Product

A trust layer for single-cell findings

VERITAS sits after the analysis and before the claim. It takes an already-stated finding — a differential-expression or a cell-type-composition claim — and independently re-checks whether it holds. It never becomes a copilot and never generates a result.

Claim type 1

Differential expression

“Gene X is up/down in condition A vs B.” VERITAS re-tests significance and direction across the multiverse; a finding survives an axis only if it stays significant AND same-direction.

Claim type 2

Cell-type composition

“Cell type Y shifts in proportion between conditions.” VERITAS re-checks the shift at both cell-pooled and donor-level proportions, guarding against donor-driven artifacts.

What it does

From claim to signed verdict

Recompute the baseline

The engine reconstructs the analysis behind a stated finding and reproduces the baseline deterministically — seeds pinned, numbers byte-identical on re-run.

Rerun defensible choices

A fixed grid of real methods re-tests the finding: QC thresholds, normalization, batch model, DE test, significance correction, re-clustering, integration, and donor-level pseudobulk.

Check design confounds

Structural sentinels measure batch/condition aliasing, pseudoreplication, covariate confounding, group imbalance, and double-dipping — each backed by a re-derivable statistic.

Attempt a correction

When a finding fails and the design is still identifiable, the engine reruns the corrected analysis its own fix implies — or reports 'unverifiable' when no analysis choice can rescue it.

Sign the certificate

The outcome is signed and reproducible. The public certificate re-checks its own signature on every read; a bundle pins data, spec, and dependencies for one-command reproduction.

Adversarial challengers (opt-in)

Claude can propose attacks grounded in the dataset's real structure, drawn from a whitelist the engine can execute. The engine runs each and computes the statistic — Claude never produces a number or a verdict.

What VERITAS is not

Not a copilot — it does not suggest or generate analyses.
Not a provenance Q&A — it re-checks, it does not narrate sources.
Not a truth oracle — it never claims a finding is biologically true.
Not a guarantee — it reports survival across defensible choices, honestly.

See the method behind the verdict

Every axis is a real method and every flag is a re-derivable statistic. Read exactly how a verdict is decided.

Read the method Try a sample