# How MunicipalBench Final Gold was obtained Final gold is the resolved annotation used for scoring. Human first passes supplied the candidate labels and boxes. Agreement rules resolved compatible submissions; disputed cases could pass through semantic matching, VLM voting or explicit human adjudication. Earlier manager selections were preserved. These routes are recorded per image, so final gold does not imply that all annotators originally agreed or that every image underwent the same review. Each row represents one usable synthetic image. Read `kind` together with `edit_summary.resolution.method_id`, `method_name` and `method_version`. The same `accepted` kind covers both historical manager selections and later explicit adjudication; kind alone is insufficient to identify the method. ## Population and method counts | Method | Ordinary synthetic | Traps | Total | | ------------------------------------- | -----------------: | ------: | ---------: | | GT-001: Auto-agreement | 4,232 | 123 | 4,355 | | GT-002: Rating resolution | 1,177 | 45 | 1,222 | | GT-003: Semantic resolution | 30 | 2 | 32 | | GT-004: Independent VLM council | 3,116 | 303 | 3,419 | | GT-005: Human adjudication | 968 | 115 | 1,083 | | GT-008: Historical manager acceptance | 1,315 | 108 | 1,423 | | **Total** | **10,838** | **696** | **11,534** | GT-006 denotes confirmed unusable images and GT-007 denotes unresolved work. Neither is included in Final Gold. Real controls are also outside this release. There are 7,836 positive and 3,002 negative ordinary synthetic images, and 619 positive and 77 negative traps: 8,455 positives and 3,079 negatives overall. ## GT-001: Original or recovered auto-agreement Export kind: `auto_agreed`; method name: `original_auto_agreement`. Require compatible report states, identical violation-label sets, plausibility and naturalness differences at most 1, and the native box-agreement statistic at least 0.7. Agreed usable negatives have no box-IoU requirement. The original submission hook considers exactly two first-pass annotations when gold is absent. It takes labels and boxes from the first annotation in sorted source-ID order and uses rounded rating midpoints. Exact half-integers use Python's nearest-even rounding. Existing records retain their original labels, boxes and ratings. This release contains 4,354 existing records and one recovered agreement. The recovered record passes label/geometry criteria for every available rater pair and rating ranges at most 1. It keeps the first sorted human's labels/boxes and the rounded midpoint of the first two humans' ratings. `resolution.origin` distinguishes `existing_gold` from `recovered_from_iaa`. ### Native geometry used by GT-001, GT-002 and GT-003 For each shared label, take the maximum IoU among all cross-annotator box pairs. Average those label maxima, round to three decimals, and compare with 0.7. This is not a one-to-one instance assignment or an unmatched-box-zero score. It can pass despite different instance counts or disagreement on extra boxes. Passing it must not be described as full instance consensus. ## GT-002: Conservative rating resolution Export kind: `rating_resolved`; method name: `conservative_rating_resolution`. Require identical labels and passing native geometry for every rater pair. At least one rating dimension has a range greater than 1; otherwise the case belongs to recovered GT-001. Keep the first sorted human's labels and boxes. Set plausibility and naturalness independently to the minimum across all supplied raters (`minimum_all_annotators`). No VLM call is needed. The minimum is an explicitly conservative policy, not evidence that the lower rating is correct. The native geometry limitation above remains applicable. ## GT-003: Direct semantic resolution Export kind: `semantic_resolved`; method name: `direct_semantic_resolution`. Use the 44 algorithmically accepted direct pairs in the distributed semantic policy. Acceptance means cosine similarity above 0.80, not expert validation. There is no transitive closure: accepting A/B and B/C does not imply accepting A/C. Lock exact labels common to all raters. Require a complete one-to-one mapping of the remaining labels where every pair in each group is identical or directly accepted by that policy. Every rater pair must pass native geometry after mapping. Search in sorted label order and choose the first passing mapping. Keep the first sorted human's original labels and boxes, without rewriting them to an umbrella class. Both ratings use the minimum across all raters, even where the original ratings were already within tolerance. ## GT-004: Independent council over human proposals Export kind: `vlm_resolved`; method name: `independent_vlm_council`; method version: `5.1.0`. | Judge slot | Model | Provider | | ---------: | ------------------------------ | ---------------- | | 0 | `deepseek/deepseek-v4.1-flash` | `deepinfra/fp8` | | 1 | `meta-llama/llama-4-maverick` | `deepinfra/base` | | 2 | `qwen/qwen3.8-27b` | `deepinfra/bf16` | The final protocol is `batched_verification_v1`, restricted to `supplied_human_annotations_only`. Judges independently see the original image, human proposal panels and violation definitions. Each judge supplies a batch of categorical ballots for the image. Models assess presence and adequacy of existing boxes; they generate no new labels, coordinates, instances or confidence scores. Missed-instance and coverage-discovery votes are ignored. Group exact labels and unambiguous direct semantic aliases before voting. Preserve an agreed instance only when the first human's box has a unique match in every other rater at IoU at least 0.7, reverse matches are unique, and every matched pair also reaches 0.7. Retain that original human box. Extra or ambiguous instances remain contested. This preservation rule is stricter than the native geometry statistic used in GT-001/002/003. Contested labels receive `present`, `absent` or `uncertain` votes. Proposals receive `keep`, `reject_absent`, `redundant`, `inadequate` or `uncertain` votes. - `joint_presence_and_box_vote`: at least two judges jointly support label presence and retaining that box. - `majority_label_absence`: at least two valid judges reject a label supplied by only some humans; remove its disputed proposals. - `majority_region_absence`: given a positive label majority, at least two valid judges reject the particular region. - `majority_existing_replacements`: at least two positive judges identify retained same-label proposals that cover the redundant proposal. All-human presence or a locked reference cannot be overturned automatically by a majority absence vote. Replacement references must identify directly retained same-label proposals; self-references, chains and cycles do not justify removal. Overlapping alternatives require two judges jointly supporting their retention. Ambiguous semantic overlaps require human review. Missing, malformed or contradictory ballots abstain for the affected label. Inadequate boxes without supported replacements, insufficient majorities and remaining contradictions go to human review. Only fully resolved images receive GT-004 gold. Zero retained violations becomes an explicit clean negative. Ratings use the minimum across all humans. `label_decisions` records ballots and decisions; `source_proposals` records the original labels, boxes and source references; `human_agreements` records mappings and locked instances. The ordered `judges` list decodes judge indices. Compatible earlier ballots may be replayed under the final protocol; `replay_source` retains the run ID, with the private local path removed. ## GT-005: Explicit human adjudication Method name: `human_adjudication`; version: `manager-review-1.0.0`. A manager explicitly approved the final labels, boxes and ratings after image inspection. Human-selected ratings supersede the automatic minimum-rating policy. | Kind | Starting basis | Synthetic | Traps | Total | | ---------- | ------------------------ | --------: | ------: | --------: | | `accepted` | One annotator, unchanged | 699 | 70 | 769 | | `edited` | One annotator, modified | 143 | 25 | 168 | | `authored` | Human union | 106 | 18 | 124 | | `authored` | Blank annotation | 20 | 2 | 22 | | **Total** | | **968** | **115** | **1,083** | The manager may add, remove or redraw boxes, change labels and ratings, or explicitly mark no violations. Union basis does not mean every union box was retained. Only completed adjudications are included; drafts and reopened reviews are excluded. The export was checked against completed review revisions. Preserve `review_revision`, `source_run_id`, `basis`, `status: manager_reviewed`, `platform_status: gold_set` and `rating_policy: human_adjudicated`. The status describes this human-reviewed route, not every automatic record. ## GT-008: Historical manager acceptance Export kind: `accepted`; method name: `existing_manager_acceptance`. A manager selected a supplied human annotation as gold in the original platform. Preserve its payload and agreement snapshot, adding the resolution wrapper with `origin: existing_gold`. Do not imply an automatic-agreement pass or retrospectively apply the later minimum-rating policy. The release-wide `accepted` kind totals 2,192: 1,423 historical GT-008 selections and 769 GT-005 approvals. GT-001/002/003/008 use wrapper version `1.0.0`; this does not identify the original annotation platform's software revision. ## Interpretation `agreement_at_gold` preserves the original snapshot, generally calculated from the first two source annotators. It is not a fresh comparison against final gold or proof of correctness. Historical null snapshots remain null. Use the supplied first-pass annotations for independent agreement analysis rather than treating gold as an additional rater. Conservative minima, model majorities and manager decisions are fallible. Report method composition and the potential dependence when an evaluated model also served on the council. Model performance is reported separately from these ground-truth construction methods. All totals here describe only the distributed synthetic population.