Platform — Alethic

The platform

Verify every claim in a paper, end to end.

Upload a manuscript and Alethic decomposes it into atomic claims, retrieves each cited source, and audits citation fidelity, statistical consistency, and figure-text alignment, with full evidence trails behind every verdict.

Verification report · 2024

Efficacy of CRISPR-Cas9 in Mammalian Gene Therapy

Chen, L. · Nakamura, K. · Patel, R. · Williams, D. · Nature Biotechnology · 47/47 refs resolved

Sections

Abstract6

Introduction11

Methods18

Results34

Discussion34

Contradicted9Misstated10Needs a citation7Supported26

Introduction11 claims

Cas9 cuts both DNA strands three base pairs upstream of the PAM, leaving blunt ends.

12factual

Supported

Homology-directed repair is absent in post-mitotic cells, ruling out precise correction in neurons.

341 paper unavailablefactual

Misstated

One AAV vector can carry SpCas9, its guide RNA and a strong ubiquitous promoter.

67methodological

Contradicted

SpCas9, its guide RNA and a strong ubiquitous promoter fit in a single AAV vector.

0 of 2 cited papers we could read state this

[6]

Packaging limits of AAV vectors for in vivo genome editingContradicts

“SpCas9 with a ubiquitous promoter exceeds the 4.7 kb a single AAV genome can hold.”

[7]

Delivery routes for CRISPR nucleases in vivoDoesn't cover

Compares viral and lipid delivery; says nothing about AAV packaging limits.

Verification engine

Every claim. Every citation.
Every figure. Verified.

Document-level scores hide what matters. Alethic decomposes a paper into its atomic claims, retrieves the underlying evidence, and verifies each one with a full audit trail.

The pipeline

Multi-agent · live

Three specialists, in sequence.

01 Extraction

47 claims identified

02 Verification

Resolving evidence 41 / 47

03 Report

Compiling output

~25%¹

of citations contain errors

κ = 0.18²

peer-review agreement (NIH grants)

~50%³

of psychology papers have p-value errors

32 mo⁴

median time to retraction

1Lukić et al. 2004, J Bone Joint Surg Am/2Pier et al. 2018, PNAS/3Nuijten et al. 2016, Behav Res Methods/4Steen et al. 2013, PLOS ONE

statisticalcitation

Partially Verified

Multi-head attention with 8 heads provides a clear improvement over single-head attention on the WMT 2014 English-to-German task²⁷.

What the paper measured

25.8→26.4+0.6

BLEU · §3.2.2 / Table 3

Why partial

BLEU figure of 26.4 matches Table 3 verbatim. The cited paper supports the qualitative direction of the claim, but the +0.6 BLEU gain is smaller than what “clear improvement” typically denotes in this literature.

Source

“Multi-head attention allows the model to jointly attend to information from different representation subspaces.”

Attention Is All You Need · Vaswani et al., NeurIPS 2017 · §3.2.2

01Atomic claims

We don't score papers. We verify their claims.

Alethic decomposes a paper into individual claims, classifies each by type, and verifies each against retrieved evidence. Eight verdict types, full rationale, source quotations — every step traceable.

8 verdictsPer-claim rationaleSource quotations

02Multimodal review

Figures that don't support their captions.

A vision-language pass evaluates whether each figure actually supports the claims made about it. Detects truncated axes, misleading scales, and visual–textual misalignment — with the same evidence trail as text claims.

Axis detectionCaption alignmentVerbatim quotes

Figure check · Figure 2Disputed

Editing efficiency by cell liney-axis truncated at 85

Caption claim

“Figure 2 shows dramatic gains across cell lines with CRISPR-Cas9 optimization.”

Y-axis spans 85–100; Δ HEK293T → T-cell = 5pp. The visual “dramatic gain” reflects the truncated scale, not the underlying difference.

Statistical recomputation · §4.2Disputed

Reported

Recomputed

mean

3.47

n=21 · scale 1–7

closest possible

3.43 or 3.52

Impossible

t(20)

2.84

t(20)

2.84Matches

0.010

0.0103Rounds to 0.010

03Statistical integrity

When the numbers don't add up.

Alethic recomputes test statistics from reported values, runs GRIM/SPRITE consistency checks, and flags effect-size implausibility. Surfaces claims that don't hold up to their own data.

GRIM / SPRITEp-value recomputeEffect-size sanity

Security & privacy

Built for the institutions
research lives in.

Manuscripts and data are processed transiently in isolated environments. Nothing is retained, nothing is used for model training, and on-prem deployment is available for institutions that need the perimeter inside.

manuscript.pdf

4.2 MB · uploaded 0:00

In memory

Verifying claims

0:18

Inputs cleared

Disk: 0 KB · Memory: 0 KB

0:30

No storage by default

Papers are processed transiently in ephemeral memory. Inputs and intermediates are purged within 30 minutes of report delivery.

Your data

Training set

# inference.config

retain_inputs: false

train_on_inputs: false

weight_updates: disabled

No training on user data

Manuscripts and data never enter our training corpus. Inference only — zero weight updates from your inputs, ever.

Your institution

alethic-app

vector-db

gpu-pool

egressBlocked

nothing leaves this perimeter

On-prem available

Air-gapped deployment inside your institution's perimeter. Same product, your hardware, no data leaving the network.

Start verifying today

Every claim,
verified.

Join researchers at 12+ institutions already using Alethic to catch what peer review misses.

Try Now Talk to us