Ernest vs ChatGPT: can you paste your inspection report into a chat?

Yes, and for a one-off question that is enough. Deciding on a purchase from 80 pages is another matter: a general assistant reads part of the PDF, summarizes the rest and answers confidently, without saying where each statement comes from.

I am not a chat with a good prompt: behind the conversation there is an engine built on more than a year of work with building inspectors. I read every page, a verifier with no AI checks every quote, I classify every finding on a 160-category taxonomy and the BNQ 3009-500 grid, and I hand you a complete plan in three to six minutes, for $29.

It reads part, summarizes the rest. A general assistant receives the PDF and works from excerpts. What it has not read, it fills in with what is likely in a typical inspection report. The result reads smoothly, and nothing tells you what was skipped.

It blends severity levels. An observed defect, a mere sign to confirm, an area not inspected and a general note end up in the same list of “issues”. Your inspector keeps them carefully apart, and that distinction changes your decision.

It answers one question at a time. You get an answer, then you have to keep prompting: what about costs? negotiation? where do I start? The conversation grows, nothing is verifiable, and nobody will find it again next week.

Steps, not one prompt.

I identify the report. Before reading, I recognize the format (Québec report, InterNACHI, ASHI, Spectora, Vesta, Lumos…), the standard followed and your inspector’s own severity legend. A chat starts from scratch every time.

I read in slices, in parallel. The report is cut into slices of a few pages, read in parallel by a large model; every finding is captured with its page and its exact sentence. Nothing is summarized before it has been read.

I verify without AI. A deterministic verifier checks every quote against the text of the stated page. A finding that cannot be found is dropped before it enters the analysis. It is a mechanism, not a promise.

I classify with the inspectors’ grid. Every finding is classified on PRISM, Lumos’s building taxonomy (160 categories, 28 domains), then qualified under the BNQ 3009-500 grid: observed apparent defect, sign of deficiency to confirm, safety risk, limitation, information. Your inspector’s severity label is kept, never contradicted.

I consolidate. The summary, the detail sheet and the recommendation often describe the same defect; I merge them by domain so each problem counts once, with all its pages.

I write with domain rules and regional grids. A justified rating out of ten, three lists priced with regional cost grids in your currency, a negotiation strategy, a first-year roadmap, questions for the inspector; with the instruction never to add a fact about the house. All of it measured on a corpus of real reports, and guaranteed 100%.

Here it is.

To understand a term (“efflorescence”, “flashing”), rephrase a paragraph or check a hunch, a general assistant does the job well, and it is free. I am for something else: reading an entire report without missing anything, and handing you a plan to decide before your inspection contingency deadline.

Not in a guaranteed way. A general assistant works from excerpts of the document and fills in with what is likely, without saying what it skipped. I read every page in short slices and check every quote against the page it cites.

They know the vocabulary but do not apply it systematically: a sign to confirm and an observed defect often land in the same list. I classify every finding with that grid and repeat your inspector’s label word for word.

Yes, but not like a chat. I am a step-by-step engine that uses a model suited to each step (a large model to read and write, a fast model to identify and merge), surrounded by components that have nothing to do with a conversation: a quote verifier with no AI, the 160-category PRISM taxonomy, the BNQ 3009-500 grid, regional cost grids, and a corpus of real reports to measure results. The model providers do not train their models on your data.

Because one analysis means dozens of calls to frontier models billed per use, an engine built on more than a year of work with inspectors, and an accuracy guarantee. $29 CAD per report, taxes included, one-time payment. General assistants are free in their basic version or sold as a monthly subscription, but they read only part of the document, verify nothing and guarantee nothing.

No, and neither do ChatGPT, Gemini or Claude. I read the report; your inspector saw the house. I prepare the questions to ask them, and they have the last word.

Full comparison: ChatGPT, Gemini or Claude, reading it yourself, consultant

When you are ready, attach your PDF report.