# Building legislation selection — dataset card

**Version:** 1.0, legal reference date 10 October 2026.
**Status:** frozen release, 2026-10-11.

The dataset tests whether a decision model can select applicable Australian
building provisions from a realistic synthetic application. It contains the
original 500 planned items across all eight states and territories, including
requests with deliberately missing location or other decisive information.
It is a benchmark of the stated rules and cases, not a representative sample of
permit applications or a basis for making live permit decisions.

## Records and labels

Each request supplies an application, optional reference notes, 4–20 candidate
provisions and typed questions. Six document styles and both JSON/text state
formats are represented. The questions ask for the governing NCC edition,
principal building class, multiple classification, construction type,
information sufficiency and a binary applicability answer for each candidate.
[LABEL_SPEC.md](LABEL_SPEC.md) defines the labels; [METHOD.md](METHOD.md)
describes generation, blind checks, adjudication and evaluation.

Gold is computed from structured facts and executable rules, then reviewed
against independent annotations and source evidence. Designed information-gap
items have a null-gold candidate excluded from scoring. Entire contested records
are exported separately. `bls-v1-0006` is currently contested because its pool
energy candidates cannot be resolved reliably from the stated equipment scope;
its provisional answers are not settled benchmark gold.

Release files are `release/{all,train,dev,test,contested}.jsonl`,
`release/questions_only/{train,dev,test}.jsonl` and `release/manifest.json`.
Request-only files contain model input. Full records also contain gold, facts,
rationales, citations, candidate mappings, trap tags, token counts and provenance.
The legacy archetype hint `facts.principal_class` is excluded from release facts;
use the physical `facts.parts` and derived `gold.building_class` instead.
The manifest identifies file and source hashes. A frozen manifest is the
authoritative release identifier.

## Coverage and splits

The current finalizer records retain all **500 planned items**: **499 scored**
and one separately contested record, `bls-v1-0006`. The original split allocation
is preserved; excluding that QLD test item leaves 250 train, 62 dev and 187 test
items. The following realised counts come from
[the released composition audit](../release/composition.json).

| Scored slice | Train | Dev | Test | Total |
|---|---:|---:|---:|---:|
| Scored items | 250 | 62 | 187 | 499 |
| Short tier | 120 | 23 | 57 | 200 |
| Medium tier | 84 | 24 | 91 | 199 |
| Long tier | 46 | 15 | 39 | 100 |
| JSON state | 159 | 39 | 101 | 299 |
| Text state | 91 | 23 | 86 | 200 |
| Reference notes supplied | 250 | 62 | 112 | 424 |
| Closed book | 0 | 0 | 75 | 75 |
| Scored questions | 3432 | 892 | 2870 | 7194 |
| Null-gold questions | 14 | 7 | 18 | 39 |

| Jurisdiction | Train | Dev | Test | Total |
|---|---:|---:|---:|---:|
| ACT | 15 | 10 | 33 | 58 |
| NSW | 35 | 11 | 32 | 78 |
| NT | 22 | 4 | 14 | 40 |
| QLD | 35 | 7 | 25 | 67 |
| SA | 23 | 8 | 17 | 48 |
| TAS | 29 | 5 | 24 | 58 |
| VIC | 47 | 7 | 24 | 78 |
| WA | 37 | 5 | 16 | 58 |
| Unknown | 7 | 5 | 2 | 14 |

| Edition gold | Train | Dev | Test | Total |
|---|---:|---:|---:|---:|
| NCC 2022 | 140 | 37 | 66 | 243 |
| NCC 2025 | 44 | 8 | 62 | 114 |
| Applicant may elect | 22 | 6 | 22 | 50 |
| Insufficient information | 26 | 8 | 8 | 42 |
| Out of scope | 18 | 3 | 29 | 50 |

| Principal class gold | Train | Dev | Test | Total |
|---|---:|---:|---:|---:|
| Class 1a | 72 | 14 | 29 | 115 |
| Class 1b | 5 | 0 | 1 | 6 |
| Class 2 | 28 | 10 | 28 | 66 |
| Class 3 | 11 | 4 | 12 | 27 |
| Class 4 | 1 | 0 | 1 | 2 |
| Class 5 | 11 | 3 | 12 | 26 |
| Class 6 | 18 | 5 | 14 | 37 |
| Class 7a | 9 | 3 | 15 | 27 |
| Class 7b | 6 | 1 | 4 | 11 |
| Class 8 | 15 | 5 | 5 | 25 |
| Class 9a | 6 | 2 | 4 | 12 |
| Class 9b | 6 | 3 | 15 | 24 |
| Class 9c | 4 | 1 | 7 | 12 |
| Class 10a | 12 | 1 | 3 | 16 |
| Class 10b | 20 | 5 | 3 | 28 |
| Class 10c | 1 | 1 | 1 | 3 |
| Insufficient information | 7 | 1 | 4 | 12 |
| No building work | 18 | 3 | 29 | 50 |

| Work, presentation and difficulty | Train | Dev | Test | Total |
|---|---:|---:|---:|---:|
| New work | 146 | 38 | 115 | 299 |
| Alteration/addition | 41 | 9 | 22 | 72 |
| Change of use | 12 | 5 | 14 | 31 |
| Minor work | 33 | 7 | 7 | 47 |
| Work type unstated/out of scope | 18 | 3 | 29 | 50 |
| BCA report | 30 | 11 | 34 | 75 |
| Certifier intake | 43 | 15 | 42 | 100 |
| Council referral | 24 | 7 | 19 | 50 |
| Email | 41 | 5 | 29 | 75 |
| Mixed bundle | 21 | 4 | 24 | 49 |
| Portal JSON | 91 | 20 | 39 | 150 |
| Difficulty 1 | 67 | 15 | 16 | 98 |
| Difficulty 2 | 110 | 30 | 60 | 200 |
| Difficulty 3 | 73 | 17 | 111 | 201 |
| Information sufficiency 0 | 181 | 40 | 137 | 358 |
| Information sufficiency 1 | 14 | 7 | 16 | 37 |
| Information sufficiency 2 | 55 | 15 | 34 | 104 |

Multiple classification is positive for 162 scored items and negative for 337.
Construction type is A for 110, B for 23, C for 144 and not applicable for 222.
The composition artifact also retains the full 31-archetype distribution and all
of these dimensions by split.

| Candidate labels | Train | Dev | Test | Total |
|---|---:|---:|---:|---:|
| Positive | 712 | 177 | 621 | 1510 |
| Hard negative | 776 | 197 | 947 | 1920 |
| Easy negative | 694 | 208 | 367 | 1269 |
| Unscored | 14 | 7 | 18 | 39 |

Across 4,699 scored candidate questions, 32.1% are positive, 40.9% are hard
negatives and 27.0% are easy negatives. The corresponding test shares across
1,935 scored candidates are 32.1%, 48.9% and 19.0%. The 39 null-gold candidate
questions occur in 39 different items and are excluded from the 7,194-question
scoring denominator. Test scoring contains 2,870 questions in 187 items.

| Hard-negative tag | Test items | All scored items |
|---|---:|---:|
| HN-CLASS | 48 | 106 |
| HN-COND | 58 | 104 |
| HN-CONFLICT | 31 | 61 |
| HN-DATE | 61 | 134 |
| HN-INJ | 30 | 36 |
| HN-JUR | 52 | 181 |
| HN-LEX | 33 | 80 |
| HN-NONADOPT | 38 | 74 |
| HN-PART | 28 | 41 |
| HN-SCOPE | 31 | 53 |
| HN-STALE | 30 | 59 |
| HN-THRESH | 36 | 75 |
| HN-TRUNC | 37 | 74 |
| HN-VER | 92 | 192 |

Hard-negative tags can overlap within an item. They exercise version, date,
scope, class, thresholds, jurisdiction, conflicting fields, stale references
and injected instructions. Their prevalence does not estimate how often these
situations arise in practice.

After label repairs and the contested exclusion, HN-COND and HN-PART fall two
test items below the guide's respective targets of 60 and 30. The original
splits and corrected labels are preserved; [GUIDE_CORRECTIONS.md](GUIDE_CORRECTIONS.md)
records the shortfall and its causes.

One item belongs to each family. Closed-book cases, selected future
pre-lodgement cases, ACT boundary dates and held-out councils are test-only.
The resulting splits differ in difficulty and edition distribution.

Tier budgets use Qwen3.5 tokens for the complete state and questions: up to 1,800
for short, 1,800–8,000 for medium and 8,000–24,000 for long. Actual scored requests
range from 1,403 to 17,155 tokens, with a median of 4,706. Actual local-model input
lengths can differ; the 8,192-token local-model limit is part of the evaluation
constraints.

The current candidate library has **268 units**, with SHA-256
`76ee50bd5048680b6cbd01c7e7194c1b85d79c20340e8cf35502ab0443f01333`.
Frozen at `2026-10-11T07:28:28.920519+00:00`. The [manifest](../release/manifest.json)
records release, source and evaluation hashes. `all.jsonl` SHA-256:
`b8f9cbbc0881e767f1e5aefdf12a76e0770ed7061e92809620c61cc5bc4bb7ce`.

## Provenance and quality controls

The initial work used Claude Opus 5.5. The user authorized the current Codex
session model and inherited sub-agents to continue when the former runner was
unavailable. Separate roles performed writing, complete-document blind
extraction, handbook-only primary annotation and evidence-based adjudication.
GLM-5.3-Flash via SCX.ai supplied the independent model-family second opinion.
One request, `bls-v1-0400`, uses an explicitly recorded `reasoning_effort=low`
retry after repeated provider timeouts and disconnects. The other 499 opinions
use the original defaults. The override is included in the annotation cache
identity and exported provenance; its input and handbook are unchanged.
Per-item provenance distinguishes original artifacts, continuation work, repairs
and question-clarification transfers; there is no claim that all items were
produced or freshly annotated by one model.

Release requires all 500 planned IDs, current extraction/application and
annotation/request fingerprints, complete A1 and GLM answers, resolved scored
disagreements, automated QC and cross-split duplicate checks. A1 transfers after
the principal-class wording clarification are explicitly recorded; they are not
fresh annotation calls. [GUIDE_CORRECTIONS.md](GUIDE_CORRECTIONS.md) and
`work/continuation_provenance.json` retain the repair history, including deck and
ancillary classification, carpark and boarding capacities, BASIX document kind,
barrier-only pool energy scope and the ACT conversion predicate.

The QDC historical-version repair added three library units, changed 31 requests
and preserved all gold labels. The roof-scope repair separated balcony/podium
surfaces from roof coverings in 46 specifications; nine applications and
requests changed. It corrected four applicability answers to positive and
three to null, with one information-sufficiency adjustment. Those nine repaired
applications received fresh blind extraction and A1 checks. In `bls-v1-0464`,
independent boundary retaining walls were made explicit and represented as a
separate Class 10b structure, changing multiple classification from 0 to 1;
principal Class 7a, Type C, dates and candidates were preserved. Fresh checking
of repaired requests uses their current fingerprints.

The latest full automated QC pass covers all 500 items, with zero blocking
items and zero near-duplicate pairs. It records 45 warning occurrences across
42 items. The [warning review](../work/_cache/qc_warning_review.json) identifies
44 source-index title/version metadata warnings and one false-positive
Australian Standard attribution warning. These warnings remain visible in the
audit. The [export integrity audit](../work/_cache/final_export_integrity_audit.json)
compares all 500 exported records with current requests, gold labels, candidate
mappings, original splits and unique families. Its 18,339 checks found zero
errors, including checks of manifest hashes, annotation provenance and agreement
counts. Every exported request matches the input used for final model evaluation.

All **500** planned applications have current blind extractions, complete A1 and GLM annotations, and resolved scored disagreements. **96 tests passed**. Current adjudications cover 136 items: 190 GOLD_CORRECT, 3 CONTESTED. Archived earlier defect findings remain in the repair audits. The one contested record is excluded from scored agreement and benchmark results.

The [verification record](../release/verification.json) records coverage, provenance, QC and adjudication counts. The [agreement artifact](../release/agreement.json) gives scored-set denominators; the all-planned audit retaining provisional contested labels is kept separately under work/_cache.

| Question family | A1–gold κ | GLM–gold κ | A1–GLM κ | Scored questions |
|---|---:|---:|---:|---:|
| ncc edition | 0.997 | 0.950 | 0.953 | 499 |
| building class | 1.000 | 0.996 | 0.996 | 499 |
| multi classified | 1.000 | 0.901 | 0.901 | 499 |
| type of construction | 0.991 | 1.000 | 0.991 | 499 |
| applies | 0.987 | 0.978 | 0.975 | 4699 |
| information sufficiency | 0.991 | 0.832 | 0.831 | 499 |

## Model evaluation

The final comparison reports Jev 1.13, `Ihor/some_model@01efaf8` native,
`Tev1-0.8B-experimental@6bb2dff`, and BM25 plus training-majority/fixed defaults.
Local models use one 6 GB RTX 3060 GPU. Context overflow and GPU memory failures
are recorded without truncation or CPU fallback; missing answers count as wrong.
The report's “items with any answer” count means at least one answer, so complete and
partial coverage and failure totals must accompany accuracy.

Use `release/train.jsonl` to fit BM25's majority components. Report held-out test
results separately from the all-items report, which includes training examples.
Scoring includes per-family accuracy, choice macro-F1, binary Brier/ECE and
recall/specificity, information-score MAE, item exact match and slice results.
Null-gold questions are omitted; all-question accuracy weights questions equally.

| System | Test all-question accuracy | Test applicability accuracy | Complete / partial / failed test items |
|---|---:|---:|---|
| Jev 1.13 | 0.747 | 0.804 | 187 / 0 / 0 |
| some_model | 0.371 | 0.412 | 126 / 9 / 52 |
| Tev1-0.8B | 0.430 | 0.462 | 148 / 0 / 39 |
| BM25 + majority | 0.512 | 0.554 | 187 / 0 / 0 |

Test denominators are **187 items, 2,870 scored questions and 1,935 scored applicability questions**. Every evaluated model has a current result or recorded failure for every scored record. Coverage counts all request questions, including null-gold questions excluded from scoring.

| System | Edition | Class | Multiple class | Construction type | Sufficiency | Item exact match |
|---|---:|---:|---:|---:|---:|---:|
| Jev 1.13 | 0.711 | 0.925 | 0.861 | 0.444 | 0.209 | 0.096 |
| some_model | 0.262 | 0.209 | 0.476 | 0.134 | 0.348 | 0.000 |
| Tev1-0.8B | 0.332 | 0.348 | 0.545 | 0.155 | 0.444 | 0.000 |
| BM25 + majority | 0.551 | 0.155 | 0.332 | 0.358 | 0.733 | 0.000 |

| System | Applicability Brier | ECE | Positive recall | Specificity | Sufficiency MAE |
|---|---:|---:|---:|---:|---:|
| Jev 1.13 | 0.127 | 0.057 | 0.609 | 0.896 | 0.864 |
| some_model | 0.555 | 0.250 | 0.138 | 0.541 | 1.226 |
| Tev1-0.8B | 0.425 | 0.167 | 0.515 | 0.437 | 1.153 |
| BM25 + majority | 0.329 | 0.282 | 0.443 | 0.607 | 0.449 |

Brier includes a loss of 1 for a missing answer; ECE covers available probabilities only. The complete JSON reports also include choice macro-F1, abstention precision/recall and per-slice metrics.

| Local model | Short complete / partial / failed | Medium complete / partial / failed | Long complete / partial / failed |
|---|---|---|---|
| some_model | 57 / 0 / 0 | 69 / 9 / 13 | 0 / 0 / 39 |
| Tev1-0.8B | 57 / 0 / 0 | 91 / 0 / 0 | 0 / 0 / 39 |

Failure counts and all-question coverage are available in the [test coverage audit](../release/evaluation_coverage_test.json) and [all-items coverage audit](../release/evaluation_coverage.json). Context and memory failures count wrong, so these results reflect the stated hardware and input budget as well as model capability.

The [held-out test report](../release/report_test/report.md) ([JSON](../release/report_test/report.json)) is the primary comparison. The [all-items report](../release/report/report.md) ([JSON](../release/report/report.json)) includes training examples: its all-question accuracies are Jev 1.13 0.801, some_model 0.395, Tev1-0.8B 0.442, BM25 + majority 0.534.

A1 and GLM scores are agreement references, not independent held-out benchmarks:
their answers helped identify defects and construct gold. Their reported Brier
and ECE use clipped self-reported confidence, not calibrated probabilities.
Local cache reuse requires an identical state, question and checkpoint, but
latency on reused records measures only newly executed questions. These cached
latencies cannot support a fair full-request speed ranking.

## Limits on interpretation

Applications are synthetic, shaped by archetypes, fixed source libraries and
deliberate traps. The dataset does not cover every applicable instrument or
real-world factual dispute. A candidate excerpt can omit surrounding definitions;
legal source review is needed when interpreting a disputed label. The legal
reference date and recorded adoption assumptions constrain future-dated cases.

Blind checks and a separate model-family opinion reduce some errors, but the
writers, primary annotators and adjudicators share model families and generation
assumptions. Their review is not independent expert certification. The final verification
checks document consistency and the stated rule framework; it does not remove
these limits on generalisation.