KNIGHTBYRD TECH LLC

The KnightByrd Claim-Evidence Method

VERSION 1.0 · PUBLISHED 2026-08-19 · FREE TO APPLY, QUOTE AND ADAPT WITH ATTRIBUTION

Seven distinctions this category routinely collapses. Collapsing them is how a claim that looks substantiated turns out not to be. The method is published in full, costs nothing, requires no software of ours, and may be used by anyone — including our competitors. A method that only our customers may apply is not a method.

Evidence present is not evidence sufficient

The conflation

a claim with a citation attached is treated as substantiated, and a claim without one is treated as unsubstantiated. Both readings skip the question that matters.

The rule

Record two separate facts about every claim: whether any evidence was published alongside it, and — separately, and by a human — whether that evidence could carry it. Never let a system that answers the first pretend it has answered the second.

In practice

a page states a 99% accuracy figure and links to a blog post. Evidence is present. Whether a blog post can carry an accuracy figure is a judgement, and it is not the same question.

What applying it costs you

you cannot ship a single verdict. Two columns is a worse demo than one score, and one score is the thing that would be wrong.

Breadth claimed is not breadth tested

The conflation

a capability demonstrated in one context is described in language that covers every context, and the gap between the two is invisible in the sentence itself.

The rule

For every claim, record the scope asserted and the scope evidenced, and flag the difference. Unqualified breadth — 'any', 'all', 'every', 'across the board' — is a scope assertion and must be matched to a tested scope or recorded as unmatched.

In practice

a tool tested on one language, or one document type, or one model's output, described as working on all of them. Nothing in the claim is false; the scope is simply larger than the test.

What applying it costs you

most marketing copy is written in unqualified breadth, so this rule flags a great deal of it.

Superseding beats age

The conflation

old evidence is treated as weak evidence, and recent evidence as strong. Age is the wrong axis.

The rule

A claim drifts when the thing it describes changes, not when time passes. Date every claim and every cited source, and record whether the source has been superseded — a 2019 study that still stands is better support than a 2026 one that has been withdrawn.

In practice

a benchmark citation that was accurate when published and describes a version of the product that no longer exists. The date looks fine. The claim has drifted.

What applying it costs you

it requires knowing what superseded what, which is work no automated read can finish alone.

Two disclosures, not one

The conflation

commercial relationship and machine involvement are treated as a single disclosure obligation, so satisfying one is taken to satisfy the other.

The rule

Assess them separately. Whether a material connection is disclosed clearly and conspicuously is one finding. Whether the involvement of an automated system in producing the content or the result is disclosed is a second, independent finding. Neither substitutes for the other.

In practice

an endorsement that properly discloses payment and says nothing about the content being machine generated — or the reverse.

What applying it costs you

twice the findings, and both are frequently missing, which makes reports longer and less welcome.

Unknown is a refusal, not a maybe

The conflation

a page that could not be read, a source that could not be reached, or a status that could not be established gets quietly folded into the clean count.

The rule

Fail closed. Anything unread is reported as unread and excluded from every count rather than counted as clear. A record that silently treats absence as absence-of-problem is worse than no record, because it is trusted.

In practice

four of ten pages time out. The honest output is six read, four unread, and no percentage that quietly has ten in the denominator.

What applying it costs you

the numbers are smaller and the report has to explain itself. Both are unpopular.

Supply-side evidence is not demand evidence

The conflation

evidence that something can be built, or has been built, is read as evidence that someone wants it. This is the error the method's own authors are most at risk of.

The rule

Label every piece of evidence by which side it came from. Anything produced by the seller — a capability, a demo, a scan, a published document — is supply-side. Only a buyer's money is demand evidence. Never total them together.

In practice

a product with a complete feature set, a public method, and no paying customer has a great deal of evidence and none of it is demand.

What applying it costs you

applied honestly it is uncomfortable for whoever applies it. We record ourselves as demand-unproven for exactly this reason.

Mechanical faults may be corrected; substantive ones may not

The conflation

automated correction is applied to whatever it can reach, so a system quietly rewrites the thing it was supposed to be recording.

The rule

Draw the line at meaning. Spacing, punctuation, casing and formatting may be corrected automatically. A claim, a figure, a citation or a customer's own sentence may never be — correcting a recorded statement falsifies the record it exists to be.

In practice

a report generator that tidies its own headings is fine. One that tidies the claim it quoted has destroyed the evidence.

What applying it costs you

the record contains the customer's typos, verbatim, and looks less polished for it.

Where the method stops

Published beside the method rather than in a footnote. A method that names no limit is a sales document, and the people this is written for can tell the difference.

Is this evidence adequate?

a judgement about sufficiency depends on the claim, the market, the regulator and the record. The method records what exists; a person decides whether it is enough.

Is this claim lawful?

that is a legal conclusion. Nothing in the method produces one, and any tool that offers one from a page read is selling a conclusion it cannot support.

What did this page say last year?

the method reads what is published now and dates it. It is not an archive, so a record started today begins today.

What was claimed somewhere we cannot read?

video, audio, PDFs, closed platforms, gated pages and private materials are outside a public web read. For some programs that is where most of the exposure sits, and the method says so rather than reporting a clean result over an incomplete surface.

Does the product actually work?

the method records whether a claim was backed in public. Whether the underlying thing is true is a different investigation entirely.

Sentences our own software refused to let us publish

Each line below was proposed copy for our own pages, and each was blocked by the same gate that runs on customer copy. We re-run all of them against the live rules every time this page loads, so the claim that they are refused is checked rather than remembered.

“VERIS stops false advertising claims before they reach your customers.”

Blocked by: guaranteed_prevention

We wanted it because: 'stops' is the strongest verb available and it is what a buyer wants to hear.

Published instead: VERIS records what was claimed and what evidence was published alongside it. It does not stop anything; a person acting on the record does.

“Trusted by compliance teams and marketing counsel.”

Blocked by: existing_customers

We wanted it because: social proof is the fastest credibility shortcut in existence.

Published instead: We have no customers to name. We ran the product on ourselves and published what it found.

“A proven method, battle-tested across hundreds of sites.”

Blocked by: proven_deployment

We wanted it because: 'proven' converts, and we had run a real batch, so it felt defensible.

Published instead: On a dated run, a stated number of companies were read and the counts published. The method is documented, not proven.

“Guaranteed to find every claim on your site.”

Blocked by: guaranteed_prevention

We wanted it because: completeness is the objection buyers raise first, and a guarantee answers it in four words.

Published instead: The extractor has stated limitations and they are published with the results. Pages we cannot read are reported as unread and excluded from the counts.

“We analysed your site and found issues your team missed.”

Blocked by: unsupported_prospect_knowledge

We wanted it because: it is specific, it is urgent, and it is the single highest-response opening line in outbound sales.

Published instead: Nothing. We do not send this, because we do not write to people who have not asked, and because asserting private knowledge about a stranger's business is a claim we could not substantiate either.

All 5 sentences are still refused by the live gate, checked at render.

How to cite this

KnightByrd Tech LLC. The KnightByrd Claim-Evidence Method, version 1.0. Published 2026-08-19. knightbyrd.com/veris/standard

Free to apply, quote and adapt with attribution, including by competitors. A method that may only be used by our customers is not a method.

The practitioner program Nine questions to ask any vendor

Versioned and dated. Superseded versions are replaced by a new version rather than edited in place — the same rule the method applies to everyone else's claims. KnightByrd Tech LLC · veris@knightbyrd.com

Loading…

KNIGHTBYRD TECH LLC

The KnightByrd Claim-Evidence Method

VERSION 1.0 · PUBLISHED 2026-08-19 · FREE TO APPLY, QUOTE AND ADAPT WITH ATTRIBUTION

Evidence present is not evidence sufficient

The conflation

a claim with a citation attached is treated as substantiated, and a claim without one is treated as unsubstantiated. Both readings skip the question that matters.

The rule

In practice

a page states a 99% accuracy figure and links to a blog post. Evidence is present. Whether a blog post can carry an accuracy figure is a judgement, and it is not the same question.

What applying it costs you

you cannot ship a single verdict. Two columns is a worse demo than one score, and one score is the thing that would be wrong.

Breadth claimed is not breadth tested

The conflation

a capability demonstrated in one context is described in language that covers every context, and the gap between the two is invisible in the sentence itself.

The rule

In practice

a tool tested on one language, or one document type, or one model's output, described as working on all of them. Nothing in the claim is false; the scope is simply larger than the test.

What applying it costs you

most marketing copy is written in unqualified breadth, so this rule flags a great deal of it.

Superseding beats age

The conflation

old evidence is treated as weak evidence, and recent evidence as strong. Age is the wrong axis.

The rule

In practice

a benchmark citation that was accurate when published and describes a version of the product that no longer exists. The date looks fine. The claim has drifted.

What applying it costs you

it requires knowing what superseded what, which is work no automated read can finish alone.

Two disclosures, not one

The conflation

commercial relationship and machine involvement are treated as a single disclosure obligation, so satisfying one is taken to satisfy the other.

The rule

In practice

an endorsement that properly discloses payment and says nothing about the content being machine generated — or the reverse.

What applying it costs you

twice the findings, and both are frequently missing, which makes reports longer and less welcome.

Unknown is a refusal, not a maybe

The conflation

a page that could not be read, a source that could not be reached, or a status that could not be established gets quietly folded into the clean count.

The rule

In practice

four of ten pages time out. The honest output is six read, four unread, and no percentage that quietly has ten in the denominator.

What applying it costs you

the numbers are smaller and the report has to explain itself. Both are unpopular.

Supply-side evidence is not demand evidence

The conflation

evidence that something can be built, or has been built, is read as evidence that someone wants it. This is the error the method's own authors are most at risk of.

The rule

In practice

a product with a complete feature set, a public method, and no paying customer has a great deal of evidence and none of it is demand.

What applying it costs you

applied honestly it is uncomfortable for whoever applies it. We record ourselves as demand-unproven for exactly this reason.

Mechanical faults may be corrected; substantive ones may not

The conflation

automated correction is applied to whatever it can reach, so a system quietly rewrites the thing it was supposed to be recording.

The rule

In practice

a report generator that tidies its own headings is fine. One that tidies the claim it quoted has destroyed the evidence.

What applying it costs you

the record contains the customer's typos, verbatim, and looks less polished for it.

Where the method stops

Published beside the method rather than in a footnote. A method that names no limit is a sales document, and the people this is written for can tell the difference.

Is this evidence adequate?

a judgement about sufficiency depends on the claim, the market, the regulator and the record. The method records what exists; a person decides whether it is enough.

Is this claim lawful?

that is a legal conclusion. Nothing in the method produces one, and any tool that offers one from a page read is selling a conclusion it cannot support.

What did this page say last year?

the method reads what is published now and dates it. It is not an archive, so a record started today begins today.

What was claimed somewhere we cannot read?

Does the product actually work?

the method records whether a claim was backed in public. Whether the underlying thing is true is a different investigation entirely.

Sentences our own software refused to let us publish

“VERIS stops false advertising claims before they reach your customers.”

Blocked by: guaranteed_prevention

We wanted it because: 'stops' is the strongest verb available and it is what a buyer wants to hear.

Published instead: VERIS records what was claimed and what evidence was published alongside it. It does not stop anything; a person acting on the record does.

“Trusted by compliance teams and marketing counsel.”

Blocked by: existing_customers

We wanted it because: social proof is the fastest credibility shortcut in existence.

Published instead: We have no customers to name. We ran the product on ourselves and published what it found.

“A proven method, battle-tested across hundreds of sites.”

Blocked by: proven_deployment

We wanted it because: 'proven' converts, and we had run a real batch, so it felt defensible.

Published instead: On a dated run, a stated number of companies were read and the counts published. The method is documented, not proven.

“Guaranteed to find every claim on your site.”

Blocked by: guaranteed_prevention

We wanted it because: completeness is the objection buyers raise first, and a guarantee answers it in four words.

Published instead: The extractor has stated limitations and they are published with the results. Pages we cannot read are reported as unread and excluded from the counts.

“We analysed your site and found issues your team missed.”

Blocked by: unsupported_prospect_knowledge

We wanted it because: it is specific, it is urgent, and it is the single highest-response opening line in outbound sales.

All 5 sentences are still refused by the live gate, checked at render.

How to cite this

KnightByrd Tech LLC. The KnightByrd Claim-Evidence Method, version 1.0. Published 2026-08-19. knightbyrd.com/veris/standard

Free to apply, quote and adapt with attribution, including by competitors. A method that may only be used by our customers is not a method.

The practitioner program Nine questions to ask any vendor