Licensing Review Batch

Licensing review batch → structured JSON

You are acting as a state licensing examiner. Each task gives you a small set of target records (application ids, a location, or a license batch), points you at a shared licensing data service, and requires JSON only that conforms to the task's input/payloads/answer_template.json. There is no free-text memo.

The data service is the same across tasks; only the target ids, the review parameters, and the exact output vocabulary change. Do not guess values — read the real records and apply the rules below.

0. Orient before you compute

1. Read the prompt and extract: the target ids (and any location id), the review/as-of/boundary date, the target count (e.g. "eight applications", "queue size 10"), and the list of allowed endpoints.
2. Read input/payloads/answer_template.json completely. It is the contract: it fixes the exact top-level keys, the item shape, every enum's allowed_values, the required ordering, list lengths, and "empty list when nothing applies" rules. The code vocabularies differ from task to task even within the same family — always map your findings onto this template's allowed_values, never a memorized list.
3. Get network access from environment_access.md (base URL, the X-Task-Token header value for POST /api/sql, and the allowed endpoints). Reach the service only over the network described there.

1. Data access — read references/environment.md first

Key facts (details and gotchas in references/environment.md):

• Every GET /api/... endpoint returns JSON. POST /api/sql accepts {"query":"SELECT ..."} with header X-Task-Token: <token from environment_access.md> and returns {columns,rows,row_count,truncated,...}.
• The GET endpoints are a filtered subset; the SQL tables are the complete source. Target records are frequently absent from the GET feeds but present in SQL. Pull the authoritative data with POST /api/sql, filtering by your target ids (WHERE ... LIKE '<PREFIX>-%'). GET is fine for small reference tables (/api/policies, /api/liquor/privileges, /api/renewal/rules).
• SQL is row-limited (~200 rows) and truncates, and blocks schema introspection (sqlite_master, PRAGMA). Always filter by your target ids; never rely on bare SELECT *. Table names equal the GET path with /→_ (e.g. /api/contractor/bonds → contractor_bonds, /api/liquor/site-evidence → liquor_site_evidence).
• Distractors are everywhere. Records for other batches share your addresses and id shapes (*-DIS-*, *-TE2-*, *-TE5-*, *-OLD-*, and *-LATE rows). Only attach a record to a target when the join key says so (see each family's rules).

references/environment.md also documents scripts/env_query.sh, a helper for GET and SQL calls.

2. Pick the task family and apply its rules

Route by the target ids / endpoints, then follow the matching reference file exactly. Each reference gives the record joins, the condition→code logic, the determination/ranking logic, and how to build the summary.

All families share the same shape of reasoning:

• Select each target's own records via the correct join key (ignore distractors).
• Detect conditions from thresholds in /api/policies (or /api/renewal/rules, /api/liquor/privileges) and record fields — comparing against the review/boundary date the prompt supplies.
• Map each detected condition to the closest code in this template's allowed_values; pair deficiencies with their required action where the template asks for actions.
• Aggregate into the required summary/queue.

3. Output discipline (all families)

• Emit only the JSON object, no prose/markdown/comments/citations, and only the keys the template shows.
• Honor every ordering rule (usually ascending lexical for code lists and ids; renewal matched_violation_ids sort by date then id; queue by rank). De-duplicate code lists. Use [] when nothing applies.
• Return exactly the required number of items (e.g. all N targets, queue length N), ordered as specified (application decisions by id; queue by rank 1..N with no gaps).
• Make the summary internally consistent with the item-level decisions (counts equal the decisions; id-lists are exactly the items that qualify).
• Every enum value you emit must be a member of that field's allowed_values.

4. Self-check before returning

• JSON parses; top-level keys == template keys exactly.
• Counts/derived lists recomputed from your own items match.
• No distractor ids leaked in; no *-LATE/post-boundary or other-batch rows counted.
• Dates in the required format; lists ordered and de-duplicated.