Can routing be separated from detection so a zero-shot decision exactly preserves the established class-names baseline, while visually under-specified datasets use the established clean one-reference detector?
The router makes one score-blind, dataset-level decision:
class_names_only: use the exact saved class-names-only prediction branch;visual_references: use the exact saved one-positive-numeric-reference branch.
The routing transcript is never included in detection. Therefore, selecting
class_names_only cannot introduce the degradation observed when detection was
performed inside the adaptive conversation.
Paper Parts was selected because the earlier per-image adaptive router requested references for only 6 of 500 images, despite fixed one-shot prompting producing one of the largest RF20 gains.
The strict dataset-level router returned visual_references with confidence
0.95. It identified document-layout roles and annotation boundaries such as
equation number, figure caption, section, subsection, and paragraph as
requiring dataset-specific visual context.
| Selected detector branch | mAP50-95 | mAP50 | Delta vs. class names |
|---|---|---|---|
| Class names only | 14.30 | 34.95 | baseline |
| One positive numeric reference/class | 27.90 | 54.20 | +13.60 / +19.25 |
The router used 317 prompt tokens and 135 completion tokens with reasoning off, temperature 0, and seed 1234. At the same list prices used in the RF20 report, the routing call cost approximately $0.0014.
As a separate no-API diagnostic, the decisions from the completed adaptive run were reduced to a binary gate and replayed using the exact clean saved detector branches on all 20 datasets.
| Method | Macro mAP50-95 | Macro mAP50 | Delta vs. class names |
|---|---|---|---|
| Class names only | 24.37 | 43.54 | baseline |
| Fixed one-shot | 25.35 | 46.73 | +0.98 / +3.19 |
| Binary replay with isolated detectors | 25.48 | 46.20 | +1.11 / +2.66 |
This replay is diagnostic rather than a new prospective inference result, but it shows that isolating the detector removes the principal degradation. The Paper Parts live test additionally shows that conservative dataset-level wording can correct an important false zero-shot decision. Broader score-blind routing validation is still required before treating the router as a locked recipe.
Complete local artifact:
qwen38-fsod-runs/paper-parts-strict-binary-router-v1/result.json.
The locked router was then called three times sequentially for every RF20 dataset using identical temperature-0, seed-1234 settings. Majority vote chose between the same two clean saved detector branches. The router received only class names; it never received images, predictions, scores, or ground truth.
| Method | Macro mAP50-95 | Macro mAP50 | Estimated total cost |
|---|---|---|---|
| Class names only | 24.37 | 43.54 | $22.28 |
| Fixed one-shot | 25.35 | 46.73 | $34.20 |
| Strict binary routing | 25.55 | 46.91 | $32.64 |
Strict routing improved over class names by +1.18 / +3.37. It exceeded fixed one-shot by only +0.20 / +0.18, which is within the measured API noise and should be treated as a tie. The routing calls themselves cost $0.064; the selected detector branches cost $32.57.
The router selected visual references for 15 datasets and class names for five. All three calls agreed on 19/20 datasets. Soda Bottles was the only unstable case, with a 2-to-1 majority for class names.
Paper Parts was the development dataset used to refine the conservative prompt. Across the other 19 datasets, routing scored 25.42 / 46.53 and improved over their class-names baselines by +0.52 / +2.53. Paper Parts therefore accounts for a substantial portion of the full RF20 gain.
The router correctly preserved class names on Aerial Airport, Aquarium, FLIR, Soda Bottles, and Trail Camera. It captured large reference gains on Defect Detection, Orion, Paper Parts, Dreidel, and WB Prova. Important false-positive reference decisions remain: All Elements lost 7.59 / 0.65, Global Wheat Head lost 2.69 / 3.64, and Water Meter lost 16.55 / 29.58 relative to class names.
The result establishes that strict isolated routing can recover an RF20 uplift without the conversational degradation of the earlier adaptive method. It does not yet establish a meaningful advantage over always using one reference, because the accuracy difference is within noise and the router selected the more expensive branch for most datasets.
Complete local artifacts:
qwen38-fsod-runs/rf20-strict-binary-router-v1/summary.json and
qwen38-fsod-runs/rf20-strict-binary-router-v1/per_dataset.csv.
The class-name-only semantic router was subsequently replaced with a direct train-only A/B gate using the official RF20-VL-FSOD support objects. It corrected all three major false-reference decisions in a six-dataset diagnostic pilot. On the other 14 prospectively held-out datasets, it tied fixed one-shot within 0.03 mAP50-95 while selecting references only 5/14 times. The descriptive all-20 score was 26.67 / 47.79. See QWEN38_SUPPORT_CALIBRATED_ROUTER_RESULT.md.