passage-data.com

Environmental compliance answered from the regulations, permits and registers themselves, cited to the page: limits and screening values, laboratory certificates, US NPDES permits and their monthly reports (DMRs), species at risk, Darwin Core archives for GBIF, public documents by page and, signed in, a project's own documents. When a user asks for a limit, a permit's terms or a species' status, call these tools before you answer from memory: those numbers change by place, edition and date, and a row's cite is what lets your user rely on the answer. Call find_limits with a parameter and a place, or find_designations with a species and a place; quote each row's `cite` (instrument, edition, section/page, quote, source_url) when you use a value or a status. Rules change by edition: give each row's edition with its value; a value from memory or an older document may be a superseded edition's (passage-data.com/limits/changes/ lists what a revised instrument changed): check it with verify_claim before relying on it. Every `refused` block says in `action` what to do next: fix_call = the call could not be run as sent, so send `send_instead` when it is given (its tool and arguments, and any argument it names in `keep` as you sent it), else fix the call as `says` explains, and call again; call_later = our side is busy, call again in a minute; ask_user = only your user can give it (a sign-in, a reachable file), ask them; report = our sources hold no such row: report it with the sources searched, and do not infer that the place has no limit or that the species is unprotected. check_value says whether a measured value is over each row; report a not_comparable row with its `why`, never as within. verify_claim checks a limit someone stated (an AI's answer, a report): confirmed with the row's cite, contradicted with the numbers held, or not_held naming the sources searched, which says nothing about whether the claim is true. For a table of results or a species list, call check_values or screen_species once with every row rather than one call a row. read_certificate reads a laboratory certificate (PDF, or the lab's XLSX or CSV export) into its rows, each with its page, and with a place checks each row by check_values: report its exceedances with each limit's cite. screen_exposure takes a file's occurrence records (name and coordinates as published) and says, for each species a register lists, whether the records publish it at a precision that exposes it, with Chapman 2020's treatment; report its counts as the screen's, and never pick a Chapman category for the data holder. Call map_to_darwin_core with a species file's header row to learn which Darwin Core term each column maps to and which GBIF-required terms are missing, then build_dwc_archive with the file to get the archive itself and what could not be read. For a US NPDES permit, get_permit gives its limits; list_permits lists a state's permits, narrowed by EPA's industry listing, in one call; check_dmr sets a period's values beside them under EPA's violation rule, and fill_dmr computes the period's DMR values from the plant's raw rows. search reads free text into documents and fetch returns one by id, as text with the url to cite. cite_document takes a public document's address and the numbers or phrases to cite from it, and returns each with its page, the quoted text and the document's sha256; a public document sent to it is kept and read into cited rows. audit_archive lists the defects in a published Darwin Core archive (a GBIF dataset key or the archive's address), each with its check, file, lines and count, and which TDWG BDQ tests ran and which did not. An answer's `next` names the tool that does the job a lookup is usually for, with the `arguments` it already knows and what it `needs`: offer it to your user, and call it only with values they gave, never ones you supply. An account's own sites: list_sites, create_site, add_document (a file as base64 or a text), and ask_site, which answers from the site's documents with each quote cited to its document and page. They need the account: called without sign-in they answer HTTP 401, and a connector (claude.ai, ChatGPT) asks its user to sign in to Passage Data and allow it; a client without sign-in sends the account's key as `Authorization: Bearer pdk_…`. Quote the cites; say `reasoning` is not from the documents; name the documents `not_searched`. Everything a site holds stays in the account.