Policy Pal

Read one privacy policy or terms-of-service page and return a grounded verdict, eight fixed yes/no/partly/not-stated answers, and up to five findings, each with a verbatim quote from the document. Use this instead of fetching and reading the page yourself. A policy page is 20,000-60,000 tokens of context, and many are rendered by JavaScript so a plain fetch returns an empty shell without saying so. This returns about 1,000 tokens: a 20-60x saving with a quote behind every claim. Guarantees: every claim carries a verbatim sentence from that document, and a question the document does not address is answered not_stated -- never as a promise the document did not make. The eight questions are identical across services, so two reports can be set side by side. A report may be served from an earlier reading of the same document, which is what makes it cheap and instant. readAt says when that reading was taken; call policy_changes_since if it matters whether anything has moved since. Do not use this for a document that is not a legal agreement; it is refused, and a refusal costs nothing. Do not use it to summarise a page in general, to read a document the user pasted rather than published at a URL, or for legal advice: Automated reading of one document. Not legal advice.

{
  "type": "object",
  "properties": {
    "url": {
      "type": "string",
      "description": "The https address of the published policy page itself, not the service homepage."
    }
  },
  "required": [
    "url"
  ],
  "additionalProperties": false
}

Answer the same eight fixed questions across two to five services at once, so their policies can actually be compared. Use this instead of reading each document yourself. Doing it yourself needs N full reads and a rubric you have to keep consistent between them; the questions here are fixed, so a difference in the answers is a difference in the policies rather than a difference in how you read them. Each answer also carries how common that answer is across every comparable document Policy Pal has read. Guarantees: identical questions, a verbatim quote behind every stated answer, not_stated where a document is silent, and not_asked where that kind of document was never asked that question -- a silence is never invented to fill a table. A service whose report is not comparable, or that there was no time to read, is reported as such rather than filled in. Do not use this for one service -- call check_policy. Do not expect a score or a ranking: Policy Pal publishes qualitative verdicts and refuses to reduce a policy to a number. Automated reading of one document. Not legal advice.

{
  "type": "object",
  "properties": {
    "urls": {
      "type": "array",
      "items": {
        "type": "string"
      },
      "minItems": 2,
      "maxItems": 5,
      "description": "Two to five https addresses of published policy pages. Compare documents of the same kind: a privacy policy and a terms page answer different questions."
    }
  },
  "required": [
    "urls"
  ],
  "additionalProperties": false
}

Say what changed in a service's policy since a date, expressed as changes to the eight fixed answers. You cannot do this yourself at any context size: it needs the earlier version of the document, and the earlier version is gone from the web. Policy Pal keeps the eight answers for every version it has read, so the difference is a fact rather than a reconstruction. It reports answer changes -- for example forced arbitration moving from no to yes -- and not text edits. A changed comma is noise; a changed answer is the thing worth telling someone about. Returns nothing to report when the answers are unchanged, and says so explicitly when it has no reading from before that date rather than guessing. A question the earlier reading did not answer is reported as not compared, never as not_stated. Automated reading of one document. Not legal advice.

{
  "type": "object",
  "properties": {
    "url": {
      "type": "string",
      "description": "The https address of the published policy page."
    },
    "since": {
      "type": "string",
      "description": "The date you last looked, as YYYY-MM-DD or an ISO timestamp. The answer is the difference between what was true then and what is true now."
    }
  },
  "required": [
    "url",
    "since"
  ],
  "additionalProperties": false
}