find_in_section
ChatGPTReturns only the verbatim paragraphs of one obligation (section) that contain the given terms. Inputs: obligation_id; terms, 1-6 short words or phrases ("cash-out", "credit score", "LTV"; abbreviations are expanded; numbers match as words or digits, so "five years" finds "five (5) years"); optional max_passages. Each passage has verbatim (the regulator's exact text), offset (its position in the section's text), canonical_citation (the pinpoint cite, where derivable) and heading_path (every heading above it, e.g. "vi. No Cash-Out Refinances > (A) Rate and Term > (i) Standard"). A match inside a table returns the whole table (table: true; table_layout: "flattened" when the source lost its layout and cells may not line up). A passage ending in a colon includes the list it introduces (list_included: true). truncated means more passages matched than were returned. No passages means no paragraph contains the terms as given; it is not a statement about the rule. currency is null or a version/status signal (dual-version, pending, revision-pending, superseded, expired...). On an FHA section HUD's Update 18 revises, version ("in-force" or "update-18-optional") says which text this is and other_version identifies the other.
get_obligation
ChatGPTReturns one obligation's (section's) verbatim regulator text with provenance. Inputs: obligation_id; optional offset, limit, compare_versions. verbatim holds the text between <<<BEGIN VERBATIM>>> and <<<END VERBATIM>>>. canonical_citation is the section's citation in the regulator's style (e.g. "4000.1 II.A.5.d", "B3-3.3-07", "38 U.S.C. 3709(b)"), or null. Paging: at most 24,000 chars per response; chars_offset + chars_returned is the position within chars_total; next_offset is the offset of the next page, or null when the section ends in this response; truncated means more text follows; started_mid_section means text before chars_offset is not included; paging gives page N of M. currency is null or a version/status signal (dual-version, pending, revision-pending, superseded, expired...) with a one-line headline. An FHA 4000.1 section HUD's Update 18 revises has two obligations, the in-force text and the revision (<id>::update-18); version says which this is and other_version identifies the other. compare_versions: true adds version_changes, the sentences Update 18 deletes and adds. figures is a best-effort extraction of figures in the text; an empty or incomplete list is not evidence that the text states no figure. canonical_id, when set, is the register this one is a same-text copy of.
get_playbook_kit
ChatGPTReturns a playbook's rule kit. FNMA AI lender letter: the verbatim LL-2026-04, InfoSec Supplement and ECOA/Reg B/FCRA/AVM rows with citations, paginated; next_offset is the next page's offset, or null when the kit is complete. FHA/VA/USDA QC: the corpus registers each check area is grounded in. The kit block's fingerprint and dates identify its version. Playbook tools accept only ids and reference names, no document text; the user's P&Ps and loan files stay in their own chat.
get_playbook_reference
ChatGPTReturns one stage reference document of a playbook (e.g. the FNMA playbook's DISCOVERY, RECONCILE, SYNTHESIS, PP-UPDATE, CHANGELOG or UPDATING stage). Inputs: id, name (one of the playbook's references). Playbook tools accept only ids and reference names, no document text; the user's P&Ps and loan files stay in their own chat.
get_reg
ChatGPTReturns a manifest of one register (source document): its provenance, currency (null or a version/status signal: dual-version, pending revision, superseded, expired...) and its obligations (sections), each with obligation_id, section, canonical_citation, chars_total and will_truncate (true when the section exceeds one 24,000-char page). The manifest contains no regulator text. A dual-version section lists both obligations, each with version ("in-force" or "update-18-optional") and other_version.
list_playbooks
ChatGPTLists the guided compliance playbooks this server provides: the Fannie Mae AI lender letter (LL-2026-04) P&P review (AI inventory, AI governance policy, risk assessment, P&P updates, change log) and FHA, VA and USDA loan-file QC against the verbatim handbooks. Returns each playbook's id, what it does, when it applies and what the user provides. Playbook tools accept only ids and reference names, no document text; the user's P&Ps and loan files stay in their own chat.
lookup_citation
ChatGPTResolves a regulatory citation (e.g. '4000.1 II.A.8.d', 'B3-4.3-04', 'Freddie Mac Guide 5306.1', '38 U.S.C. 3737', 'M26-7 Chapter 4 Topic 9', 'SBA Notice 5000-872764', 'SOP 50 10 8 B.Ch1.A.2', 'HB-1-3555 11.3') to a register_id. Errors: not_found (the cite parses but is not in the corpus; parsed shows the parse), ambiguous (it could mean several registers, listed in candidates; no guess is made), unparsed (not recognizable as a citation). Near matches are not suggested. A handbook section resolves to the handbook's own register; documents that cite or amend it (e.g. Mortgagee Letters) are listed in related and are not the section's text. Identical copies of a section resolve to one register and are listed in also_at. A whole-handbook name returns handbook, an outline of its parts. A section the regulator renumbered returns kind "renumbered": register_id is the section under its new number, renumbered_from the old register, and no text is given under the old number.
search_regs
ChatGPTFull-text search of US mortgage regulator text (Fannie Mae, Freddie Mac, FHA, VA, USDA, Ginnie Mae, SBA, CFPB). Inputs: query; optional family (one agency), limit (1-25), currency_kind (only sections carrying that currency signal; "any" = any signal). Common abbreviations (DTI, LTV, IRRRL, W-2, VOE, PITI, BK...) are expanded to the regulator's terms. Each result is one section: title, obligation_id, canonical_citation (the citation in the regulator's style, or null) and excerpt, a verbatim fragment near the match, not the whole rule. match is "all_terms" or "partial"; a partial result lists the query words it lacks in missing_terms. Order: rank_score, highest first = text score (title-weighted) x term coverage x title fit, with superseded, pending, incorporated or expired text x0.2, a special program's section (special_program: Home Possible, HomeReady, 203(k), HECM...) x0.5 unless the query names the program, and obligation_type "note" (source-credit notes) or "index" (a chapter heading) x0.2 unless the query asks for them. section_status "deleted" marks a deleted-section stub, ranked last. A query that is a citation returns the cited section first and sets cited; a bare handbook name ("HUD 4000.1") adds handbook, an outline of its parts. currency is null or a version/status signal whose kinds include dual-version, pending, revision-pending, superseded, superseded-in-part (a Mortgagee Letter replaced some provisions, named in headline), modified-by (a later issuance changes it, optional until its mandatory date) and expired. version "in-force" or "update-18-optional" marks the two texts of an FHA section HUD's Update 18 revises (mandatory_date for the revision); other_version is the other text's obligation_id. Empty results carry reason (empty_query, no_match, no_match_above_threshold) and may carry guidance naming shorthand in the query.
start_playbook
ChatGPTReturns one playbook's full workflow instructions, its stages, the stage reference names and the rule-kit version: the Fannie Mae AI lender letter (LL-2026-04) P&P review, or FHA, VA or USDA loan-file QC. Input: the playbook id; an unknown id returns unknown_playbook with valid_ids. Playbook tools accept only ids and reference names, no document text; the user's P&Ps and loan files stay in their own chat.
verify_quote
ChatGPTChecks whether a passage appears in a regulator's source document; accepts text from any source. Inputs: text (up to 5,000 words); optional register_id (without it, up to 3 candidate documents are located by search). outcome is one of: exact (contiguous in the source, quote marks and dashes as the source writes them; a line break counts as a space); normalized (matches after folding quote and dash typography, typography_folded; ignoring page footers or running heads, furniture_ignored; spacing where the stored source splits or joins words, spacing_ignored; or added list markers, list_markers_ignored); composed (every word is in the source, in source order, but from non-contiguous pieces, as with an overview or an ellipsis; omissions lists the short source phrases skipped between pieces); not_found (no source support; unsupported lists the unsupported spans, and on a near miss differences gives the closest source text, the source words the text lacks and the words it adds); not_checked (over 5,000 words). coverage is the share of words found. checked_against names the documents checked (document, edition, retrieval date; ids in checked_snapshot_ids); matched_against names the one matched, matched_register_id / matched_obligation_id the stored text containing the passage. matched_version is current; superseded (the text is a prior version of the rule; superseded_by is the current register); or update-18-optional (HUD's Update 18 revision of an FHA 4000.1 section, optional until mandatory_date; in_force_obligation_id is the text in force by default). obligation_currency lists currency signals on the obligation holding the text (e.g. superseded-in-part). rescinded: true (with rescinded_by) means the passage is text a later issuance removed or replaced.
