The Flagship
Ask About Any Medical Device. Read the Source Beside the Answer.
GearDock AI is a specialist in medical equipment, devices, and consumables that also writes. Ask it a question or request a product page. Every factual statement includes its source. Wording that makes no factual claim is labelled clearly.
Write a short product description for the Acme Medical AM-4200 infusion pump.
The AM-4200 is a volumetric infusion pump for adult and paediatric use. Manufacturer verified
It delivers between 0.1 and 999 mL/h across a 50 mL to 1,000 mL container range. Manufacturer verified
The device holds 510(k) clearance under product code FRN. Regulator verified
It is designed for wards where pumps move between beds during a shift. Wording only
2 statements withheld
Battery runtime and weight were drafted but no admitted source states them.
Grounded coverage
100%
Provenance accuracy
100%
Sources
4
The Problem
A General Model Will Answer Confidently and Be Wrong Quietly
Ask a general assistant for the specifications of a patient monitor and it will produce a clean, plausible, well-formatted answer. Some of it will be correct. Some of it will be the specifications of a different model in the same family, and nothing in the reply will tell you which parts are which.
- It fills gaps from memory. A partly documented product comes back fully specified, and the invented half looks exactly like the sourced half.
- It blurs a family into a model. A console, its software tier, and its transducers are different products with different numbers.
- It cannot show its work. When a buyer challenges a figure months later, there is nothing to go back to.
- It treats regulatory words as selling words. A 510(k) clearance is not an approval, and the difference is not decorative.
The Difference
The Source Sits Beside the Sentence
This is the whole argument. A statement on the left, the passage that supports it on the right, and a join between them that a reviewer can follow in one glance rather than by opening a document library.
The AM-4200 delivers between 0.1 and 999 mL/h and accepts containers from 50 mL to 1,000 mL.
“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.”
How It Works
What GearDock AI 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.
It works out what you are asking about
Before anything is written, the request has to resolve to a real product. GearDock recognizes 77 manufacturers and 460 product families, and it reads the identifier shapes this industry actually uses. For example, a Philips transducer is written aperture first, while a GE probe carries a connector generation suffix. When the request is genuinely ambiguous, it asks a focused question instead of guessing.
Identity resolutionIt goes and finds the sources
It searches the documents in your own workspace first, because a document your team attached and reviewed outranks anything found on the open web. Where it needs more, it goes to the manufacturer's own site before anywhere else, and holds what it finds at a lower tier than what you supplied.
Governed source discoveryIt writes only what the sources carry
Language is free; facts are not. The model may write “designed for” and “delivers” as freely as any copywriter. It may not write a number, an identifier, a manufacturer, or a regulatory, clinical, safety, compatibility, or performance statement that its sources do not contain.
Claim FirewallEvery statement is checked one at a time
Each factual atom is extracted out of the prose and judged on its own against the evidence cited for it. The result is a ledger, not a verdict: this statement, this class of claim, this source, supported or not, and why. Statements that fail are withheld rather than softened.
9 claim classesThe draft is edited before you see it
The answer is scored on 7 dimensions: grounding, structure, clarity, specificity, readability, intent match, and format match. It is sent back for revision when it falls short. You get the revised draft, not the first attempt.
Editorial QAYou can challenge it
If you disagree with a statement, say so. The claim is checked against its evidence again. You get a direct account of what supports it, what does not, and whether the original answer should stand.
Fact challenge
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.
Every fact is labelled with where it came from
Manufacturer verified, Regulator verified, Industry source, GearDock knowledge, Your workspace, You provided this, or Wording only when a sentence carries no fact.
Two separate confidence numbers, not one
Grounded coverage asks whether each statement resolves to a real admitted source. Provenance accuracy asks whether that source is labelled correctly. An answer can be fully covered and entirely mislabelled, so they are reported apart.
27 named deliverables, each with its own editorial playbook
A specification summary and a buyer guide are not the same document. The requested type selects the playbook, so a brief that resolves to no type does not get written by a generic prompt.
It refuses work outside its field
Recognition runs as a cascade. A request is refused only when every stage comes up empty and the request clearly concerns something else. The thresholds also avoid rejecting real medical supplies such as electrodes.
- 77
- Manufacturers
- 460
- Product families
- 9
- Claim classes
- 27
- Deliverables
Recognized by name, alias, or the identifier shape they use.
IntelliVue, LOGIQ, and Tegaderm are families, not models.
A description is not adjudicated the way a regulatory status is.
Each with its own editorial playbook, not one generic prompt.
Boundaries
What It Does Not Do
Stated here rather than discovered later.
- GearDock AI writes. A person decides what the catalog says and approves content before it is released or exported.
- It will tell you it could not find a reliable source rather than fill the gap. A short answer with four sourced facts is the correct output when only four facts are supported.
- It is not a system for patient data. Do not send it patient records, clinical notes, or any protected health information.
Common Questions
GearDock AI, Answered
Is GearDock AI just ChatGPT with a medical prompt?
No. The writing step is the only part that reaches a language model at all; everything that decides what is true is deterministic code. Identity resolution, source discovery, evidence tiering, claim adjudication, and provenance labelling all run before and after the model, and any statement the evidence does not carry is removed rather than rephrased.
What happens when there are no good sources for a product?
You get told. GearDock AI reports that it could not find reliable sources rather than producing a complete-looking answer built from category knowledge. You can then supply a document yourself, and it will use that as your workspace's own evidence, ranked above anything it found on the web.
Can I use its output directly on my storefront?
Not without a person approving it. Drafting, approval, release, and export are separate steps in GearDock by design, and generated content sits as a draft until somebody with the right role releases it.
Does it know about my products specifically?
It reads the documents and product records in your own workspace, which stay inside your organization's boundary. Evidence you supply is treated as the highest tier of source, above regulatory records and manufacturer material, because someone on your team attached it deliberately.
Go Deeper
Related Reading
Private Beta
Ask It About a Product You Already Sell
Private beta workspaces are set up with you rather than through open signup. Tell us what you sell and how you publish today, and we will show you GearDock AI answering on your own catalog.