The short version
EdGauge uses language models from outside providers for a few specific jobs: answering questions in Ask EdGauge, checking contributed figures for plausibility, matching job titles to standard positions, reading public documents, and search. No model makes a decision about a person, and every figure the assistant states cites where it came from. This page says where AI is used, which providers run it, what they receive, and how we keep the assistant's answers honest. Our privacy policy covers personal information in full, and our subprocessors page lists the providers.
Where AI is used
| Use | What it does | Provider | What is sent |
|---|---|---|---|
| Ask EdGauge, on the web and through MCP | Answers questions from the platform's data | OpenAI by default; Anthropic when our configuration selects it | Your question, up to the last 20 messages of the conversation, the results of the tools the model calls, your organization's kind and numeric id (not its name), your entitlement tier, your typed position title and the catalog position it maps to, and your own district's name and code |
| Plausibility review of contributed data | Scores each submitted value; writes a review brief when a submission is flagged | TypeSafe's Jev model through OpenRouter for the scoring; the chat provider above for the brief | For scoring: the district's type and size band (not its name), the measure, the value, its history and cohort statistics, validation results, the other values in the submission, and your notes with emails and phone numbers removed and cut to 500 characters. For a flagged submission's brief: the full context, including the district's name, the contributor's account name and the notes unredacted |
| Position title mapping | Suggests which standard position a district's printed job title matches | Jev through OpenRouter, with a fallback model (an OpenAI model, also through OpenRouter) for titles Jev is unsure of | The employer's name and county, the job title, department, number of positions and salary range, and the employer's other titles. Never a person's name |
| Document extraction, in our data pipeline | Reads public documents: salary schedules districts post, DIR public-works files, and audit report findings | OpenAI or Anthropic; findings from public audit reports go to Anthropic | Pages of a public document, a page or a section at a time, with the list of measures we are looking for. Public documents can name officials, auditors and signatories |
| Embeddings for search | Turns text into vectors so search can match meaning, not only words | OpenAI | Measure definitions, public guidance passages, standard position names, payroll job titles, and the search text the assistant writes from your question |
| Smaller uses | Pulls recommendations out of public reports (FCMAT, audit and state reports); writes the prose of a district brief | The chat provider above | Passages of public documents; a district's name, school year and peer-group label. No user data |
No request the app sends to a model carries a user id field, your email or your IP address. Your name is not sent with an assistant question. Even so, a position title with one holder, together with your district, can point to you, so we do not claim that a provider cannot tell who is asking.
The contributor's account name does reach the chat model today: in the review brief for a flagged submission, and when a tool returns a contributed value's evidence details. The evidence file's original filename goes with it, and a filename can contain a person's name. Evidence document contents are never sent to any model.
Our staff use Anthropic's Claude as a tool to operate and maintain EdGauge. What their database queries and log reads return is sent to Anthropic, and that can include account details, problem reports and IP addresses.
Training and retention
- Our requests to OpenAI's Responses API are sent with
store: false, which asks OpenAI not to keep the response on its side. - Under their commercial API terms, OpenAI and Anthropic do not train on API data by default, and may keep API data for a limited time to monitor for abuse.
- We turn off OpenRouter's prompt logging for our account. The model host behind each route OpenRouter picks has its own terms, and those terms apply to that host.
- Beyond that, each provider's own terms govern what it does with what it receives. We do not claim more than they do.
How Ask EdGauge answers are kept honest
The assistant never writes a database query. It calls typed tools, and those tools run as you, with your organization's permissions. The assistant cannot see data you cannot see.
Every figure in an answer is checked against what the tools returned before you see it. A figure must carry a citation to the observation it came from, and you can open that citation. A citation the tools did not return is removed. A figure the check cannot match gets one corrective try; if it still does not match, the answer is shown with that figure flagged and a notice.
Suppressed stays suppressed. When a peer group has fewer than the platform's small-group threshold of districts (five by default), the tools return no statistic for it, so the assistant cannot report one, not even an approximation. It says the statistic was withheld.
The assistant names only districts you named, your own, or ones a tool returned, never one from memory. It describes; it does not advise. It gives no legal, HR, labor-relations or financial advice and no recommendation of its own, though it may report what a published review recommended, attributed to that review.
Models make mistakes. The assistant can misread a question, pick the wrong measure or period, or leave something out. Check the citations before you rely on an answer.
Contributed data
No model rejects a contributed value. The AI review can send a submission to a person, or let it continue.
Every submitted value passes our validation rules and the AI plausibility review. A person on our staff reviews a submission when:
- the AI review flags it;
- it comes from an organization that is new to contributing, or that we have placed under watch;
- a validation rule blocks it;
- a note, value or file name in it reads like instructions to a model;
- the AI review was made by a model version we have not yet measured;
- it is drawn for our random audit sample.
If the AI review does not finish, nothing is cleared automatically; the submission waits for a person. A submission that passes every check, is not flagged, comes from an established organization and is not drawn for the audit sample can clear without a person looking at it.
Title mapping
A model suggests which standard position a district's printed job title matches, so that salaries can be compared by position across districts. A suggestion is accepted without review only when the model is confident, its pick is among the closest positions by meaning, and the title's salary range sits inside the range already seen for that position. Everything else goes to our staff. A random share of accepted suggestions is audited, and automatic acceptance pauses itself if the audit finds too many mistakes. The salary study shows how much of each comparison was mapped by AI. No person's name is ever sent: we never read the names in the payroll records we draw titles from.
What we do not do
- No AI decision about a person. No model scores, ranks, profiles or makes a decision about you, your staff or any official.
- No student data. EdGauge does not ask for it and has no model built to receive it. Do not type it into a question, a note or an upload; our terms say the same.
- No control over your own AI client. If you connect our tools to your own AI client through MCP, your client's model receives the results. What it does with them is between you and that provider.
Questions
Write to [email protected].