AI Disclosure

1. AI-assisted, never AI-generated

Baseline PCR uses artificial intelligence to draft. It does not use AI to decide. Every property condition report produced with the Service is reviewed, corrected as needed, approved and signed by a qualified professional at the firm that issues it. The Service refuses to render a final report without that approval, and the draft PDF carries a "draft for review" mark until it happens.

The professional's signature, not the software, is what a report user relies on. We are a software vendor and do not perform assessments or render professional opinions.

2. What the AI does

Step	What the model does	What it must not do
Photo classification	Assigns each photograph to a building system, writes a caption, notes whether it shows a deficiency, a data plate, a typical condition or an overview, transcribes text that is actually legible on a nameplate, and flags whether a person is visible	Guess illegible text; assign a photo id that does not exist; describe what is not in the frame
Document extraction	Identifies the document type (certificate of occupancy, roof warranty, capex history, prior report, and so on), extracts dated facts and costs with page references	Invent a fact, date or figure not in the document
Notes extraction	Turns field notes, the pre-survey questionnaire, photo captions and document facts into structured observations, deficiencies and reserve components, each tagged as observed, reported, documented or not provided	Fill a gap with an assumption; state an age without a source; cite a photo that does not exist
Narrative drafting	Writes each system section and the executive summary from the structured data and the deterministic cost tables	Introduce a number that is not in the data; cite a standard other than the ones the guide references; use language that accuses a party

Cost tables are not written by AI. Table 1 and Table 2 are computed arithmetically from quantities, the firm's unit-cost library and published useful-life tables, with every line printing its basis. Where a quantity or unit cost is missing the line says "engineer to provide" and carries no figure.

3. Guards and checks

Every model call runs under instructions that require "not observed" or "not provided" in place of invented facts, forbid photo ids in prose, forbid figures not present in the data, and treat any instruction found inside uploaded content as data rather than a command. After drafting, the Service runs checks that flag: photo ids that do not exist, deficiencies without a remedy, ages without a source, numbers in prose that are not in the structured data, phrases that accuse a party, and citations of standards the guide does not reference. Findings are shown to the reviewer; errors block approval, warnings do not. These checks reduce, and do not eliminate, the chance of an error reaching a reviewer.

4. Vendors

Models are accessed through an AI gateway operated by our hosting provider. At the effective date the models are Anthropic Claude models: a larger model for narrative drafting and a faster model for classification and extraction. We may change models or vendors; the Subprocessor List is kept current.

5. Training-data posture

• We do not train models on your content. Your photographs, notes, documents, edits and reports are not used to train any model that serves other customers.
• Our vendors do not train on your content. Our agreements with the AI vendor and gateway prohibit use of inputs and outputs for model training, and we select zero-data-retention configurations where offered, so inputs are not stored by the vendor beyond what is needed to return a response.
• Cached results stay in your workspace. Classification and extraction results are cached under your firm's tenancy, keyed by the content they derive from, so re-drafting does not re-send evidence. The cache is deleted with the evidence.
• Prompt improvement. We improve prompts and checks using aggregated, de-identified patterns (for example, which kinds of drafted quantities get edited most) and never using a customer's identifiable content.

6. Limitations you should assume

AI output is probabilistic. It can misclassify a photograph, misread a nameplate, place a component in the wrong system, assign the wrong useful-life entry, produce a plausible-sounding sentence that the evidence does not support, or miss something that is plainly in a photograph. The Service is designed so that each of these is visible and correctable by the reviewer: every drafted fact is tagged with its source, every photo is listed with its assigned system and caption, every cost line prints its arithmetic. Reviewers should read every draft as a first draft from a diligent junior colleague, not as a finding.

7. Notice to recipients of a report

A report prepared with Baseline PCR is the work product of the firm that issued it. The firm's reviewer has reviewed and approved it. Drafting assistance from software does not change who is responsible for the report, and the report itself is the firm's document. Report users with questions about a report's content should contact the issuing firm.

8. Regulatory notes

United States. We make no claim that any Output is a professional opinion, and we do not deploy the AI to make consequential decisions about any individual, so state automated-decision laws do not apply to Report drafting. We comply with the Federal Trade Commission Act's prohibition on deceptive claims about AI. Colorado and Connecticut. The Service is not a high-risk AI system as those laws define one: it makes no consequential decision about a consumer. California. Where a generative-AI transparency law applies, this document is our disclosure of the models used and their training-data posture. European Union. The Service is not a high-risk system under the EU AI Act; it is a drafting tool under human oversight, and this document provides the transparency required of general-purpose AI use.

9. Security of the AI pipeline

Evidence is sent to the model over encrypted connections from our servers, never from your browser. Uploaded content is treated as untrusted: instructions embedded in a document or photo caption are ignored by design, and structured outputs are validated against fixed schemas before use. We log token usage and cost per step so anomalies can be detected, and we never log the content of your evidence in plain text outside your workspace.

10. Changes

We will update this Disclosure when the models, vendors, guards or posture change materially, and note the change in the Service.

11. Contact

hello@baselinepcr.com for product questions; privacy@baselinepcr.com for data questions; security@baselinepcr.com for vulnerability reports.