AI Transparency

Extract text and visual content from uploaded documents, including charts, diagrams, tables, and images. The model "sees" the document as a human would, understanding both text and visual elements.

Evaluate document sections against quality indicators. Determines relevance, extracts supporting evidence quotes, and assigns scores based on alignment with indicator criteria.

Generate the final assessment report including indicator findings, executive summary, and actionable recommendations. Synthesizes all assessment data into coherent, useful narratives.

In our instant web demo, a visitor can speak their school's name and city instead of typing a website. This model transcribes that short audio clip to text. Audio is held in memory only for the duration of the request and is never stored.

Resolve a spoken or typed school name and location to the school's official public website, so the demo can analyze the right site. Uses a single Claude call with web search.

Score a school against quality indicators using only its public website, in a single call, with evidence quotes. This powers the ~2-minute instant demo (a preview of the full upload-based assessment).

Redact personally identifiable information (such as names, emails, and phone numbers) from crawled public web content before it is scored or shown, even though the source pages are already public.

Judge each criterion in an authorizing framework against the evidence assembled for it, and return one verdict per criterion: met, partial, or unmet. This is how CharterRead reviews a charter application, and how a renewal review is produced for frameworks that rate rather than score.

Remove personal information from a report before it leaves the platform, whether it is downloaded as a Word document or opened through a share link. An exported file cannot be recalled once it is sent, so redaction runs on every export and there is no setting that turns it off.

An organization can opt into stricter handling of its school data. The setting changes what is kept, what is de-identified, what is sent to a model that reads images, and which model processes the data. Every organization that has not opted in keeps the standard handling described in the cards above.