# Supplementary studies

These studies examine how scoring rules, prompts and repeated inference affect
benchmark results. Semantic relaxation uses the released 11,534 images. All
other studies use older frozen ground truth and populations described below;
their scores must not be read as results on the current release.

Historical results are aggregate evidence. Their real-control images and
per-image records are not distributed, so the release cannot reproduce those
experiments end to end. The JSON files retain the original scores and source
fingerprints. This guide explains their populations, findings and limitations.

## Semantic relaxation

**Question:** How much does exact class naming affect the measured score? Some
classes describe related defects in different settings, such as faulty estate
lights and faulty street lights. The fixed BGE-large-en-v1.5 policy accepts 44
undirected class pairs with cosine similarity strictly above 0.80. It uses no
transitive closure or manual overrides. These are algorithmically accepted
pairs, not expert-confirmed equivalences.

The analysis keeps predictions fixed, matches exact labels first, then matches
remaining eligible pairs one to one. Localisation also requires the selected
intersection-over-union (IoU) threshold. On all 11,534 shared-valid images,
classification micro F2 changes as follows:

| Model            |  Exact | Relaxed |
| ---------------- | -----: | ------: |
| YOLO-World       | 0.1193 |  0.1490 |
| Gemma 4 12B      | 0.3764 |  0.4131 |
| GLM 5.3 Flash    | 0.6264 |  0.6679 |
| GPT-5.6 Luna     | 0.5937 |  0.6391 |
| Muse Glimmer 30B | 0.5990 |  0.6424 |
| Gemini 3.8 Flash | 0.6710 |  0.7106 |

The increases reflect changed acceptance rules, not improved model predictions.
Exact scoring remains primary. The separate mapping-opportunity analysis selects
images whose gold contains a class with an accepted alternative; it scores all
targets and predictions on those images. Its denominator differs from the full
population. Sensitivity varies IoU from 0.10 to 0.90 and semantic acceptance
from 0.50 to 1.00 on the same retained detections, without confidence filtering.

In [semantic results](../evaluation/supplementary/semantic/results.json),
`models[].gold_scopes.all.common` contains full-population comparisons,
`mapping_opportunity` contains the conditional analysis, and `sensitivity`
contains the threshold grids. The frozen matrix and scoring code are supplied;
the primary scoring example checks exact scores only.

## Historical model comparisons

The preliminary benchmark compared ten models on 302 images; the expanded
benchmark compared thirteen on 1,200. All models within each published
comparison have valid predictions for its full population. The expanded study
adds negative controls, allowing false alarms on clean images to be measured.

| Benchmark   | Synthetic images | Undistributed real controls | Shared valid |
| ----------- | ---------------: | --------------------------: | -----------: |
| Preliminary |              301 |                           1 |          302 |
| Expanded    |            1,093 |                         107 |        1,200 |

These are separate experiments with older gold, different composition and model
rosters. Differences between their scores are not paired improvements on the
current dataset. See the
[302-image results](../evaluation/supplementary/historical/benchmark302/results.json)
and
[1,200-image results](../evaluation/supplementary/historical/benchmark1200/results.json).
`experiment` describes each cohort; `models[].scopes` contains valid/common
scores, and the adjacent CSV files give compact metric tables.

## Gemma prompt grouping

**Question:** Does asking about fewer classes in each prompt improve detection?
Gemma 4 12B receives all 97 classes in one prompt, five disjoint class bundles,
or thirteen domain prompts. Every image receives every prompt; gold labels do
not select which prompts are used. The one-pass baseline is reused.

Five-pass and domain-pass each have 1,198 valid responses out of 1,200, leaving
1,196 shared-valid images across all three strategies:

| Strategy               | Classification F2 | Localisation F2 at IoU 0.50 |
| ---------------------- | ----------------: | --------------------------: |
| One pass               |            0.4071 |                      0.1645 |
| Five passes            |            0.4668 |                      0.1129 |
| Thirteen domain passes |            0.4958 |                      0.1331 |

More passes increase classification recall and F2, but reduce localisation F2.
They also increase false alarms on the 200 shared-valid negative images: 6 for
one pass, 20 for five passes and 19 for domain passes. Each configuration was
observed once. Per-pass output caps mean that total output capacity changes with
the number of passes; this does not isolate prompt grouping alone.

[Gemma strategy results](../evaluation/supplementary/gemma-strategies/results.json)
record the population in `study`, prompts in `strategies[].configuration`, and
these scores in `strategies[].scopes.common`. Per-model valid scores use their
own populations and can differ from this table.

## Original-settings repeatability

**Question:** Does the same deployed configuration reproduce its outputs? The
original run and two additional runs cover the same 1,200 images for each of
four models, with all responses valid. Gemma uses greedy decoding; GLM requests
temperature zero; GPT-5.6 Luna uses its provider default. YOLO retains its
original inference settings.

| Model         | Identical class sets across all three runs |
| ------------- | -----------------------------------------: |
| YOLO-World    |                                     100.0% |
| Gemma 4 12B   |                                     100.0% |
| GLM 5.3 Flash |                                     40.08% |
| GPT-5.6 Luna  |                                     57.00% |

Gemma and YOLO also reproduce their boxes exactly. These are agreement rates,
not accuracy: repeating an error still counts as agreement. Hosted provider
routing varied, so this measures the deployed configuration rather than
isolating intrinsic sampling variance. Three runs provide descriptive evidence,
not a precise estimate of long-run variability.

[Repeatability results](../evaluation/supplementary/repeatability/results.json)
contain `models[].pairs` for each run pair, `all_three` for three-way agreement,
and `variation` for F2 ranges and sample standard deviations. Exact-box
agreement includes matching empty predictions; pairwise box F1 instead allows
same-class spatial matches at IoU 0.50. Performance is scored against gold,
separately from agreement between runs.

## Sampling at temperature 0.3

**Question:** How stable are predictions when sampling is explicitly enabled?
Gemma 4 12B, GLM 5.3 Flash and Llama 4 Maverick each have three new runs at
temperature 0.3, top-p 1.0 and top-k 0, with seeds 24092401–24092403. The image
population, taxonomy and one-pass prompt are fixed. Hosted providers are pinned
for this study; a requested seed does not guarantee identical hosted execution.

| Model            | Shared valid | Same classes in all three runs | Classification F2 range |
| ---------------- | -----------: | -----------------------------: | ----------------------: |
| Gemma 4 12B      |        1,198 |                          75.0% |           0.3991–0.4058 |
| GLM 5.3 Flash    |        1,200 |                          42.5% |           0.5848–0.5969 |
| Llama 4 Maverick |        1,200 |                          77.0% |           0.4822–0.4877 |

Individual predictions vary more than aggregate F2. Gemma has one invalid
response in each of its first two runs, on different images; these failures
explain its smaller common population. Pairwise comparisons can therefore have
different denominators. Old and new hosted deployments differ in provider
selection as well as sampling settings, so their difference is not an isolated
temperature effect. One temperature and three seeds do not establish a general
temperature-response curve or a causal model-size effect.

[Sampling results](../evaluation/supplementary/temperature-sampling/results.json)
use the same `runs`, `pairs`, `all_three` and `variation` structure as the
repeatability study; `sampling` records the settings.

## Parking prompt definitions

**Question:** Do explicit annotation definitions help the model recognise
parking violations? GLM 5.3 Flash is evaluated on 47 parking-positive images
from the older 1,200-image cohort. Each of U62–U64 has 16 positive images, with
one image belonging to two classes. Both arms receive all 97 categories; scoring
considers only the three parking classes. The comparison reuses the original
predictions and adds a new arm with class definitions, examples and exclusions,
without image-specific answers.

| Task                     | Original prompt F2 | With definitions F2 |
| ------------------------ | -----------------: | ------------------: |
| Classification           |             0.1515 |              0.6444 |
| Localisation at IoU 0.50 |             0.1126 |              0.5622 |

Both arms have 47 valid responses. The improvement is specific to this selected
subset. There are no clean-scene parking controls, so the study cannot measure
false-alarm rates on those scenes. Prompt length also changes, and single runs
on different dates do not control provider drift or run-to-run variation.

[Parking results](../evaluation/supplementary/parking-prompts/results.json)
contain `arms[].scores`, `classes` for support, `prompts` for the actual text,
and `delta_f2` for absolute score differences. Scores range from 0 to 1: 0.4929
is a 49.29-percentage-point increase, not a 49.29% relative increase.