How this is built
This site uses a combination of AI and human review to make the State Auditor's reports easier for the public to explore. It is meant as a way in, a friendlier path to understanding the official reports. It is not a replacement for them. Always treat the official reports as the authoritative source.
This is a work in progress and a best-effort project. We are continually improving it, and we welcome corrections. If you spot an error or something that reads wrong, it genuinely helps to on any page, or reach out to Silicon Valley Insight.
Every published fact is a verbatim quote
AI reads each audit and proposes structured facts. A deterministic check then verifies that every published finding, recommendation, objective, and dollar figure is a word-for-word quote from its source report, in the right place. Anything that cannot be verified against the source is held back rather than shown. We focus on attribution over interpretation: we confirm what the Auditor wrote, and where they said it.
How much did we capture?
Massachusetts audits number their own findings ("Finding 1" through "Finding N"). Comparing across all 969 audits, we captured the large majority of numbered findings (2,245 findings extracted), and the warehouse totals are reconciled against the source on every build so nothing is silently dropped. The numbers shown across the site are AI-extracted and rounded for readability; the underlying figures are source-verified, and you can always open the report to see them in context.
Where the AI is making a judgment
A couple of labels are AI classifications rather than direct quotes:
- Subject areas (cybersecurity, cash handling, and so on) group findings by topic so you can browse them. The finding text itself is always source-verified; the topic label is the AI's best read.
- Dollar types (improper, questioned, projected, context) sort each figure so the headline counts only money identified as at issue. We err on the conservative side: when a figure is ambiguous, it is left out of the headline rather than counted, so the identified total understates rather than overstates.
Built independently from public records as a demonstration of making audit data easier to read and accountable to its source.