README.md

---
license: cc-by-4.0
language:
  - en
tags:
  - benchmark
  - decision-model
  - ranking
  - ab-testing
  - calibration
size_categories:
  - 1K<n<10K
task_categories:
  - text-classification
---

# REVEALED

Do decision models know what people actually did?

18,485 headline pairs from randomized A/B tests. Every pair of arms inside
every test. The label is a measured click rate, never a human's opinion and
never a model's judgment.

## Why

Every published number on decision models comes from a vendor or a one-off with
a few dozen samples. There is no shared benchmark.

Every one of them is also supervised by a stated preference. Laya distills from
a teacher. Guardrails use human safety annotations. Rerankers use relevance
judgments. The label is always what somebody, or something, *said* was better.

Here the label is what 538 million assignments *did*.

## Leaderboard

Rows measured on different pair sets do not compare, so the pair column says
which. A judge is asked every pair twice and only counts when both orders
agree, which is why some rows carry fewer pairs than the set holds.

| System | Params | Pairs | Decidable | Undecidable | Shift |
|---|---|---|---|---|---|
| VERA | 435M | 18,485 | **0.842** | 0.626 | 0.522 |
| Gemini 3.5 flash-lite, 20 per call | frontier | 12,494 | 0.739 | 0.601 | — |
| Gemini 3.1 Pro, one per call | frontier | 1,788 | 0.795 | 0.641 | — |
| Laya typed-decisions, one per call | 421M | 1,131 | 0.527 | 0.449 | — |
| length baseline | — | 18,485 | 0.531 | 0.496 | 0.507 |
| chance | — | — | 0.500 | 0.500 | 0.500 |

Reproduce the last two rows with `python score.py --demo`.

Batching twenty pairs into one call lets a model see the other nineteen while
answering each. That is a different task from the isolated-pair protocol, and
the row says so.

Open a PR to add a row.

## Three numbers, not one

**Decidable** is the 5,385 pairs whose click-rate difference reaches
significance at 5%, 29% of the set. It asks whether a system can rank at all.

**Undecidable** is the other 13,100. The arms are statistically indistinguishable,
so the recorded winner is weak evidence. It asks what a system does when the
question has no confident answer.

Read the undecidable column carefully. 0.500 is the floor if those labels are
pure noise. A failed significance test also covers a real gap the experiment
lacked power to prove, and Gemini posts 0.641 there having never seen these
labels, so the ceiling is unknown. The column separates systems. It does not
certify one.

**Shift** is 117,118 Reddit title pairs from SNAP's resubmission set: one
image, many titles, a measured score. Same shape, different domain. It asks
whether anything survives leaving the training distribution.

## Protocol

A ranker emits a scalar per headline. The higher one wins.

A judge is asked twice, once with the winner placed first and once with it
placed second. A pair counts only when both orders name the same headline.
Models favour the first option regardless of content, so a single-order run
scores itself on a subset it chose. Gemini 3.1 Pro picked slot A 53.2% of the
time and contradicted itself on 16.3% of pairs. Flash-lite contradicted itself
on 32.4%, Laya on 47.1%.

Report the self-consistency rate alongside the accuracy.

## Data

`holdout-pairs.csv`

| Column | Meaning |
|---|---|
| `winner` | the headline with the higher measured click rate |
| `loser` | the other arm from the same test |
| `decidable` | whether the difference reaches p<0.05 |
| `*_impressions`, `*_clicks` | the evidence behind the label |

Impressions and clicks are included so you can redraw the stratum boundary if
you disagree with 5%.

Both arms ran in the same randomized test on the same article, against
comparable random samples of the same audience.

No model's scores appear in this file. Publishing one system's logits beside
the answers invites fitting to them.

## Construction

From the holdout split of [the Upworthy Research Archive](https://osf.io/jd64p/),
which assigns tests to splits at random. The three splits hold disjoint test
ids, so no arm appears on both sides of any training run.

Arms are aggregated per headline within a test and filtered to at least 500
impressions. Every within-test pair is kept; nothing is filtered by effect
size. Experiments between 2013-06-25 and 2014-01-10 are excluded, since the
maintainers disclosed a randomization failure in that window in 2024.

The first release of this dataset held 2,137 pairs and claimed the same. Those
were one pair per test, the highest-CTR arm against the lowest, which is the
widest gap a test offers. 70% of them cleared p<0.05 against 29% here. Every
number published against that set ran several points high.

## Limits

One publisher, one medium, 2013 to 2015. A system that wins here has shown it
reads Upworthy's voice. VERA's 0.522 on the shift column is what that is worth
elsewhere.

Reddit scores are not click rates and nothing randomized them. Pairs form
inside one image and one subreddit, and the repost-decay curve is residualised
out per subreddit, which leaves time of day and submitter reputation
uncontrolled. A null result there cannot separate "no transfer" from "the
confounds swamped it."

## Citation

```bibtex
@misc{khimani_2026_revealed,
  author = {Khimani, Aliasgar},
  title  = {REVEALED: do decision models know what people actually did?},
  year   = {2026},
  url    = {https://huggingface.co/datasets/NovusEdge/revealed}
}
```

Cite the archive as well. This is a slice of their data.

Matias, J., Munger, K., Le Quere, M.A., Ebersole, C. (2021), *The Upworthy
Research Archive, a time series of 32,487 experiments in U.S. media*, Nature
Scientific Data. CC BY 4.0.