AI can’t tell what’s proven from what’s just repeated.

Veriscia is cloud infrastructure that traces a scientific claim back to the experiments that actually established it. It keeps what the evidence proved separate from what later papers merely restated. Research tools, pharma R&D, and frontier AI call it to ground what they claim in what was demonstrated.

Chain of custody

2023 · stated by an AI system

Microglia cause Alzheimer’s.

Unsupported
restated as settled fact

2019 · review, now the standard citation

Microglia drive neurodegeneration.

scope dropped: “may”, “in mice”

2015 · the original experiment

In mice, microglial activation may contribute to synaptic loss.

Traced to originating evidence

Nothing is dropped silently: every step records what it left behind.

The retrieval problem

Today, every answer is built from scratch.

Scientific papers were written to be read by people, one at a time, by an expert who supplies the context. The systems now reading the literature at scale, LLMs included, get unstructured prose and no machine-readable memory of what it established. So every question starts here:

01

A question arrives

with no record of what was already established.

02

Papers are ranked by citations

a measure of how much attention a paper has drawn.

03

An answer is assembled

weighing the same evidence again.

04

The reasoning is discarded

leaving the next question to start from zero.

Slow, expensive, and nothing guarantees that the next answer agrees with the last one.

The provenance problem

A claim gains authority every time it’s repeated, not every time it’s proven.

Claims rarely fail all at once. One paper reports a cautious result. The next cites it a little more confidently. A review restates it as established. By the time an AI system reaches it, a hedged “might” under narrow conditions is being reported as plain fact, and the original experiment is never read again. The text was real at every step; what eroded was the provenance.

How one claim escalates

  1. A small animal study reports a tentative result
  2. A later paper cites it as support
  3. A review describes the effect as probable
  4. An AI system states it as fact

Six ways a claim degrades as it travels

Hedged claims harden

Qualifiers like "may" and "in this model" drop away with each citation until nothing marks the original uncertainty.

Findings jump between species

A result established in one model organism is restated as though it were shown in humans.

Citations replace the evidence

Later work cites the review rather than the study, and the underlying data stops being read at all.

One dataset looks like many

Multiple papers drawing on a single cohort read as independent confirmations of the same effect.

Contradicting evidence detaches

Failed replications and null results are published but never linked back to the claim they bear on.

Conclusions outlive their foundation

The finding underneath is retracted or quietly corrected, and the work built on top carries on unchanged.

None of this is visible in the paper in front of you. Catching it means holding a claim’s whole history: where it came from, and everything since that has propped it up or undercut it. It means keeping that history current. That is what Veriscia builds.

Our solution: The Epistemic Engine

We index scientific claims, not papers.

The Epistemic Engine reads every paper for what it demonstrates, which is rarely the same as what it says. Each demonstrated result becomes an episteme, the atomic unit of evidence, and it retains what later summaries strip out: the scope the result held under, the magnitude of the effect, the conditions attached to it, and the hedging the authors chose. A claim is then judged against the full set of epistemes that bear on it, and against how well each one actually supports it. The engine is live and in production use today, with the second generation now in development.

Originating evidence

The specific epistemes a claim rests on, held apart from everything that has since cited it.

Confidence, recomputed

How well the evidence supports the claim as stated. Derived from replication history and how independent the groups behind it actually are, then recomputed as new papers land so it never goes stale.

Claim drift

How a claim’s wording and scope shifted as it moved through the literature.

Conflicts and dependencies

What supports it, contradicts it, depends on it, or supersedes it entirely.

How it works

Read once. Reused by every question after.

Every paper is resolved into epistemes and reconciled into a living graph that gets deeper with each addition. The expensive reasoning happens a single time, and every query after it reads the result.

Papers enter once

Layer 1

Document layer

Raw papers, figures and citations become machine-readable objects.

Layer 2

Epistemic Engine

What each paper demonstrates is extracted as epistemes, then reconciled into one evolving graph.

Layer 3

Reasoning layer

Client agents query the graph over API or MCP and get structured evidence back for their own reasoning to work from.

Reused by every query

Two questions about the same claim get the same evidence.

Who it’s for

For work where being wrong is expensive.

AI and research infrastructure

Veriscia runs underneath. Your agent calls the engine over MCP or API to check a claim before it repeats it. What it generates then traces to originating evidence. The evidence-integrity layer for scientific AI.

Pharma and biotech R&D

Know whether a scientific consensus reflects independent evidence, or the same result cited a hundred times. See what a target actually rests on: what has replicated independently, and what still traces back to one group.

Clinical and translational research

Check whether the evidence under a programme or guideline is solid, or narrows to a single underpowered study. Be told when new work weakens it.

See it work

Ask it something science thinks it knows.

The demo shows the same output the platform returns, with the reasoning that produced it left visible.

How solid is the evidence behind this target?
  1. Traced the claim past the reviews to 3 originating papers.
  2. Two are single-group findings with no independent replication.
  3. One has been contradicted by a later, better-powered study.
Lowacross the originating evidence

The claim is not unsupported. It is supported more thinly than the literature makes it look. Veriscia shows which part is load-bearing, and what happens to everything above it if that part gives way.

Team

Founded by scientists building the tools they needed.

Both founders spent years inside the process the engine is built to audit.

Richard Ronayne, PhD

Co-Founder

Postdoctoral neuroscientist, Queensland Brain Institute

  • Cognitive neuroscience
  • Electrical engineering
  • Machine learning

Built the scientific extraction, grounding and auditing architecture.

Jayce Rushton

Co-Founder

AI engineer. Mathematics and neuroscience, UQ

  • Machine learning
  • LLM systems
  • Mathematical modelling

Built the confidence modelling, retrieval and navigational index.

Brisbane, Australia · Incorporated 2026

Enterprise

Put an evidence layer under your research.

If your work or your model’s output depends on the literature being read properly, and on knowing when it hasn’t been, we should talk.