Legislation Retrieval Research:

1. Data-selection logic

Source of truth. golden_dataset.gds_legislation_v4: 21,790 rows, 534 publications (V3 had 21,927 rows and 549 publications; V4 drops 137 V3 rows, all other rows match V3 on the 19 shared fields).

Definitions agreed on 27 Oct. "New" means the first occurrence of a legislation citation within a publication. Every later occurrence is "update". The key is (publication, legislation LNI), ordered by addition-release date. This is first observed in partial GDS history, not verified first-ever. Relabelling the 21,790 rows gives 14,872 + 548 from original "new" and 5,973 + 397 from original "updated" (Appendix 14). 469 publication/citation-text groups span several legislation LNIs, so identity or version differences may affect labels.

Dates, not release numbers. Calendar dates decide chronology and cutoffs. Release numbers are labels only. Nine pub/release keys have conflicting dates and are retained, not resolved.

Candidate rules (from the C/D framework).

A: the frozen baseline population per slice.
C-12m / C-24m: first occurrence of the citation inside [T−12m, T] or [T−24m, T], where T is the selected checkpoint date.
D (before-citation): not cited by T, first cited in (T, T+12m], with the law's effective date before B (the earliest selected T). Strict E < B; equality fails.
Cap: a seeded cap of 20 rows per publication, applied jointly across C and D.

Slices.

S2: cutoff 2024-12-31, W = 3 months. 180 searchable pubs, 10,758 A rows, B = 2024-09-30.
S3: cutoff 2023-12-31, W = 6 months. 251 pubs, 10,278 A rows, B = 2023-06-30.
S1: cutoff 2026-03-31, 12-month freshness. Candidate onboarding list only (Appendix 25), pending Anton's approval.

Candidate-set sizes (v3.1, Appendix 13, new and new rows):

Set	S2	S3
A	7,248	7,655
C-12m (inclusive)	1,456	1,055
C-24m	2,143	2,314
D	229	273

These are frozen from V3 and not regenerated. Historical section existence and content are not verified.

3. Index-first V1 (current preferred rule)

Instead of choosing a slice and then asking what is indexed, I started from the pub/release pairs actually in the DEV index, joined to GDS on exact labels (no release substitution), and applied the C/D rules.

Inventory (3 Nov 2026 18:54 UTC): 1,201,169 chunks, 682 pubs, exactly one release per pub.
Exact join to GDS: 3,313 rows (15.2% of 21,790), 243 pubs. 242 have an unambiguous calendar T (pub 569 has conflicting dates).
V1 primary (Appendix 16): 1,611 examples, 659 laws, 162 pubs. 1,223 new / 388 updated. Targets 1,592 present / 19 absent. Top pub share 8.38%.
V1 cap20: 968 examples, 540 laws, 162 pubs; top pub share 2.07%.
Target-present, eligible: 1,592 examples, 649 statutes, 159 pubs. This is the set I locked for the lexical baseline.
Sensitivity variants (V0, V2, V3) are reported only for comparison and are never mixed into V1. D-only gives 30 rows in 1 pub (Florida pub 1297), which is not usable alone.
Overlap with the frozen slices is small: V1 candidate checkpoints overlap 9/21, and the frozen S2/S3 inventories retain 18/180 and 27/251 checkpoints.

Caveat. The snapshot is provisional until Anton names the collection to freeze. Cross-section only, since there is one release per pub.

4. Version and release overlap (the "exact publication+release" work)

Vamshi's reference list has 688 pub/release pairs. The DEV inventory has 682 pubs. Publication overlap is 682. Exact pair overlap is 616.
The 72 list-only pairs are 66 pubs whose indexed release differs plus 6 pubs absent from the index (231/2024, 779/67, 2746/2, 7431/2025, 68439/2022, 80886/51). Failed onboarding is unproven for those 6. The same 66 appear as DEV-only pairs.
A name match alone is not enough. A pub can be present at a different release.
Release alignment for the 950 at/after examples (Appendix 17): the indexed date equals the GDS new_rel date for 133, is later for 754, and is unresolved for 63. The indexed release is usually later than the GDS release, which is why the "citation is already in the indexed text" problem matters.
The collection grew from about 239K chunks and 144 pubs (21 Oct) to 1.08M chunks and 636 pubs (early Nov), then 1.2M chunks and 682 pubs.

5. Statute XML and query-text findings

XML inventory (Appendix 24, audit 6 Nov 2026): of 649 eligible statute LNIs, 132 have verified XML (227 examples, 63 pubs). 517 not found in checked local inputs (1,365 examples, 128 pubs). Not found does not mean missing. The largest gaps are Florida (72 statutes), United States (56), California (49), Indiana (47), Ohio (45) and North Carolina (32). Only one payload version exists per statute, and its legal-version suitability is unverified. Possible suppliers, not yet called: EDMS, the CiteDef resolver, GDS.
Query text (Appendix 3): the pilot's q1 used the entire XML text, including editorial annotations. 1,072/2,354 queries (45.54%) were truncated at embedding. Mean token loss was 31.03% overall and 68.15% among truncated examples. Notes contain secondary-source references and should be excluded from clean queries. The pilot text cache holds flattened .txt, not raw XML.

6. Retrieval experiments run so far

All numbers use whole-collection search (no pub/release filter) with passage 1.0, title_path 0.8, footnote 0.5 weights, candidate_k = 50, top_k = 100, unless stated. These are exploratory, not a baseline.

6.1 Pilot vs mask-pilot baseline (Appendix 2)

On the same 682 examples, R@100 fell from 0.604 (22 Oct pilot) to 0.433 (mask-pilot original arm). Hit@100 fell 0.735 → 0.576, MRR 0.289 → 0.229. The dominant cause is the collection growing (493 new pubs acting as distractors), shown by re-running the same q1 on 20 examples (R@100 0.580 → 0.433). Query text is secondary, and scoring was ruled out. Takeaway: pilot and mask-pilot baselines are not comparable. Read deltas only against the arm that produced them.

6.2 Publication-parts ablation (Appendix 4)

Frozen 422 DEV / 986 HOLDOUT, LNI-disjoint. DEV winner A2 (heading + hierarchy + dc:title) was selected before HOLDOUT. On HOLDOUT, A2 lost R@100 and MRR versus the full-text V0 for both types. A5 (V0 minus annotations and primlawhist) had recall differences whose CIs include zero, and new-type MRR fell. Whole documents versus selected parts is the open question you were asked to look at. The honest summary is that selecting parts did not beat the whole document on HOLDOUT.

6.3 Masking (Appendix 5, Appendix 6)

Exploratory, 255 statutes (682 repeated source rows, so not independent). Original vs masked R@100 was 0.434 vs 0.441, which is within noise. A 12-statute check of the "target cites the law" case is anecdotal. The M0 preflight marks all four branches NO-GO (current exact corpus/targets, pinned serving version, pub-aware scoring adapter, verified keyword-to-book-field lookup still missing).

7. Citation-location audit (950 at/after examples)

Appendix 19, DEV snapshot 3 Nov 2026. Each example counts once.

Measure	Examples	Share
Footnote citation present (F)	604	63.6%
Direct/body citation present (D)	296	31.2%
Either (union)	885	93.2%

All 281 D-only examples show the full citation in the cached passage text. These are availability figures, not retrieval accuracy. The passage-only audit (275 examples) found the body citation lists populated in 625 of 626 rows. This links directly to the 9 Nov Jcite issue: the citation path currently converts only footnote citations into Jcite keywords, so direct body citations are missed. Test collections will keep footnote-derived and body-derived keyword lists in separate fields, and Sesha's overlap approach moved hybrid recall from about 62% to about 64%.

8. Earlier V4 selection (superseded context)

final_legislation_v4/ (20 Oct): 37 examples (28 new, 9 amendments), 11 publications and checkpoints, 35 legislation LNIs, 52 section-label mappings. Live checks passed on metadata. 143 earlier-checkpoint candidates were kept separate and are ineligible for scoring. Historical execution was recorded as BLOCKED on content/version checks and a retrieval-filter adapter.

9. Known failures and limits

Collection drift: results change as the index grows. Re-run baselines only against a confirmed frozen snapshot.
517 of 649 eligible statutes have no located XML.
Cross-section only (one release per pub), so no time-series conclusions.
Historical content/version not verified for any slice.
Unknown targets must not be counted as retrieval failures (3 of 682 pilot examples had no target present; 19 V1 rows have absent targets).
Repeated source rows inflate paired counts (682 rows = 255 statutes).
Cadence figures describe observed change-bearing releases, not true publication cadence. All S2/S3 pubs have PARTIAL history.
Transmitted-file parity for the review package is not verified.
JSite non-prod connectivity blocks the amendment flow.

10. Suggested first steps for you

Read 0, then 6.2 and 7, which are most relevant to whole-document vs part-based queries.
Do not start the lexical baseline until Anton confirms the frozen collection. The settings are locked in Appendix 26, including the target-present eligible set of 1,592 examples.
For amendments, use the Oct 20 V4 mapping (9 amendment examples) and the legislation-change architecture note as your starting point, then validate against the 50 JSONs once JSite is open.
Use the legislation-structure education you and Anton discussed (US Code organisation) to interpret the XML findings.

11. Open questions I am taking to Anton (they affect your work)

Which collection and snapshot should be frozen?
Which sampling policy applies (natural frequency, cap, stratified)?
Are direct and footnote citations both ground truth?
How do we validate release when a pub has many versions and the index has one?
Who owns the XML sourcing gap?

12. Contacts and handoff

Anton: scope and acceptance criteria.
Neha: production citation workaround and the test-schema backfill.
Sesha: hybrid plus Jcite overlap and the unified RAG output.
Vamshi: 688-pair publication list.
Ganesh: publication-distribution spreadsheet (top-10 shares 17.5% case law / 31.7% legislation, historical tables).

I am happy to walk through any section live. Per Alfiya, we are not creating a KT-only story. The research spike Anton proposed covers it.

APPENDICES: full contents of the reports referenced above

Each report is included verbatim, so this one file is self-contained. Some reports mention other working files by name in their own text; those files are not needed to follow the findings. Large reports contain embedded CSV tables.

Appendix index

GDS Legislation V4: publication, checkpoint and document selection report
Baseline gap (pilot vs mask pilot) and masking/citation-location subsets
Pilot legislation XML and query-text inventory
Publication-parts ablation
Exploratory masking pilot
M0 masking preflight
Cadence groups across GDS V4
S2/S3 publication cadence
CAD review: matching validation and D diagnostics
Release summary by publication/release
Simulated publication/checkpoint selection
Publication summary (cadence per publication)
Candidate-set choice table (v3.1)
Revised new/update counts
Index-first candidate audit (V1)
Index-first V2 comparison
Release alignment, DEV jurisdiction_test
Prefilter inventory comparison
JCite-field presence audit
Passage-only citation audit
Footnote trace sample
Footnote chunk examples and worked chunks
Citation examples shared with Neha
Statute XML inventory
S1 publication/checkpoint onboarding candidates
Lexical baseline: locked settings
Solr schema decision (768-d configset)
Legislation RAG architecture
Legislation amendment architecture

===== APPENDIX 1: GDS Legislation V4: publication, checkpoint and document selection report =====

GDS Legislation V4: Publication, Checkpoint and Document Selection Report

Review date: October 20, 2026 (America/New_York). Evidence timestamp: 2026-10-20T18:15:52.508210+00:00. Authoritative source: golden_dataset.gds_legislation_v4. Solr collection: us_legal_tracker_books_gds_test_snowflake_v3.

1. Executive summary

The finalized selection contains 37 examples: 28 new legislation and 9 amendments, spanning 11 publications and 11 historical checkpoints. It identifies 35 distinct legislation document LNIs and 30 distinct expected publication-document LNIs. The target inventory contains 52 section-label mappings, corresponding to 42 example–LNI scoring pairs and 40 distinct publication/release/LNI/chapter/section combinations.

Live V4 check: PASS. Solr target metadata: PASS at all 11 checkpoints. Complete target sets remain aligned with V4. No selected example currently requires replacement. This conclusion concerns GDS relationships and indexed metadata. It does not establish that the indexed text or embedding represents the correct historical content version.

The 143 earlier-checkpoint candidates remain separate and ineligible for scoring. Selection deliverables are complete; historical retrieval execution remains BLOCKED. Required historical content/version checks and retrieval-filter integration are unresolved. No baseline retrieval evaluation or scoring function was executed during this review.

2. Anton's requirements and how each was addressed

Requirement	Status	Evidence and remaining work
Identify suitable V4 publications	COMPLETED — selection	11 selected publications; all 37 examples checked against authoritative V4. See sections 4–5 and the detailed CSV.
Select N−1 or earlier checkpoints	COMPLETED — metadata	Each selected checkpoint is the GDS prior release. All 37 have prior-release number exactly one below the later reference release and earlier checkpoint dates. Earlier alternatives remain review-only.
Identify associated legislation documents	COMPLETED — identifiers	35 distinct legislation LNIs, preserving type, citations, publication, releases and all expected targets. Legislation full-text/version availability is not independently verified.
Determine indexed historical documents and onboarding needs	PARTIAL — historical content review pending	Publication, release and expected-section metadata checked live. No metadata-driven gap was found if all checkpoint checks pass; correct historical text/version and complete publication inventory remain unverified.
Separate new/amendment datasets for baseline testing	COMPLETED — data preparation; BLOCKED — execution	28 and 9 examples stored separately. Historical adapter and content/version evidence are required before baseline execution.

The story IDs were supplied in the request. Their full tracker acceptance criteria and current states were not independently available. Recommendations in section 9 are against Anton's stated requirements, not a tracker status update.

3. V3 vs. V4 comparison

Measure	V3	V4
Records	21,927	21,790
New legislation	15,500	15,420
Updated/amendment records	6,427	6,370
Publications	549	534
Columns	19	20

The approved comparison found that all 21,790 V4 rows match V3 on all 19 shared fields. V4 removes 137 V3 rows: 132 with blank section labels and five with blank publication PCSI, comprising 80 new and 57 updated records. V4 adds statute, a coarse source identifier matching the first two JCite tokens in every row, not statute full text or a unique document key. Its 185 values must not replace document LNIs.

All 37 chosen examples retain their original queries, citations, publication/release references and complete targets. Live review checked V4 rather than silently using V3 as the authoritative source. Comparison evidence is in v3_v4_comparison.json; current database results are in selection_review_evidence.json.

4. Selected publications and historical checkpoints

Historical Release is GDS pub_rel_number; Later Release is GDS new_rel_number. They are publication release identifiers, not versions of the legislation document. The source's prior-release relationship, numeric N−1 consistency and date ordering all pass. An independent exhaustive release inventory was not reconstructed.

Publication ID	Historical Release	Later Release	Historical Checkpoint Date	New	Amendments	Distinct expected section LNIs
38	37	38	2019-02-28	7	0	2
119	39	40	2024-09-30	0	3	1
309	79	80	2018-10-31	9	0	5
340	132	133	2024-06-30	0	2	2
347	93	94	2024-06-30	2	2	4
464	43	44	2018-09-30	2	0	2
547	57	58	2017-09-30	1	0	1
565	70	71	2024-04-30	0	1	4
1542	16	17	2024-04-30	2	1	3
82302	21	22	2018-08-31	3	0	4
82813	20	21	2018-08-31	2	0	2

The checkpoints are retained in historical_checkpoints.jsonl. Corpus membership is constrained to the selected publication, exact historical release and permitted publication-date strings. The full publication corpus is not independently certified complete merely because required targets exist.

5. Legislation-to-expected-section mapping

Identifier roles: legislation_lni identifies the legislation document associated with a change. GDS lni identifies the publication/book document expected to be retrieved. section_key and chapter describe the publication location and are not the statute section in citation. A single expected LNI may contain multiple section/subsection labels; the current guard scores unique LNIs rather than independently scoring those labels.

The full 37-row mapping CSV includes the eight requested columns plus example ID, citation variants, full section mapping objects, effective/query dates and historical filters. JSON cells preserve complete arrays rather than selecting one target. The 52 mappings are distributed as 38 new-legislation mappings and 14 amendment mappings. Their LNI scoring pairs are 30 and 12 respectively.

PASS_METADATA_ONLY; CONTENT_UNVERIFIED means the complete V4 mapping and current Solr target metadata passed. It is not a historical full-text certification.

Publication ID	Historical Release	Later Release	Legislation LNI	Type	Expected Section LNIs	Historical Checkpoint Date	Validation Status
38	37	38	8V0M-1Y72-D6RV-H0VP-00000-00	new_legislation	5R2B-F7W0-R03N-33DH-00000-00	2019-02-28	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
38	37	38	8VW3-1D12-D6RV-H207-00000-00	new_legislation	5R2B-F7W0-R03N-33DH-00000-00	2019-02-28	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
38	37	38	8V65-6WN2-D6RV-H0D4-00000-00	new_legislation	5R2B-F7W0-R03N-33DH-00000-00	2019-02-28	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
38	37	38	8SMM-P9Y2-8T6X-741D-00000-00	new_legislation	5R2B-F7W0-R03N-33D6-00000-00	2019-02-28	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
38	37	38	8VC2-J3N2-8T6X-716C-00000-00	new_legislation	5R2B-F7W0-R03N-33DH-00000-00	2019-02-28	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
38	37	38	8VRD-YF52-8T6X-70MR-00000-00	new_legislation	5R2B-F7W0-R03N-33D6-00000-00	2019-02-28	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
38	37	38	8SC9-MJP2-8T6X-737M-00000-00	new_legislation	5R2B-F7W0-R03N-33DH-00000-00	2019-02-28	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
119	39	40	6GH3-WRK3-RRVN-0158-00000-00	amendment	5PCN-K5B0-R03N-11WR-00000-00	2024-09-30	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
119	39	40	6GGG-2SM3-RRRC-N3PF-00000-00	amendment	5PCN-K5B0-R03N-11WR-00000-00	2024-09-30	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
119	39	40	6GH3-VXV3-RRVM-43PV-00000-00	amendment	5PCN-K5B0-R03N-11WR-00000-00	2024-09-30	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
309	79	80	8VH8-FTK2-D6RV-H45K-00000-00	new_legislation	55S4-FM50-R03K-42VB-00000-00	2018-10-31	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
309	79	80	8RBV-HY92-D6RV-H4SS-00000-00	new_legislation	55S4-FM50-R03K-42VB-00000-00	2018-10-31	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
309	79	80	8S2Y-WJD2-D6RV-H005-00000-00	new_legislation	55S4-FM50-R03K-42VB-00000-00	2018-10-31	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
309	79	80	8V50-N842-D6RV-H190-00000-00	new_legislation	55S4-FM50-R03K-42VB-00000-00	2018-10-31	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
309	79	80	8RNN-DT92-D6RV-H25F-00000-00	new_legislation	55S4-FM50-R03K-42X0-00000-00	2018-10-31	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
309	79	80	8V62-W762-8T6X-750D-00000-00	new_legislation	55S4-FM50-R03K-42VG-00000-00; 5KGJ-2C70-R03N-007C-00000-00	2018-10-31	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
309	79	80	8RWS-8TR2-D6RV-H3T1-00000-00	new_legislation	5KGJ-2C70-R03N-007D-00000-00	2018-10-31	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
309	79	80	8S1C-TW22-D6RV-H1DM-00000-00	new_legislation	55S4-FM50-R03K-42VB-00000-00	2018-10-31	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
309	79	80	8T18-C402-D6RV-H2DJ-00000-00	new_legislation	55S4-FM50-R03K-42X0-00000-00	2018-10-31	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
340	132	133	6C0D-2143-RSG7-G085-00000-00	amendment	51R5-YM20-R03N-P0GR-00000-00	2024-06-30	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
340	132	133	6BNK-9YC3-RXB9-23RN-00000-00	amendment	51R6-0PT0-R03M-516H-00000-00	2024-06-30	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
347	93	94	6C0D-2143-RSG7-G085-00000-00	amendment	55T7-Y770-R03M-J45F-00000-00	2024-06-30	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
347	93	94	6BNK-9YC3-RXB9-23RN-00000-00	amendment	55T7-YJC0-R03J-W4FF-00000-00	2024-06-30	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
347	93	94	6BYR-7X23-RRWX-J259-00000-00	new_legislation	55TF-9FF0-R03N-1226-00000-00	2024-06-30	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
347	93	94	6BM4-62B3-RS0S-C3XD-00000-00	new_legislation	55T7-Y770-R03M-J465-00000-00	2024-06-30	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
464	43	44	8RJT-2YJ2-D6RV-H1KT-00000-00	new_legislation	53YB-YP90-R03K-C1C2-00000-00	2018-09-30	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
464	43	44	8RHF-7XS2-D6RV-H17J-00000-00	new_legislation	53YB-YP90-R03K-C1CC-00000-00	2018-09-30	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
547	57	58	8RDR-WN32-D6RV-H4CV-00000-00	new_legislation	556T-6G30-R03M-11DV-00000-00	2017-09-30	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
565	70	71	6C4M-TST3-RV9G-W319-00000-00	amendment	524X-GMN0-R03N-C40V-00000-00; 524X-GYS0-R03N-04NV-00000-00; 524X-H8F0-R03N-N48C-00000-00; 524X-HGH0-R03N-04V8-00000-00	2024-04-30	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
1542	16	17	6DWW-FX63-RS3H-0531-00000-00	amendment	54J7-0VM0-R03J-K1FG-00000-00	2024-04-30	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
1542	16	17	67R1-T243-CGX8-00VM-00000-00	new_legislation	54J7-0VM0-R03J-K1GK-00000-00	2024-04-30	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
1542	16	17	6FFW-S0K3-RYMY-50VM-00000-00	new_legislation	5R6K-DF40-R03K-H1XS-00000-00	2024-04-30	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
82302	21	22	8RH0-54H2-D6RV-H2D5-00000-00	new_legislation	55FF-DVK0-R03M-T33R-00000-00	2018-08-31	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
82302	21	22	8SKG-7TF2-8T6X-73YX-00000-00	new_legislation	5JXF-04Y0-R03N-N0WX-00000-00	2018-08-31	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
82302	21	22	8RFY-14W2-D6RV-H3K4-00000-00	new_legislation	55FF-DVK0-R03M-T364-00000-00; 55FF-DVK0-R03M-T36Y-00000-00	2018-08-31	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
82813	20	21	8VJ8-TN02-D6RV-H51J-00000-00	new_legislation	55CR-K280-R03M-71TR-00000-00	2018-08-31	PASS_METADATA_ONLY; CONTENT_UNVERIFIED
82813	20	21	8TKW-WC52-8T6X-72S6-00000-00	new_legislation	55CR-K280-R03M-71RY-00000-00	2018-08-31	PASS_METADATA_ONLY; CONTENT_UNVERIFIED

6. Solr availability and onboarding readiness

Live checks used only GET /select, q=*:*, metadata filters, rows=0 counts or explicit metadata fields (id,lni,pub_no,release_num,pubdate). No vector similarity request, ranking evaluation, text retrieval or external write was made. Stored expected chunk identities were compared with current publication/release/LNI/date metadata. A missing query response is classified as unverified, never as zero missing assets.

Publication	Historical Release	Indexed publication chunks	Indexed release chunks	Metadata status	Historical content	Onboarding readiness/action
38	37	748	748	PASS	UNVERIFIED	No metadata-driven onboarding gap identified. Historical content review remains required.
119	39	696	696	PASS	UNVERIFIED	No metadata-driven onboarding gap identified. Historical content review remains required.
309	79	4668	4668	PASS	UNVERIFIED	No metadata-driven onboarding gap identified. Historical content review remains required.
340	132	2462	2462	PASS	UNVERIFIED	No metadata-driven onboarding gap identified. Historical content review remains required.
347	93	2278	2278	PASS	UNVERIFIED	No metadata-driven onboarding gap identified. Historical content review remains required.
464	43	409	409	PASS	UNVERIFIED	No metadata-driven onboarding gap identified. Historical content review remains required.
547	57	5954	5954	PASS	UNVERIFIED	No metadata-driven onboarding gap identified. Historical content review remains required.
565	70	3626	3626	PASS	UNVERIFIED	No metadata-driven onboarding gap identified. Historical content review remains required.
1542	16	1692	1692	PASS	UNVERIFIED	No metadata-driven onboarding gap identified. Historical content review remains required.
82302	21	454	454	PASS	UNVERIFIED	No metadata-driven onboarding gap identified. Historical content review remains required.
82813	20	447	447	PASS	UNVERIFIED	No metadata-driven onboarding gap identified. Historical content review remains required.

For the selected targets, passing metadata checks identify no mandatory onboarding solely because a publication, release, section LNI or saved expected chunk is absent. That does not establish that no onboarding will be needed. A content owner must establish that each indexed chunk was generated from the correct historical source and that the section label and text correspond to that version. If the metadata points to refreshed text, omitted historical source material, or incompatible embeddings, version-specific onboarding or reindexing may be required.

The 35 legislation-document identifiers are verified in GDS, not independently downloaded or certified as available historical versions. The publication Solr collection is not assumed to be a legislation full-text repository. Any later change-summary/full-text query workflow requires separate source-document availability and version verification.

7. Dataset validation results

Check	Result
Dataset file checksums	PASS; existing manifest entries match
Required example and target fields	PASS; zero missing required values
Example IDs and exact duplicates	PASS; 37 unique example IDs; no identical duplicated example records
Within-example complete mapping tuples	PASS; no identical repeated LNI/chapter/normalized-section tuples
Complete V4 expected target sets	PASS; current authoritative groups compared with all retained expected targets
Historical release/date consistency	PASS; all 37 N−1 and chronology checks pass
Metadata guard validation	PASS; 111 wrong-publication, wrong-release or later-release test inputs rejected; no scoring invoked
Selected ambiguous mappings	None; known V4 issues remain excluded
Earlier candidates isolated	PASS; 143 separate review-only candidates, all `scoring_eligible=false`
Historical content / legislation-version availability	UNVERIFIED; manual/source-provenance review required
Baseline retrieval execution	NOT RUN

The 52 section-label entries are not 52 distinct target documents: there are 42 example–LNI pairs and 30 distinct target LNIs. Ten extra label entries share an LNI within their example. Five contextual target mappings recur across different examples, giving 40 distinct contextual mappings; shared relevance targets are not duplicate examples. Related documents must stay together via leakage_group_id when any train/test split is later created.

The availability rule is query_as_of_date = max(pub_rel_date, effective_date), strictly before new_rel_date. For 22 examples the effective date is after the historical checkpoint; for 15 it is on or before it. This ensures ordering under a proxy, but does not prove the actual first-publication, enactment, amendment-text or exact legislation-version availability date. Citation queries were selected retrospectively from GDS. That remaining assumption must be accepted or replaced with dated source evidence.

Future-release corpus leakage is prohibited by all three stored retrieval filters and the post-retrieval metadata guard. The guard rejects rather than silently removes invalid hits. These safeguards must be integrated before candidate generation/ranking. Filtering only after a globally ranked search can distort top-k results. Existing code still lacks that integration. label_provenance contains later-release metadata for retrospective labels only and must not be passed into the query/model input.

8. Exclusions and unresolved issues

V4 still has 13 ambiguous rows involving publication 4757, chapter 04757-ch0021, release 33, labels § 21.10 and § 21.11. It also has four invalid publication-document LNIs. These 17 rows are documented in mapping_issues.jsonl and are absent from the selection.
Publication 1210 has three PCSI values and remains excluded from the clean selection. See publication_identity_issues.json.
143 earlier-checkpoint candidates (62 new, 81 amendments) are stored separately under review_only/. They involve 14 checkpoints, 129 distinct legislation LNIs and 111 distinct expected-section LNIs. Metadata establishes their candidate status only; historical section text/relevance, availability and cross-checkpoint duplication still need review. They are not part of the 37-example CSV.
Historical full text, source-to-index lineage, section-label correctness and historical embedding versions remain unverified. Solr metadata and stable document identifiers cannot prove content immutability.
The current solr_multi_vector_search() / _knn_single_field() path in solr_client.py has no historical filter argument, issues unfiltered KNN requests, and omits pubdate from returned fields. historical_eval_guard.py is separate and is not automatically integrated into that retrieval path. No retrieval code was changed.
Metadata inspection is timestamped, not an immutable server-side content snapshot. Recheck corpus/target metadata if the index changes. scoring_eligible=true in final examples denotes metadata-policy eligibility; retrieval_execution_ready=false remains authoritative for execution readiness.

9. Readiness assessment for baseline retrieval testing

Deliverable or decision	Assessment
Publication/checkpoint/document selection	COMPLETED at GDS relationship and Solr metadata level
Separate finalized new/amendment files	COMPLETED; 28 and 9 examples, with complete target sets
Missing selected Solr assets requiring onboarding	None identified by passing metadata checks; historical version/content suitability still UNVERIFIED
Full satisfaction of historical-content availability requirement	PENDING MANUAL REVIEW; not established by indexed metadata
Story 3187140 closure	NEEDS REVIEW — do not close unconditionally. Ready for acceptance review of the selection deliverables; closure requires content/onboarding evidence or explicit owner agreement that these are tracked follow-ups outside the story scope.
Story 3185687 baseline execution	BLOCKED — do not start retrieval evaluation until historical filter/guard integration and the content/availability acceptance gate are resolved. Adapter planning and implementation can be separate subsequent work.

No issue tracker was changed. These recommendations assume 3187140 covers selection/onboarding readiness and 3185687 covers baseline execution, as implied by this request; they do not substitute for the actual story acceptance criteria.

10. Recommended next steps

Anton/story owner reviews this report, the mapping CSV and the distinction between metadata-complete selection and unverified historical content. Confirm the acceptance boundary for story 3187140 before closure.
Content/onboarding owners verify historical source-to-index lineage, section labels/text and embedding provenance at the 11 selected checkpoints. Record asset-level decisions. Onboard or reindex only where that review establishes a missing or incorrect historical version.
Verify historical legislation query/version availability using dated source evidence, or explicitly approve the effective-date proxy and citation-only scope. Keep the 143 earlier candidates out of scoring pending their separate review.
In a separately authorized implementation task, provide a historical retrieval adapter that applies publication, release and permitted-date filters during candidate generation and returns all guard-required metadata. Keep existing datasets and the current implementation unchanged during this review.
After those gates pass, revalidate live target metadata and approve story 3185687's baseline execution. Score document-LNI targets using the 42 example–LNI pairs; do not describe them as 52 independently scored sections. No such evaluation was run for this report.

Evidence and scope

Final selection manifest, new legislation, amendments, historical checkpoints.
63 selected V4 source rows, saved validation, V4 schema, V3/V4 comparison.
Timestamped live review evidence, including per-publication metadata and protected input hashes.
Detailed mapping CSV. One row per selected example; JSON-valued columns retain full arrays and subsection provenance.

Only new report/evidence/CSV artifacts were created. Existing datasets, manifests and scripts were not modified. No package was installed and no database, Solr or S3 write or baseline retrieval evaluation occurred.

===== APPENDIX 2: Baseline gap (pilot vs mask pilot) and masking/citation-location subsets =====

baseline_gap.md

Baseline gap: 22 Oct pilot vs mask pilot (MP) original arm

Caveats: the rerun in step 4 uses n=20 examples, so the attribution is directional. The pilot's cached 22 Oct query embeddings were reused, so no embedding API call was made. Read-only throughout.

1. Same 682 examples, pilot q1 (footnote 0.5) vs MP original

Group	n	R@10	R@100	Hit@100	MRR
ALL	682	0.427 → 0.331	0.604 → 0.433	0.735 → 0.576	0.289 → 0.229
new	306	0.404 → 0.294	0.609 → 0.418	0.719 → 0.546	0.307 → 0.257
updated	376	0.445 → 0.362	0.600 → 0.446	0.747 → 0.601	0.273 → 0.207
before	663	0.428 → 0.333	0.609 → 0.435	0.744 → 0.582	0.293 → 0.233
at	8	0.375 → 0.250	0.500 → 0.500	0.500 → 0.500	0.150 → 0.177
after	11	0.364 → 0.273	0.364 → 0.273	0.364 → 0.273	0.122 → 0.044

Expected-target counts per example match the pilot exactly (0 mismatches). The gap is a real drop on the same examples.

2. Query text (20 statutes, seed 20261031)

Pilot q1 is the full legislation text (13k–91k tokens), truncated to 8000 tokens at embedding time.
E2 A5 is a compact text (200–9.5k tokens).
The MP original is the M0 frozen A5 span. It equals A5 for 18 of 20 statutes, and the other 2 differ only by the 8000-token truncation.
q1 never equals A5 or the MP original.

3. Collection now vs 21 Oct

Chunks: 239,174 → 1,082,019 (about 4.5×).
Pubs: 144 → 636 (493 new, 1 gone, 27 identical, 116 with changed chunk counts).
Every pub has exactly one release now.
Targets: 270 of 272 distinct target LNIs are present now. 679 of 682 examples have all targets present, and 3 examples have none.

4. Rerun of pilot q1 on 20 examples (current collection, original arm)

Arm	R@10	R@100	Hit@100	MRR
Saved 22 Oct hits, re-scored	0.458	0.580	0.75	0.287
Same q1, rerun now	0.391	0.433	0.65	0.209
MP original, same examples	0.391	0.452	0.75	0.200

The saved-file and re-scored metrics are identical for 20/20 examples. Top-10 overlap between the saved and current hits is 4.6 of 10.

Cause

Collection changed (dominant). The same q1 on the same examples drops from 0.58 to 0.43 R@100 after the collection grew. 493 new pubs now act as distractors.
Query text (secondary). Moving from q1 to A5 shifts R@100/Hit@100 only slightly. The MRR difference is small and mixed.
Targets missing: negligible (3 of 682 examples).
Scoring: ruled out.

So the pilot and MP baselines are not comparable, and the MP deltas should be read only against the MP original arm.

mark_cite_present.md

Masking where the target cites the law: 12-statute subset (not the planned 150)

Caveats first:

Only 12 statutes (19 examples) could be run, not 150. Only 15 of the 466 pilot at/after statutes have an M0 plus M-self v3 pair. Building more would need new source-identity and mask work from raw XML. I did not do that, so this is anecdotal.
These are the MP rows (footnote 0.5, original vs masked) restricted to the pilot's at/after examples. I ran no new embeddings, so the key was not needed.
Relaxed temporal rule and current statute XML, as in MP.
contains_exact_cite is a case- and whitespace-normalised match of the source-attested aliases (source_identities.json) in the passage or normcite_list of the target chunks. Targets were fetched with 1 read-only Solr request.
Statute-level means. |delta| < 0.01 counts as noise.

Overall (12 statutes)

	orig	masked	delta
R@10	0.333	0.333	0.000
R@100	0.500	0.583	+0.083
Hit@100	0.500	0.583	+0.083
MRR	0.146	0.165	+0.018

R@100: 1 statute better, 0 worse, 11 noise. MRR: 2 better, 0 worse, 10 noise.

Split by contains_exact_cite

contains_exact_cite	statutes	R@100 orig → masked	MRR orig → masked	R@100 better/worse/noise
yes	2	0.50 → 0.50	0.100 → 0.100	0/0/2
no	10	0.50 → 0.60	0.156 → 0.178	1/0/9

Reading

The target section cites the law in only 2 of 12 statutes, despite the at/after label. For these two, masking changed nothing.
The whole gain comes from one statute (6579-3783-GXF6-81P4, R@100 0 → 1) and two small MRR moves. None of them has the exact cite in the target.
No statute got worse. The one slightly negative MRR (62TS-2KJ3-CH1B-T213, -0.007) is under the noise floor.
With 2 cite-present statutes, nothing can be concluded about whether masking hurts when the cite is in the target.

Files: cite_present_12.py, cite_present_per_statute.csv, cite_present_summary.json.

To reach 150

Approve building M0 + M-self v3 masks for more at/after statutes. Only 29 of the 466 have an E2 A5 query now, so A5 would also have to be generated first. New embeddings would then use the key.

cite_where.md

Where the GDS citation appears in the target sections (jurisdiction_test)

Caveats: DEV collection as of today (1.08M chunks, one release per pub), not the 22 Oct state. All 950 pilot "at"/"after" examples (466 statutes, 382 target sections, 30,865 chunk rows; 6 read-only Solr requests). Matching is string-based: exact = raw substring of the GDS citation; normalised = case / § / Sec. / Section / punctuation / either order; weak = section number within 80 chars of the code words (reported only as a flag, not a bucket). "Footnote hit" also counts the GDS CiteDef key (jcite_cite_def) in footnote_citation_keyword_list, which is code+section level and can be broader than the exact subsection. Buckets use the section's first chunk (lowest chunk_seq); the any-chunk view is below. Sections are large (about 30 chunks per example row), so a hit in a late chunk does not mean it is in the text a query matches. No retrieval or masking was run.

Stored fields (3 chunks, `fl=*`)

Holds text	Fields	Vectorised
Body	`passage`	yes (`passage_embedding`, 3072)
Heading	`title_path`, `title`, `section_title`, `chapter_title`	`title_path` only (`title_path_embedding`, 3072)
Citations in body	`citation_text_list`, `normcite_list` (codes such as `33_1311(a) 1828`)	no
Footnotes	`footnote_text_list`, `footnote_citation_text_list`, `footnote_citation_keyword_list`, `footnote_normcite_list`	no (no footnote embedding field exists)
Metadata	pub, release, jurisdiction names/urns, dates	no

Buckets (first chunk; 950 examples)

Bucket	All	new	updated	at	after
in-passage (exact)	56	40	16	11	45
in-passage (normalised only)	3	3	0	0	3
in-footnote field (first chunk)	108	74	34	11	97
in another chunk of the same section	724	632	92	107	617
nowhere in Solr	59	57	2	4	55
(n examples)	950	806	144	133	817

Any chunk of the section: passage 294 (31%), footnote fields 616 (65%), both 19, passage only 275, footnote only 597, neither 59 (6%). By type: new 240 passage / 522 footnote of 806; updated 54 / 94 of 144. By relation: at 42 / 91 of 133; after 252 / 525 of 817. Only 2 of the 59 "nowhere" examples show a weak number-near-code hit.

"Nowhere" examples (5, seed 20261031; `nowhere_examples.csv`)

Example	Pub	GDS new rel / pub rel (new/pub)	Indexed rel	CiteDef	New rel in collection?
11698	54	54 / 53	55	6 U.S.C. § 1501(9)(A)	no (only 55)
07000	41	36 / 35	41	N.Y. Ins. Law § 2601	no (only 41)
06992	41	36 / 35	41	42 Pa. Cons. Stat. § 7401 to 7411	no (only 41)
03450	1626	10 / 9	12	Fla. Stat. § 784.049	no (only 12)
06999	41	36 / 35	41	Mich. Comp. Laws § 500.2006	no (only 41)

In all 5 the section is from the indexed (later) release, not the release where the GDS row added the citation, and the text is general commentary without the statute (chunk starts are in the CSV). The index may not carry the GDS label for these.

Answer for Anton

The citations are mostly in the footnote fields (65% of examples, 597 of them footnote-only), far less in the passage text (31%), and absent from the index for 6% (59). Those footnote fields are stored but not vectorised; only passage and title_path have embeddings.

What would change if footnotes were vectorised or searched lexically

Hypothesis only (not tested here): for about 600 examples the cite is visible only in footnotes, so a footnote embedding or a lexical match on footnote_citation_text_list could pair the query's cite with the target; the 59 absent cases and the 19 passage-and-footnote cases would be unaffected or already visible. This also means any retrieval gain from a cite-bearing query could come from footnote matching, not from the passage.

Files: cite_where.csv (example × chunk), cite_where_examples.csv, nowhere_examples.csv, cite_where_summary.json, cite_where.py.

===== APPENDIX 3: Pilot legislation XML and query-text inventory =====

xml_inventory

Pilot legislation XML inventory

Read-only audit prepared October 22, 2026. Sampling seed 20261023 (a seed, not the audit date). Existing pilot inputs/code were read only; all new files are under b0_eval/e2/. No retrieval, embeddings, index updates, or onboarding were run.

Findings

q1 used the entire XML text, including editorial annotations and metadata. 1,072 / 2,354 example queries (45.54%) were truncated. Mean document loss was 31.03% of tokens across all examples, or 68.15% among truncated examples. No dedicated section-summary element was found in the 40-document inventory. Notes contain secondary-source references and should be excluded from clean section queries. The 10-document review found no observed citation to its selected target publication; this limited check does not establish that annotations are safe.

Cache and sample

The pilot cache ../pilot/legislation_text/ contains 1,022 .txt files, not raw XML: check_prereqs.py discarded the markup after parsing. These cached texts were reused for all q1 measurements. Raw markup was missing, so exactly the 40 selected source URLs were fetched into raw_xml/; no other legislation was downloaded. 40/40 fetched XML flattenings match the pilot cached text exactly. raw_fetch_manifest.json records fetch status.

sample_documents.csv contains 40 distinct legislation LNIs: 20 new and 20 updated, with a global maximum of two per source jurisdiction. Population: input-ready pilot examples with a cached legislation text file. Sort unique LNIs, shuffle using Python random.Random(20261023), then accept a source example under the type quotas and jurisdiction cap; multiple source rows for one document are ordered by type then numeric source row. Thus this is a capped descriptive sample, not a proportional estimate. Jurisdiction comes from the saved GDS source snapshot, not the target publication. sampling.json records the exact allocation.

Exact q1 construction

../pilot/check_prereqs.py:202–221: parse XML response bytes with xml.etree.ElementTree.fromstring(body). Compute payload = " ".join(" ".join(root.itertext()).split()). This walks every element’s .text and every child .tail in XML document order. It inserts spaces between text pieces, then collapses all whitespace to one ASCII space and trims the ends. XML entities are decoded by the parser. Tags, namespace declarations and attributes disappear; no element whitelist, section selection, summary extraction, note removal or metadata removal is applied. Comments/processing instructions are not returned as ordinary text by this parser. Attribute-only dates, IDs, links and normalised citations disappear unless also represented as text.
In all 40 observed documents the top-level order is statcode:head → statcode:body → doc:metadata. Head contributes self-citation and jurisdiction text. Body contributes section name/designation, nested operative section text, annotations/editorial and research notes, history/source notes, and publishing/copyright information in their original XML order. Metadata contributes document title/date/IDs, subjects, hierarchy headings and classifications. Nested structures retain their original interleaving; a global regrouping by part was never performed.
The normalised payload is saved as legislation_text/<sha256(datalake_url)>.txt. ../pilot/run_initial_baseline.py:63–71 reads that entire UTF-8 file and makes one q1 subquery [text]. q0 summaries and GDS citation fields are not appended. Identical query arrays share a SHA-256 query-set key. baseline_queries/*.json stores full text, not the truncated embedding input.
At embedding time ../../openai_embs_utils.py:_truncate first takes 64,000 Python characters (8000 * 8), then encodes using tiktoken cl100k_base, disallowed_special=(). If over 8,000 tokens it decodes only tokens[:8000]. This is a prefix cut, potentially inside a sentence or element. There is no 30,000-character cap when the tokenizer is available. The pilot runtime asserted tokenizer availability and used baseline_embeddings_token8000/; the older embedding directory is not the completed token-capped run. These counts reproduce the saved code/defaults; no exact gateway request-body log was retained.

Denominator	Queries/text files	Truncated	Mean token loss (all)	Mean token loss (truncated)
Pilot example q1 occurrences	2354	1072	31.03%	68.15%
Distinct cached text files	1022	411	26.76%	66.55%

Mean character loss is 30.34% per example (26.14% per distinct file). These are arithmetic means of each document’s lost fraction, not a pooled token fraction. The pooled fraction of original tokens removed is 90.35% across examples (87.48% across files), because very long documents dominate that denominator. Seven of 2,361 input-ready examples have no q1 and are excluded. Counts include repeated source examples sharing text; the distinct-file row avoids that weighting.

q1_truncation.csv records the token/character counts, loss fractions, final retained 150 characters and first omitted 150 characters for every q1 example. Tag locations cannot be recovered for all 1,022 texts without refetching raw XML; this audit maps exact paths for the requested 40 only. sample_cut_locations.csv: 23 untruncated; 16 cut inside annotations; one cut inside operative primlaw:bodytext/.../base:text. All mappings use the verified identical XML text.

Candidate query parts and actual markup

Lengths use cl100k_base. Main six groups partition normalised text segments: nearest annotation/history ancestry takes precedence, then metadata, then citation, authority/date, heading, and operative text. Median is per-document total tokens among documents containing textual content in that part. Thus inline operative citations are separate; annotation citations and annotation headings remain in the annotation group. Definitions and purpose are overlapping semantic subsets and must not be added to the six main groups. Wrapper-path lengths in the appendix include descendants and likewise are not additive.

Part	Actual tags / selection	Text coverage	Median tokens, present docs
Body / operative text	`primlaw:bodytext`, nested `primlaw:level`, `base:p`, `base:text`, lists/tables; exclude annotation/history ancestors. Inline `lnci:*` text measured separately.	40/40 (100%)	836.5
Cross-references / citations	`ref:citations/ref:cite4thisresource/lnci:cite/lnci:content/lnci:span`; operative inline `lnci:cite`, `lnci:content`, `lnci:span`. Citation structures also include `lnci:stat/statref`, `lnci:case`, `lnci:book/bookref`, `lnci:lawrev/lawrevref` in annotations.	40/40 (100%)	30.0
Summary / synopsis / digest	No dedicated summary/synopsis/digest/purpose/abstract element observed. Annotation case summaries use ordinary `annot:annotation/base:bodytext/base:p/base:text`; these are case analysis, not a section synopsis.	0/40 dedicated	0
Jurisdiction / authority / dates	`statcode:head/jurisinfo:legisbodyinfo/jurisinfo:jurisdiction/jurisinfo:location/location:state` or `location:country`; `jurisinfo:system`; `primlaw:levelinfo/primlawinfo:primlawinfo/primlawinfo:dates/primlawinfo:effdate`, `primlawinfo:enddate`.	40/40 (100%)	2.0
Notes / annotations / history	`annot:annotations/annot:annotationsourcegroup/annot:div/annot:annotation-grp`, nested `annot:annotation-grp`, `annot:annotation`, `annot:bodytext`, `base:bodytext`, `base:heading/base:title`, `base:toc`; `primlawhist:primlawhist/primlawhist:histgrp/primlawhist:histitem`; embedded `lnci:`, `ref:lnlink`, `form:`.	40/40 (100%)	1791.0
Other metadata / taxonomy	`doc:metadata`, `dc:metadata/{dc:title,dc:date,dc:identifier,dc:language,dc:source,dc:subject,dc:coverage}`, `dcterms:isPartOf`, `dcterms:references`, `doc:hier/doc:hierlev/base:heading`, `classify:classname`, `pubinfo:publicationname`, `base:copyright`, `lndocmeta:lnlni`.	40/40 (100%)	205.5
Title / heading / catchline	`statcode:body/primlaw:level/base:heading/{base:title,base:desig}` and headings on nested operative `primlaw:level`; optional `base:subtitle`, `base:emph`. Annotation and metadata headings excluded.	40/40 (100%)	62.0
Definitions	Ordinary `primlaw:level/base:heading/base:title` identifying the section/subsection; content under `primlaw:bodytext`, `base:p/base:text`. Semantic heading selector; no definition-specific tag observed.	7/40 (17.5%)	1641
Purpose / findings / intent in operative section	Ordinary `primlaw:level/base:heading/base:title` identifying the section/subsection; content under `primlaw:bodytext`, `base:p/base:text`. Semantic heading selector; no definition-specific tag observed.	0/40 (0%)	0

Effective-date elements occur in 40/40 documents, generally with @normdate, @year, @month, @day; their lack of text explains the very small metadata-part length. jurisinfo:system uses attributes such as @systemcode, @normlongname; jurisinfo:legisbodyinfo/@legisbodycode identifies authority. Preserve these as structured metadata rather than relying on q1’s flattened text.

Definitions require semantic selection: e.g. West Virginia “Definitions.”, Delaware subsection “Definitions. —”, and South Dakota “Definitions. [Effective until July 1, 2026]” use ordinary heading/body tags. Purpose/intent words also occur in annotation headings and taxonomy; those are not proof of a stored section summary. semantic_subparts.json records the operative heading matches. Annotations contain short case digests, research references and sometimes forms (form:form, form:bodytext, form:document, form:p, form:text); these are not a safe summary substitute.

Leakage review: 10 documents

Select five new and five updated from the 40 using a fresh random.Random(20261023) and sample within type in new-then-updated order. Reviewed full XML text and citation/link attributes; checked saved target publication titles, exact expected target LNIs, secondary-source book/law-review citations, and research-reference headings. Matching a statute’s own citation is not a match to the target secondary publication. The source XML normally cites primary law/cases and can also cite secondary works.

Document LNI / type	Selected target publication	Observed secondary references	Target-publication result
5F0Y-BKN1-6F13-04J9-00000-00 / new	1622 — Title absent in saved metadata	None found	No observed target-publication citation
6C5V-XJ73-RYBY-K2MM-00000-00 / new	41 — Art of Advocacy: Settlement	None found	No observed target-publication citation
8RBJ-1Y92-D6RV-H3BD-00000-00 / new	119 — COLLIER LENDING INSTITUTIONS AND THE BANKRUPTCY CODE	None found	No observed target-publication citation
8VFG-YDB2-8T6X-713R-00000-00 / new	629 — WAGES & HOURS	None found	No observed target-publication citation
8V0M-1Y72-D6RV-H0VP-00000-00 / new	38 — Art of Advocacy: Preparation of the Case	None found	No observed target-publication citation
6C9J-YKT3-S47J-T4BG-00000-00 / updated	469 — NO-FAULT AND UNINSURED MOTORIST AUTOMOBILE INSURANCE	New Hampshire Practice; N.H.P.	No observed target-publication citation; secondary-source references present
6F0T-DTD3-RRYV-S276-00000-00 / updated	371 — Family Law and Practice	Ohio Transaction Guide: Family Law & Forms § 11.22	No observed target-publication citation; secondary-source references present
6GN5-C8D3-RY61-C16M-00000-00 / updated	626 — EMPLOYMENT DISCRIMINATION	11 Del. J. Corp. L. 435 (Winter 1987); 14 Del. L. Rev. 69 (2013); 6 Del. L. Rev. 311 (2003); 18 Delaware Lawyer 27; 15 Del. L. Rev. 71 (2014)	No observed target-publication citation; secondary-source references present
6GD4-HMY3-RRTW-N0GR-00000-00 / updated	2918 — American Parentage Law	45 Brandeis L.J. 59 (Fall, 2006); 50 Buff. L. Rev. 341 (Winter, 2002); 37 Fam. L.Q. 35 (Spring, 2003); Massachusetts Probate Manual (MCLE) § 13.1.2	No observed target-publication citation; secondary-source references present
696X-3N93-RXV4-D1TY-00000-00 / updated	46 — CALIFORNIA JUVENILE COURTS PRACTICE AND PROCEDURE	None found	No observed target-publication citation

0/10 exact expected-target LNI matches; 0/10 observed target-publication citations. Publication 1622 lacks a saved title in the inspected metadata; its document has no book/law-review citation, and no target LNI match, but title/alias matching for that target is unresolved. No target-title normalisation catalog was fetched. Citation normpubcode values are not assumed to equal GDS publication IDs. Text inspection distinguishes references such as New Hampshire Practice (not target 469), Ohio Transaction Guide: Family Law & Forms (not target 371), and Massachusetts Probate Manual (not target 2918). These examples demonstrate secondary-source content, not confirmed target leakage. leakage_audit.json retains all reviewed evidence.

Unsafe for clean queries: exclude the entire annot:annotations subtree (including annot:annotation-grp, annot:annotation, annot:bodytext, nested base:bodytext, base:toc, embedded lnci:book, lnci:lawrev, ref:lnlink, and forms). It contains case analysis and secondary references that could name retrieval targets in other documents; the sample does not demonstrate a direct target hit. Also exclude doc:metadata (dc:subject, hierarchy/classification labels and related-content links) and history/editorial material for a section-text baseline. This is a conservative query-design flag, not a claim of confirmed target leakage in these ten. Inline operative citations to other law remain a separately selectable part; no selected target match was observed there. Self-citation and section name are explicit citation-bearing inputs and should be evaluated as a separately declared query variant if retained.

A defensible candidate is the section heading plus operative primlaw:bodytext, preserving legal definitions and excluding all annotation/history/metadata descendants before token budgeting. No experiment or replacement q1 was run.

Complete element/path inventory

1,161 distinct namespace-qualified element paths observed across the 40 documents, including empty and attribute-only elements. Paths omit sibling indices but preserve every nesting level. Coverage counts document presence (including empty elements); token lengths aggregate all occurrences of that exact path within a document, including descendants, before tokenising. Both present-document and all-40 medians are available in element_paths.csv. Each sample below is the first nonempty flattened subtree, capped at 150 Unicode characters; ∅ means no text. Parent and child rows overlap. Namespace prefix base denotes the XML default namespace; namespaces.json maps all prefixes to full namespace URIs.

<details><summary>All paths, coverage, median tokens and 150-character samples</summary>

Element path	Docs / 40	Coverage	Median tokens (present)	Sample (≤150 chars)
/statcode:statcode	40	100.0%	3566.5	19 Del. C. § 711 Delaware Unlawful employment practices; employer practices. § 711. Definitions. — (a) As used in this section: (1) “Certifying body o
/statcode:statcode/doc:metadata	40	100.0%	205.5	Delaware Code Annotated Copyright © 2026 All rights reserved. urn:contentItem:6GN5-C8D3-RY61-C16M-00000-00 § 711. Unlawful employment practices; emplo
/statcode:statcode/doc:metadata/doc:docinfo	40	100.0%	191.5	urn:contentItem:6GN5-C8D3-RY61-C16M-00000-00 § 711. Unlawful employment practices; employer practices. 2026-06-06 en-US urn:contentItem:5JHW-94P1-DYB7
/statcode:statcode/doc:metadata/doc:docinfo/classify:classification	40	100.0%	5.0	Decisis Appellate
/statcode:statcode/doc:metadata/doc:docinfo/classify:classification-grp	40	100.0%	40.5	NEGATIVE NEWS SEX OFFENSES DISCRIMINATION COMPENSATION & BENEFITS EMPLOYEE TRAINING CRIME, LAW ENFORCEMENT & CORRECTIONS CRIMINAL RECORDS EMPLOYMENT A
/statcode:statcode/doc:metadata/doc:docinfo/classify:classification-grp/classify:classification	40	100.0%	40.5	NEGATIVE NEWS SEX OFFENSES DISCRIMINATION COMPENSATION & BENEFITS EMPLOYEE TRAINING CRIME, LAW ENFORCEMENT & CORRECTIONS CRIMINAL RECORDS EMPLOYMENT A
/statcode:statcode/doc:metadata/doc:docinfo/classify:classification-grp/classify:classification/classify:classitem	40	100.0%	40.5	NEGATIVE NEWS SEX OFFENSES DISCRIMINATION COMPENSATION & BENEFITS EMPLOYEE TRAINING CRIME, LAW ENFORCEMENT & CORRECTIONS CRIMINAL RECORDS EMPLOYMENT A
/statcode:statcode/doc:metadata/doc:docinfo/classify:classification-grp/classify:classification/classify:classitem/base:meta	40	100.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:docinfo/classify:classification-grp/classify:classification/classify:classitem/base:meta/base:metaitem	40	100.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:docinfo/classify:classification-grp/classify:classification/classify:classitem/classify:classitem-identifier	36	90.0%	43.0	NEGATIVE NEWS SEX OFFENSES DISCRIMINATION COMPENSATION & BENEFITS EMPLOYEE TRAINING CRIME, LAW ENFORCEMENT & CORRECTIONS CRIMINAL RECORDS EMPLOYMENT A
/statcode:statcode/doc:metadata/doc:docinfo/classify:classification-grp/classify:classification/classify:classitem/classify:classitem-identifier/classify:classcode	36	90.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:docinfo/classify:classification-grp/classify:classification/classify:classitem/classify:classitem-identifier/classify:classname	36	90.0%	43.0	NEGATIVE NEWS SEX OFFENSES DISCRIMINATION COMPENSATION & BENEFITS EMPLOYEE TRAINING CRIME, LAW ENFORCEMENT & CORRECTIONS CRIMINAL RECORDS EMPLOYMENT A
/statcode:statcode/doc:metadata/doc:docinfo/classify:classification-grp/classify:classification/classify:classitem/topic:topicReference	40	100.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:docinfo/classify:classification/classify:classitem	40	100.0%	5.0	Decisis Appellate
/statcode:statcode/doc:metadata/doc:docinfo/classify:classification/classify:classitem/classify:classitem-identifier	40	100.0%	5.0	Decisis Appellate
/statcode:statcode/doc:metadata/doc:docinfo/classify:classification/classify:classitem/classify:classitem-identifier/classify:classcode	40	100.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:docinfo/classify:classification/classify:classitem/classify:classitem-identifier/classify:classname	40	100.0%	5.0	Decisis Appellate
/statcode:statcode/doc:metadata/doc:docinfo/dc:metadata	40	100.0%	112.0	urn:contentItem:6GN5-C8D3-RY61-C16M-00000-00 § 711. Unlawful employment practices; employer practices. 2026-06-06 en-US urn:contentItem:5JHW-94P1-DYB7
/statcode:statcode/doc:metadata/doc:docinfo/dc:metadata/dc:coverage	40	100.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:docinfo/dc:metadata/dc:coverage/globalentity:entityReference	40	100.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:docinfo/dc:metadata/dc:coverage/location:country	2	5.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:docinfo/dc:metadata/dc:coverage/location:state	38	95.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:docinfo/dc:metadata/dc:date	40	100.0%	6.0	2026-06-06
/statcode:statcode/doc:metadata/doc:docinfo/dc:metadata/dc:identifier	40	100.0%	65.0	urn:contentItem:6GN5-C8D3-RY61-C16M-00000-00 urn:contentItem:5JHW-94P1-DYB7-W51H-00000-00 6GN5-C8D3-RY61-C16M-00000-00
/statcode:statcode/doc:metadata/doc:docinfo/dc:metadata/dc:language	40	100.0%	2.0	en-US
/statcode:statcode/doc:metadata/doc:docinfo/dc:metadata/dc:source	40	100.0%	4.0	181 5080
/statcode:statcode/doc:metadata/doc:docinfo/dc:metadata/dc:subject	37	92.5%	17	pregnancy unlawful employment practice religion military gender sex sexual orientation national origin genetic marital status color individual’s class
/statcode:statcode/doc:metadata/doc:docinfo/dc:metadata/dc:title	40	100.0%	14.5	§ 711. Unlawful employment practices; employer practices.
/statcode:statcode/doc:metadata/doc:docinfo/dc:metadata/dcterms:isPartOf	2	5.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:docinfo/dc:metadata/dcterms:references	16	40.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:docinfo/doc:alertmanagement	8	20.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:docinfo/doc:alertmanagement/doc:alertitem	8	20.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:docinfo/doc:alertmanagement/doc:alertitem/doc:referenceitem	8	20.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:docinfo/doc:bookseqnum	27	67.5%	6	45020080000020833
/statcode:statcode/doc:metadata/doc:docinfo/doc:changemanagementcode	40	100.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier	40	100.0%	22.5	Labor Title 19 General Provisions Part I Employment Practices Chapter 7 Discrimination in Employment Subchapter II
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev	40	100.0%	22.5	Labor Title 19 General Provisions Part I Employment Practices Chapter 7 Discrimination in Employment Subchapter II
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/base:heading	40	100.0%	7.0	Labor Title 19
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/base:heading/base:desig	32	80.0%	3.0	Title 19
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/base:heading/base:title	40	100.0%	3.0	Labor
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev	40	100.0%	15.0	General Provisions Part I Employment Practices Chapter 7 Discrimination in Employment Subchapter II
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/base:heading	40	100.0%	8.0	General Provisions Part I
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/base:heading/base:desig	39	97.5%	4	Part I
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/base:heading/base:title	40	100.0%	5.0	General Provisions
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev	25	62.5%	9	Employment Practices Chapter 7 Discrimination in Employment Subchapter II
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev/base:heading	25	62.5%	6	Employment Practices Chapter 7
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev/base:heading/base:desig	18	45.0%	3.0	Chapter 7
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev/base:heading/base:title	25	62.5%	4	Employment Practices
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev	7	17.5%	14	Discrimination in Employment Subchapter II
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/base:heading	7	17.5%	7	Discrimination in Employment Subchapter II
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/base:heading/base:desig	6	15.0%	3.5	Subchapter II
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/base:heading/base:title	7	17.5%	4	Discrimination in Employment
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev	3	7.5%	14	Wards—Temporary Custody and Detention Article 15.
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/base:heading	3	7.5%	6	Wards—Temporary Custody and Detention Article 15.
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/base:heading/base:desig	3	7.5%	4	Article 15.
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/base:heading/base:title	3	7.5%	2	Wards—Temporary Custody and Detention
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev	2	5.0%	9.0	Administration Liability of Landowners Part 733
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/base:heading	2	5.0%	5.5	Administration
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/base:heading/base:desig	1	2.5%	4	Part 6.
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/base:heading/base:title	2	5.0%	3.5	Administration
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev	1	2.5%	8	Liability of Landowners Part 733
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/base:heading	1	2.5%	8	Liability of Landowners Part 733
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/base:heading/base:desig	1	2.5%	3	Part 733
/statcode:statcode/doc:metadata/doc:docinfo/doc:hier/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/doc:hierlev/base:heading/base:title	1	2.5%	5	Liability of Landowners
/statcode:statcode/doc:metadata/doc:docinfo/doc:legacy-metadata	22	55.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:docinfo/doc:legacy-metadata/base:meta	22	55.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:docinfo/doc:legacy-metadata/base:meta/base:metaitem	22	55.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:docinfo/doc:related-content	1	2.5%	13	2012 Ct. ALS 133 -- See section 32.
/statcode:statcode/doc:metadata/doc:docinfo/doc:related-content/doc:related-content-item	1	2.5%	13	2012 Ct. ALS 133 -- See section 32.
/statcode:statcode/doc:metadata/doc:docinfo/doc:related-content/doc:related-content-item/doc:related-content-link	1	2.5%	13	2012 Ct. ALS 133 -- See section 32.
/statcode:statcode/doc:metadata/doc:docinfo/doc:related-content/doc:related-content-item/doc:related-content-link/ref:lnlink	1	2.5%	13	2012 Ct. ALS 133 -- See section 32.
/statcode:statcode/doc:metadata/doc:docinfo/doc:related-content/doc:related-content-item/doc:related-content-link/ref:lnlink/ref:locator	1	2.5%	0	∅
/statcode:statcode/doc:metadata/doc:docinfo/doc:related-content/doc:related-content-item/doc:related-content-link/ref:lnlink/ref:locator/ref:locator-key	1	2.5%	0	∅
/statcode:statcode/doc:metadata/doc:docinfo/doc:related-content/doc:related-content-item/doc:related-content-link/ref:lnlink/ref:locator/ref:locator-key/ref:key-name	1	2.5%	0	∅
/statcode:statcode/doc:metadata/doc:docinfo/doc:related-content/doc:related-content-item/doc:related-content-link/ref:lnlink/ref:locator/ref:locator-key/ref:key-value	1	2.5%	0	∅
/statcode:statcode/doc:metadata/doc:docinfo/doc:related-content/doc:related-content-item/doc:related-content-link/ref:lnlink/ref:marker	1	2.5%	13	2012 Ct. ALS 133 -- See section 32.
/statcode:statcode/doc:metadata/doc:sourcedocinfo	40	100.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:sourcedocinfo/lndocmeta:docinfo	40	100.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:sourcedocinfo/lndocmeta:docinfo/lndocmeta:lnlni	40	100.0%	0.0	∅
/statcode:statcode/doc:metadata/doc:sourcedocinfo/lnsys:doc-history	2	5.0%	0.0	∅
/statcode:statcode/doc:metadata/pubinfo:pubinfo	40	100.0%	17.0	Delaware Code Annotated Copyright © 2026 All rights reserved.
/statcode:statcode/doc:metadata/pubinfo:pubinfo/base:copyright	40	100.0%	9.0	Copyright © 2026 All rights reserved.
/statcode:statcode/doc:metadata/pubinfo:pubinfo/base:copyright/base:copyright-year	40	100.0%	2.0	2026
/statcode:statcode/doc:metadata/pubinfo:pubinfo/pubinfo:publicationname	40	100.0%	8.0	Delaware Code Annotated
/statcode:statcode/statcode:body	40	100.0%	3353.5	Unlawful employment practices; employer practices. § 711. Definitions. — (a) As used in this section: (1) “Certifying body or organization” means an i
/statcode:statcode/statcode:body/primlaw:level	40	100.0%	3353.5	Unlawful employment practices; employer practices. § 711. Definitions. — (a) As used in this section: (1) “Certifying body or organization” means an i
/statcode:statcode/statcode:body/primlaw:level/annot:annotations	35	87.5%	3058	Cross references. Research References and Practice Aids Revisor’s note. Section 24 of 77 Del. Laws, c. 90 , provided: “The inclusion in this act of th
/statcode:statcode/statcode:body/primlaw:level/annot:annotations/annot:annotationsourcegroup	35	87.5%	3058	Cross references. Research References and Practice Aids Revisor’s note. Section 24 of 77 Del. Laws, c. 90 , provided: “The inclusion in this act of th
/statcode:statcode/statcode:body/primlaw:level/annot:annotations/annot:annotationsourcegroup/annot:div	35	87.5%	3058	Cross references. Research References and Practice Aids Revisor’s note. Section 24 of 77 Del. Laws, c. 90 , provided: “The inclusion in this act of th
/statcode:statcode/statcode:body/primlaw:level/annot:annotations/annot:annotationsourcegroup/annot:div/annot:annotation-grp	35	87.5%	3058	Cross references. Research References and Practice Aids Revisor’s note. Section 24 of 77 Del. Laws, c. 90 , provided: “The inclusion in this act of th
/statcode:statcode/statcode:body/primlaw:level/annot:annotations/annot:annotationsourcegroup/annot:div/annot:annotation-grp/annot:annotation	29	72.5%	60	19 Del. C. Pt. I, Ch. 7, Subch. II
/statcode:statcode/statcode:body/primlaw:level/annot:annotations/annot:annotationsourcegroup/annot:div/annot:annotation-grp/annot:annotation-grp	31	77.5%	4927	Revisor’s note. Section 24 of 77 Del. Laws, c. 90 , provided: “The inclusion in this act of the words ‘sexual orientation’ is intended to ensure equal
/statcode:statcode/statcode:body/primlaw:level/annot:annotations/annot:annotationsourcegroup/annot:div/annot:annotation-grp/annot:annotation-grp/annot:annotation	29	72.5%	1423	Section 24 of 77 Del. Laws, c. 90 , provided: “The inclusion in this act of the words ‘sexual orientation’ is intended to ensure equal rights and not
/statcode:statcode/statcode:body/primlaw:level/annot:annotations/annot:annotationsourcegroup/annot:div/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp	23	57.5%	855	Civil penalties, § 715 of this title. Department of Transportation projects, § 2003 of Title 2 . Exclusion from lottery games, § 4835 of Title 29 . Pr
/statcode:statcode/statcode:body/primlaw:level/annot:annotations/annot:annotationsourcegroup/annot:div/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation	23	57.5%	844	Civil penalties, § 715 of this title. Department of Transportation projects, § 2003 of Title 2 . Exclusion from lottery games, § 4835 of Title 29 . Pr
/statcode:statcode/statcode:body/primlaw:level/annot:annotations/annot:annotationsourcegroup/annot:div/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp	8	20.0%	6004.0	—Age discrimination. Professor adduced sufficient evidence to support a claim of age discrimination; the shifting nature of the university’s explanati
/statcode:statcode/statcode:body/primlaw:level/annot:annotations/annot:annotationsourcegroup/annot:div/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation	7	17.5%	4443	Professor adduced sufficient evidence to support a claim of age discrimination; the shifting nature of the university’s explanations for the professor
/statcode:statcode/statcode:body/primlaw:level/annot:annotations/annot:annotationsourcegroup/annot:div/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp	4	10.0%	970.0	— — Pension restrictions. Exclusion of county employees’ United States military employment years of service (as well as other United States years of s
/statcode:statcode/statcode:body/primlaw:level/annot:annotations/annot:annotationsourcegroup/annot:div/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation	4	10.0%	961.0	Exclusion of county employees’ United States military employment years of service (as well as other United States years of service) from a pension eli
/statcode:statcode/statcode:body/primlaw:level/annot:annotations/annot:annotationsourcegroup/annot:div/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp	1	2.5%	464016	Government properly charged lawful permanent resident with inadmissibility because nothing in Immigration and Nationality Act required border officer
/statcode:statcode/statcode:body/primlaw:level/annot:annotations/annot:annotationsourcegroup/annot:div/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation	1	2.5%	240886	Government properly charged lawful permanent resident with inadmissibility because nothing in Immigration and Nationality Act required border officer
/statcode:statcode/statcode:body/primlaw:level/annot:annotations/annot:annotationsourcegroup/annot:div/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp/annot:annotation-grp	1	2.5%	222414	Board of Immigration Appeals properly found citizen of India excludab