source analysis · trace · in development

Everything a source says. And what it supports.

Orves VERI reads one structured source at a time — ORIS output or your own JSON/Markdown — and captures every claim inside it: the exact span it came from, who asserted it, the qualifiers attached to it, the relationships stated within it, the places where it contradicts itself, and the claims it does not support. One source in, one verifiable result out. In development; the API contract is being finalized.

claims & evidenceexact spansqualifierswithin-source relationshipsABSTAINED

stateless · one source at a time · in development — API contract in definition

source analysis · illustrative preview
# one source in · one verifiable result out
# illustrative preview — the API contract is in definition
{ "source": "macro_outlook_may.pdf", "scope": "within_source",
  "claim": "IPCA forecast for 2026: 4.2%",
  "says": "asserted", "supports": "SUPPORTED",
  "span": "p.8 · para 2 · bbox[72,410,486,433]",
  "asserted_by": "Institution A · 2026-05-14",
  "qualifiers": ["preliminary", "subject to revision"],
  "contradicted_within_source": "claim c_31 · annex p.24",
  "abstains_when": "insufficient_evidence" }
The problem

Extraction returns text. It does not return what the text supports.

Pipelines pull statements out of a document and drop everything that makes them defensible: where exactly they came from, who asserted them, the hedge attached to them, and the passage two pages later that contradicts them. What is left reads like fact and cannot be checked.

spans discardedqualifiers flattenedauthorship lostinternal contradictions mergedguesses where evidence is missing
What you get
less hallucinationauditabilitycompliancereproducibilityexplainabilitydefensible citationsinstitutional trust
Within one source

VERI reads one source. Nothing else.

Every result is produced from a single structured source, with no access to your other files, no memory of previous runs and no prior state to consult. That constraint is the product: the same source always yields the same result, and anyone can check it against the file itself.

What VERI receives

  • one structured source — ORIS output or your own JSON/Markdown
  • the text, tables, layout and positions inside it
  • whatever that source declares about itself — author, date, scope

One source in, one verifiable result out.

What VERI never receives

  • your other sources
  • the results of previous runs
  • the current Corpus state
  • anything the source does not contain

Relating sources to each other is Orves MESH.

One source, end to end
structured sourceclaims located by spansays versus supportsqualifiers and authorshipwithin_source relationshipsone verifiable result

Nothing is inferred from outside the file — where the source does not support a claim, the state is ABSTAINED.

Says × supports

What a source says is not what it supports.

A document can assert something it never backs, hedge a number into meaninglessness, or state a figure on page 8 and a different one in the annex. Flattening all of that into a list of facts is how indefensible claims enter an audit.

one source · two passages
# macro_outlook_may.pdf · Institution A
p.8 para 2:  "IPCA forecast for 2026: 4.2%"
                 preliminary, subject to revision

annex p.24:  "2026 scenario built on IPCA of 4.4%"
                 no qualifier attached
the source-level result
# the same file disagrees with itself
├── two claims · same subject · same source
├── span: p.8 para 2 · bbox | annex p.24
├── qualifiers: preliminary | none
├── relationship: contradicts · scope: within_source
└── both preserved — neither averaged nor dropped
Relationships within a source

VERI captures these relationships within each source, every one carrying scope: within_source. Orves MESH discovers the same relationships across the Corpus, carrying scope: cross_source. Same vocabulary, two different scopes — never mixed, never silently promoted from one to the other.

supportscontradictssupersedesrefinesexplainsdepends_onderived_fromrelated_to
What it produces

One source, analysed to the bottom.

Claims with their evidence

Every claim keeps the exact span it came from — page, region, paragraph or audio timestamp — plus who asserted it and when, as declared in the source itself.

Says, and supports

What a source states and what it actually backs are different facts. Both are recorded, with the qualifiers and hedges that separate them. Insufficient evidence means ABSTAINED, never a guess.

Relationships within the source

Claims that support, refine, contradict or supersede one another inside the same file, each carrying scope: within_source. Relationships across sources belong to Orves MESH.

Benchmarks

Measured before claimed.

What we measure: claim extraction, span accuracy, qualifier fidelity, internal-contradiction detection and abstention calibration.

Public results are released only after pre-registered evaluation — until then, no number appears on this site. How we benchmark →

A property, not a product

Every artifact is verifiable.

Certification runs through everything Orves produces. Verification is designed to be public, reproducible and independent of an Orves account.

✓ provenance✓ integrityversionevidenceresolution stateverificationtrust →
API first

Integrate in minutes.

Everything is delivered through the managed API — no additional infrastructure to deploy.

REST APIPythonTypeScriptGoJavaOpenAPIExamples

VERI what your sources actually say.

Start with ORIS output or connect your own structured source — 15,000 free credits, no credit card.