Sign in Try a sample

The independent trust layer

Does the finding survive re-analysis?

VERITAS takes an already-stated single-cell finding — a differential-expression or cell-type-composition claim — recomputes the baseline, reruns it across every defensible analysis choice, checks experimental-design confounds, and issues a signed certificate. It verifies a result; it does not generate one.

14,274 cell-level Alzheimer’s hits became 26 under donor-level re-analysis — approximately 549× fewer. The difference between a significant result and a trustworthy one can be the experimental unit.

Murphy, Fancy & Skene, eLife 2023 · Original cohort: Mathys et al., Nature 2019

Try a sample Verify a certificate

Before a finding becomes a figure, make it survive.

“Try a sample” signs you in, then runs a bundled synthetic ground-truth case against the live engine — no data of your own needed. No expensive anonymous run is triggered.

Verification certificate

This finding survives the multiverse.

Signature valid — verdict is byte-identical to the engine's

Robustness: 15 / 16 axes
Critical axes: held
Confound sentinels: 0 flagged
Determinism: byte-identical

Illustrative certificate. Open a live one at /verify.

Bundled benchmark

Graded against known truth

Current bundled benchmark; synthetic unless labelled real.

19

Benchmark datasets / cases

18 synthetic ground-truth stress tests plus one experimentally labelled real benchmark.

18 / 18

exact synthetic-DE verdicts reproduced

Current bundled benchmark; synthetic unless labelled real.

0

false greens (a fatal finding never reads SURVIVES)

Current bundled benchmark; synthetic unless labelled real.

0

false positives on clean controls

Current bundled benchmark; synthetic unless labelled real.

100%

determinism — byte-identical certificate on re-run

Current bundled benchmark; synthetic unless labelled real.

~0.78

Doublet score AUC (real HTO benchmark)

~20% recall at the explicit expected-rate cutoff. Real labelled data — VERITAS surfaces doublet risk, it does not detect every doublet.

Explore the validation atlas →

How it works

One stated finding in. One signed verdict out.

Step 1

Recompute the baseline

The engine reconstructs the analysis behind the claim and reproduces the baseline result deterministically — same seed, same numbers, every run.

Step 2

Rerun the multiverse

The finding is re-tested across a fixed grid of real methods. It survives an axis only if it stays significant AND same-direction. Critical axes can veto a green.

Step 3

Sign the certificate

Confounds are checked with re-derivable statistics; a correction is attempted when identifiable; the outcome is signed and publicly re-checkable by id.

The multiverse

Real methods, not proxies

Every axis is a genuine re-analysis — a design-matrix batch model, real Leiden re-clustering, real Harmony integration, donor-level DESeq2 pseudobulk. When a method can't be run honestly (no donor structure), the axis skips with a caveat rather than faking a number.

Read the method →

QC threshold
Normalization
Design-matrix batch model
DE test
Benjamini–Hochberg significance
Leiden re-clustering (0.25 / 0.5 / 1.0)
Harmony integration
Donor-level DESeq2 pseudobulk

A finding survives an axis only if it stays significant and same-direction. Verdict thresholds: green ≥ 0.9, amber 0.6–0.9, red < 0.6, with critical-axis overrides.

Honest evidence

Four kinds of evidence, never blurred

VERITAS distinguishes what it is graded against. A planted synthetic answer, a real wet-lab label, an external published mechanism on a public cohort, and an unscored demonstration are different claims — and the site labels them as such.

Synthetic

Planted synthetic truth

Negative-binomial single-cell data (2000 genes) with a known answer engineered in. The verdict can be graded exactly against the planted truth.

Real · labelled

Experimental labels (real)

A real dataset carrying experimental hashtag-oligo (HTO) multiplet labels. Scored against those wet-lab labels, not a planted answer.

Real · literature

External published mechanism

A published biological claim re-checked against a public, same-experimental-design cohort standing in for a data-use-controlled dataset. Evaluates the finding's robustness, not a planted label.

Demo

Unscored demonstration

A real-workflow walkthrough with no graded outcome — shown to illustrate the pipeline end-to-end, never counted as a benchmark result.

Every verdict is one of

SurvivesAt riskDoes not survive Unverifiable

Outcomes are never conveyed by color alone — each carries its word and glyph.

See a verdict for yourself

Verify a bundled sample in the live console, or open any signed certificate by its public id — no account needed to read one.

Try a sample Verify a certificate