Licensing Review Decisions

Licensing-review decisions

These tasks give you a role (licensing examiner / renewal unit), a set of target ids, a policy/rules table, several record tables, and an answer_template.json. Your job is to return only JSON that conforms to the template. The domains differ but the working method is the same.

0. Orient before you compute

1. Read answer_template.json first. It is the contract. Extract for every field: exact key names, allowed_values (the enum vocabulary), list ordering, required_length, how to represent "none" (usually an empty list), and any date format. Output only those keys — no prose, markdown, citations, or extra keys.
2. Enum vocabularies are per-task. Two tasks in the same domain often use different code spellings/sets (e.g. bond_cancelled vs no_active_bond; POLICE_MEMO_CONFLICTING vs police_memo_identity_note). Always map your findings onto this template's allowed_values; never carry codes over from another task.
3. Pull the target list and any review/boundary date out of the prompt.
4. Identify the domain from the prompt + template (contractor eligibility / restricted liquor package / renewal review queue) and follow the matching checklist in references/families.md.

1. Gather complete, correct data

• Use the data services named in your environment-access notes. Query filtered to the target ids (and to any predecessor/related ids you need).
• Bulk/list responses can be capped (e.g. truncated at a row limit), which silently drops records for later target ids. Do not trust a raw list dump for completeness — retrieve per target id (or otherwise confirm every target and every one of its related rows is present) before deciding anything.
• Load the policy/rules table and index it by domain + class/family. It holds the numeric thresholds and the flags you reason with, plus a prior / legacy baseline row used for "did the current standard change the outcome" comparisons.

2. Signal vs. distractor (applies to every domain)

These datasets are deliberately salted with distractors. Trust structured fields and dates; treat contradictory free text as noise.

• Dates over status text. Currency of coverage (bonds, insurance) is decided by comparing effective / expiration / cancel dates to the review/boundary date, not by a status string that may say "active" on an expired policy.
• Authoritative code fields over notes. Where a record has a coded field (e.g. an inspection finding_code, an evidence status, verified_by_agency, severity), that field decides; human-readable notes/result text is frequently inconsistent on purpose.
• Only "open" items count. Resolved/dismissed violations, closed incidents, expired/cancelled old bonds, and superseded rows are history, not findings.
• Correspondence never resolves a structured deficiency. A verified "registry corrected" letter does not clear a shortfall/suspension/violation computed from the records; it only feeds the "stale/unverified correspondence" summary. Do not let a letter override a records-based finding.
• Reused identifiers are traps. The same address string, or a "successor" license, or ids with a -LATE/late suffix, or rows under a different license number, are planted to lure a wrong match. Match on the stable identifier (exact license/application id) first.

3. Decision skeleton (contractor & liquor share this shape)

1. Look up the governing policy row for the item's class/type → thresholds & required attributes.
2. Compute each finding/deficiency from structured records + the review date.
3. Map each finding to its remedy/action code and to a risk contribution using this template's vocabulary.
4. Choose the top-level determination/posture from a fixed precedence (a hard block → deny; any open fixable issue → hold / request-follow-up; otherwise approve / issue).
5. Compute policy-impact / same-premises / basis flags by comparing against the prior baseline policy, not just the current one.
6. Build the summary as pure roll-ups of the per-item decisions (counts and id lists), and make it internally consistent (counts must equal the decisions you made). Sort every list exactly as the template dictates.

See references/families.md for the concrete per-field derivation of each domain, and references/distractors.md for the full trap catalog.

4. Before returning

• Every target id present exactly once; list lengths match required_length.
• Each list sorted/deduped per the template; empty lists (not nulls) for "none".
• Summary counts and id lists are consistent with the per-item decisions.
• JSON only — the exact keys of the template, nothing else.