feat(laws): reach agency adjudications — committee decisions and 행정심판 재결 (WI-P7)

[tjdwls101010]

Why

The oversight question this package could not answer was "did the supervising body actually act?"

The statute was reachable. The court ruling was reachable. The disposition in between was not — a 개인정보보호위원회 의결, a 공정위 과징금 처분, a 조세심판원 재결. The audit doc listed all of them as optional, gated on demand being proven; a harness whose purpose is asking whether a regulator did its job is that proof.

What

Seventeen bodies, one code path. Twelve committees (ppc ftc fsc sfc kcc nhrck acr nlrc eiac iaciac oclt ecc), the general 행정심판 docket (decc), and four special tribunals (acr adap tt kmst). The request shape is identical everywhere; only the vocabulary differs, and that's normalization's job rather than seventeen near-copies of one method.

search-committee-decisions [query] --committee <code>   → committee_decision_hit_list
get-committee-decision --id <id> --committee <code>     → committee_decision_text
search-administrative-appeals [query] [--tribunal]      → administrative_appeal_hit_list
get-administrative-appeal --id <id> [--tribunal]        → administrative_appeal_text

All seventeen probed live on 2026-07-19 with the shared default OC — every one answers list and detail, no 활용신청 needed (this was the plan's blocking pre-check).

Authority is a field, not decoration

A 위원회 결정 is an administrative disposition by the regulator itself. An 행정심판 재결 reviews another agency's disposition and can itself be overturned in court. Neither is precedent.

They get separate kinds and separate authority sentences. Sharing one is how a reviewable disposition ends up cited as a settled one.

The zero-hit discipline matters most here

This family gets used to ask whether a regulator failed to act, so reading 0건 as "never disposed of anything" is wrong in exactly that direction — 비공개, 미접수, and wrong-docket all return zero. The signal says so, and points at the special tribunals, whose 소청·조세·해양안전 cases are absent from the general decc list.

Four normalization traps, all found by probing rather than reading

trap	consequence if missed
Special tribunals answer acrSpecialDecc with rows keyed decc	matching on the requested target returns nothing → reads as "no records"
Several targets return the literal string "null" for absent fields	renders as a plausible case number in an answer
Committees write dates as 2020.6.8. (unpadded, trailing dot)	compact_date passes it through; sorts wrong — and in oversight the date is the finding
Field values arrive as string / array / single-key object per body	str() yields a Python repr that travels as the agency's wording

Also

--brief now resets each field to its own declared default rather than a blanket "". text: str and full_text: str | None mean different things by absence, and blanking a nullable field to "" turns "not loaded" into "loaded and empty" — a claim about the document rather than about the request. --brief also drops reasoning, the bulk of an adjudication.

Verification

Live: ppc search+load (id 9745), ftc search+load (id 17363), decc search+load (id 273321), 조세심판원 special docket. 511 deterministic tests pass (25 new).

🤖 Generated with Claude Code