the frontier. A fair comparison needs human-labelled ground truth, which we are not in a position to publish. No submitter identity and no analytics signal are included either — the models saw only the claim text and the as-of date.

Limitations

• No ground truth. This measures disagreement, not accuracy. A claim all five models get wrong looks like perfect agreement here.
• One run per model. Re-running would shift the numbers; model output is not deterministic in practice even at a fixed tier.
• One column is not a single model. 14.2% of claude-fable-5_verdict values came from claude-opus-4-8 on retry. The claims config cannot tell you which; filter fallback_used in the responses config first, or treat that column as one panel seat rather than one model's behaviour.
• reasoning is a stated justification, not a causal trace. It is what the model wrote when asked to explain its verdict, which is not necessarily why it produced that verdict. Treat it as text to analyse, not as ground truth about the model's process.
• Claims are not independent. Users cluster submissions around news events and often submit several related claims in one session. Confidence intervals computed on these rows assume independence and should be read as a precision floor, not guaranteed coverage.
• Not a random sample. The corpus is what Lenz users chose to submit, skewed toward claims someone already found worth checking.
• The retrieval models were not perfectly isolated from Lenz. An exclusion list was sent to every provider that accepts one, but the two agentic models did not fully honour it: lenz.io appears among returned sources on ~25 Sonar Deep Research and ~16 Grok 4.5 calls out of 1,000, including 10 rows where the cited page was Lenz's own verdict on the exact claim being rated. In those 10 the model matched Lenz 6 times and diverged 4. Gemini's grounding tool accepts no exclusion parameter under a standard API key, so it ran without this guard and its sources are opaque redirects that cannot be audited either way. See the harness README for the full accounting.
• Equal spacing is an assumption. Treating the five buckets as evenly spaced is a simplification.
• claim_url may 404. Some pages are gone if the submitter later deleted or privatised them. The claim text in atomic_claim is the durable record.

Provenance

DOI	http://doi.org/10.5281/zenodo.21829261
Snapshot version	v1_1
Data as of	2026-07-18T11:09:40.272025+00:00
Frozen at	2026-07-22T17:32:23.270428Z
Corpus	lenzhq/lenz-research @ tag llm-disagreement-v1.1
Snapshot SHA-256	69566029… (see the manifest; supersedes af380f41…, corrected below)

The DOI resolves to the Zenodo deposit of record, which carries the paper PDF and the per-claim table. The deposit predates the text correction described under Erratum, so its copy of the per-claim table still contains the superseded claim text. For the corrected table, use this page. Cite the DOI for the paper; cite this page for the data until a new deposit version is issued.

aggregates/llm_disagreement_v1_1.json is the full frozen snapshot — every table and figure in the paper is computed from it, and the manifest hash verifies it. claims.csv is derived from that snapshot. It is no longer byte-identical to the copy served at https://lenz.io/research/llm-disagreement/v1.1/data.csv, which still serves the superseded text pending redeploy.

Erratum — claim text corrected

What was wrong. In the first published version of this dataset, atomic_claim in the claims config was not the sentence the models were shown on 181 of 997 rows. Most differences were rewordings; at least one changed a date (130f1005 read "July 12, 2026" where the rated sentence, and the as-of date given to the models, was May 18, 2026).

Why. Lenz's pipeline begins with an LLM framing step that rewrites a submission into an atomic claim. A July 2026 rerun over 410 production claims re-ran framing and overwrote the stored text; the snapshot freeze then read that live field rather than the harvest's own record of what was sent.

What changed. atomic_claim was restored from the harness corpus pinned at lenzhq/lenz-research tag llm-disagreement-v1.1 — the exact file the runner read — on 181 CSV rows and 9 embedded examples. The snapshot hash changed accordingly; the manifest records both the new and superseded values.

What did not change. No verdict, no confidence value, and no statistic. Those come from the harvest record, which framing never touched. The 63% headline, its interval, Krippendorff's α and the distance distribution are all unaffected, and max_pairwise_bucket_distance still recomputes from the verdict columns on all 997 rows. The responses config was always correct and is unchanged.

The v1.1 tag points at the corrected commit.

Citation

The paper:

@misc{lenz2026disagreement,
  author    = {Jordanov, Kosta and Yordanov, David and Jordanova, Yana},
  title     = {Beyond Benchmarks: Disagreement Among Frontier LLMs on Real-World Fact-Checks},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21829261},
  note      = {Research version v1.1},
  url       = {http://doi.org/10.5281/zenodo.21829261}
}

The dataset:

@misc{lenz2026disagreementdata,
  author    = {Lenz},
  title     = {Lenz Frontier-LLM Disagreement Benchmark (v1.1)},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/DavidYor06/llm-disagreement},
  note      = {Dataset. Archived copy of record: doi:10.5281/zenodo.21829261}
}

Links

• DOI (cite this): http://doi.org/10.5281/zenodo.21829261
• Paper and interactive figures: https://lenz.io/research/llm-disagreement
• PDF: https://lenz.io/research/llm-disagreement/v1.1/pdf
• Harvest harness and corpus: https://github.com/lenzhq/lenz-research