Skip to content
GearDock

How the Writing Is Governed

A Draft Can Read Well and Still Be Wrong.

Good prose is not evidence of accuracy. GearDock separates each sentence from the facts inside it, checks every fact against the cited document, and shows the supporting passage next to the claim.

Attribute ledger6 attributes · AM-4200
AttributeValueEvidence
ManufacturerAcme Medical
Source-backedAM-4200 datasheet, p.1
ModelAM-4200
Source-backedAM-4200 datasheet, p.1
Flow rate range0.1-999 mL/h
Source-backedAM-4200 datasheet, p.4
Weight2.4 kg
Source-backedService manual rev. C, p.12
Battery runtimeNo source
GapNo source supports this
Power input100-240 V AC
Conflict2 sources disagree

The Problem

In a Spreadsheet, a Verified Value and a Guess Look Identical

The cost of an unsupported claim is not paid when it is written. It is paid when a buyer asks where a number came from, or when a regulator asks what a product page meant by a clinical term, and the only answer available is an email thread from eleven months ago.

  • A well-written description attached to the wrong configuration is worse than an empty field.
  • The facts are locked inside PDFs, and nothing connects the page a number is printed on to the field it ends up in.
  • Reviewers approve fluency because fluency is what they can see. The evidence is in a folder somewhere else.
  • Every downstream system holds its own copy, and none of them records why it changed.

The Seam

One Claim, One Passage, One Join

A reviewer should be able to check a statement without leaving the statement. The passage that supports it sits beside it, tagged with the tier of the source it came from.

Drafted contentUnder review

The AM-4200 delivers between 0.1 and 999 mL/h and accepts containers from 50 mL to 1,000 mL.

Claim classSpecification
AM-4200 datasheet · p.3Tier 3 · Manufacturer document

“Flow rate is adjustable from 0.1 mL/h to 999 mL/h in 0.1 mL/h increments. The pump accepts standard containers of 50 mL, 100 mL, 250 mL, 500 mL and 1,000 mL.”

Manufacturer verifiedAdmitted · fit confirmed

How It Works

What Evidence-Based Content Actually Does

Each step depends on the one before it. The label on the right names the part of the system that does the work.

  1. Sources are ranked before they are used

    Not every document carries the same authority. GearDock runs a 4 tier ladder: evidence your team reviewed and attached, then regulatory records, then the manufacturer's own material, then secondary sources. Finding a dealer listing does not make it trusted evidence.

    Source tiering
  2. A source has to be about this product

    Being a real document is not enough. A datasheet that covers a whole family, or a page that lists a sibling product beside yours, is exactly how a fact about the wrong model gets attached to the right one. Fit is assessed before the passage is admitted as evidence.

    Evidence admission
  3. Facts are separated from prose

    Each drafted sentence is broken into the facts it asserts. Those facts are classified into 9 classes: identity, specification, compatibility, regulatory status, clinical use, safety, performance, materials, and description. A description is not checked the same way as a regulatory status.

    Assertion plane
  4. Compatibility is decided separately, on purpose

    Which device a product fits is not read out of source prose. Eight consecutive independent reviews established that no reading of free text reliably separates a product's own fit from the fit of the sibling listed next to it on the same datasheet, so compatibility is settled by governed relationships instead.

    Compatibility plane
  5. The result is a ledger you can read

    Not “the description was rejected” but “the description said 99 hours and no source states it”. A reviewer can disagree with that, and their disagreement is actionable feedback rather than a complaint about the software.

    Claim ledger

Why You Can Rely on It

The Mechanisms Behind the Claim

Statements about how the system is built, not adjectives about how good it is.

The rules are versioned

A refusal recorded under one version of the Claim Firewall says the sentence failed those rules. A reviewer looking at last quarter's rejection is shown which rule set applied, not today's answer to yesterday's question.

Language is free, facts are not

An earlier design treated any word absent from the evidence as an unsupported claim. That rejected nearly every complete sentence and left customers reading raw fragments. The current check focuses on factual claims instead.

Evidence you supply outranks evidence it finds

A document someone on your team attached and bound to a product is Tier 1. Anything retrieved rather than supplied starts at Tier 4 and has to earn its way up.

Withheld beats softened

A statement the evidence does not carry is removed from the answer. It is not hedged, qualified, or rewritten into something vaguer that would survive the check.

Boundaries

What It Does Not Do

Stated here rather than discovered later.

  • The Claim Firewall checks statements against the sources cited for them. It cannot tell you that a manufacturer's own datasheet is wrong.
  • Evidence quality limits content quality. A product with no documents produces a short, honest, mostly empty record. That is the correct result.
  • This is a review instrument, not a certification. GearDock holds no FDA clearance, SOC 2 report, or HIPAA attestation, and does not present its checks as one.

Common Questions

Evidence-Based Content, Answered

What stops the AI from just inventing a specification?

Every factual atom in a drafted sentence is extracted and checked against the specific evidence cited for it before the sentence is shown to anyone. A number, identifier, manufacturer, or regulatory, clinical, safety, compatibility, or performance statement that the sources do not contain is withheld from the output entirely.

Can I see the actual passage behind a claim?

Yes. That is the point of the design. The supporting passage sits beside the claim during review rather than in a document library you would have to go and search, which is what makes reviewing faster rather than slower.

What happens when two documents disagree?

The conflict is shown rather than resolved silently. Source tier informs which is likelier to be right, but a genuine contradiction between a manufacturer datasheet and a regulatory record is a decision for a person, and GearDock surfaces it as one.

Does this slow the writing down?

It changes where the time goes. Drafting is fast and review is easier because the evidence is already attached to each claim. Teams no longer have to reconstruct what a sentence was based on months later.

Private Beta

Bring a Product Nobody Can Currently Verify

The fastest way to see this is on a record you already argue about internally. Send us the product and the documents you hold, and we will show you what is supported and what is not.

Source Ladder

Not Every Document Is Worth the Same

Where a fact came from decides how far it can be trusted, and for which kind of statement. Discovery does not become evidence by being found.

  1. Tier 1

    Verified internal evidence

    A document someone on your team reviewed and attached to this product.

  2. Tier 2

    Regulatory record

    An FDA, GUDID, or CE/MDR record for the device itself.

  3. Tier 3

    Manufacturer document

    The manufacturer's own IFU, service manual, or datasheet.

  4. Tier 4

    Secondary source

    A dealer listing or third-party page. Held, but not trusted on its own.

Claim Classes

What the Firewall Tells Apart

Each factual atom in a drafted sentence is sorted into one of these before it is judged, because a description and a regulatory status are not the same kind of statement and cannot be checked the same way.

  • Product identity
  • Specification
  • Compatibility
  • Regulatory status
  • Clinical use
  • Safety
  • Performance
  • Materials
  • Description

Drafts are then scored on 7 editorial dimensions and sent back for revision when they fall short, so what reaches a reviewer is the revised draft rather than the first attempt.