| # Frozen-prediction robustness analyses |
|
|
| Post-hoc exclusion sensitivity, not new ground truth or retraining. IoU 0.5; 10,000 paired whole-cluster resamples. All predictions and rules are frozen for this analysis. |
|
|
| | subset | pages | base clusters | ProofJudge F1 | Argon F1 | delta 95% CI | |
| |---|---:|---:|---:|---:|---| |
| | unfiltered | 2871 | 486 | 0.478 | 0.360 | [+0.088, +0.148] | |
| | published | 2765 | 469 | 0.481 | 0.372 | [+0.079, +0.140] | |
| | audit_excluded | 2680 | 460 | 0.489 | 0.377 | [+0.080, +0.141] | |
| | structural_quarantine | 2581 | 460 | 0.506 | 0.395 | [+0.078, +0.143] | |
| | convention_sensitivity | 1766 | 457 | 0.627 | 0.397 | [+0.187, +0.271] | |
| |
| Exclusions are cumulative after published: author-audit exclusions; drop/reorder; component transplants and local-icon. Entire records are quarantined, never relabelled clean. The last subset changes task coverage and is not a replacement benchmark. |
| |
| Base clusters merge source generator + base page across general/local/multi/negative variants; requirement clusters further merge generator/design language/theme. Real pages use URL/page identity, then hostname/app. Neither grouping corrects test-informed selection. |
| |
| ## Published-pool diagnostics |
| |
| | model | F1 | clean flagged | scoring errors | normalized kind right / matched known | aspect right / matched known | |
| |---|---:|---:|---:|---|---| |
| | rolejudge-v7f-r128soup-cut | 0.481 | 7.4% | 0 | 839/848 | 267/304 | |
| | gemini-4-argon | 0.372 | 17.8% | 0 | 709/853 | 282/504 | |
| | claude-opus-5-5-medium | 0.308 | 78.3% | 0 | 898/1064 | 283/611 | |
| | gpt-6.1-sol | 0.300 | 36.6% | 3 | 710/805 | 196/395 | |
| | rolejudge-v7f-r128soup | 0.455 | 51.6% | 0 | 1038/1058 | 385/443 | |
| | rolejudge-v7f-2500-cut | 0.405 | 6.9% | 0 | 679/688 | 182/223 | |
| | qwen3.5-9b-base | 0.043 | 87.8% | 282 | 428/603 | 214/408 | |
| |
| Kind maps the consistency-prompt label inconsistent to the stored label brand; no other alias is mapped. Kind/aspect accuracy is conditional on matched boxes with known labels. It is not explanation correctness. Errors follow the existing scorer (non-dict answer), not a strict schema validator. |
| |
| ## Requirement-cluster intervals |
| |
| { |
| "clusters": 121, |
| "min_size": 1, |
| "max_size": 414, |
| "vs": { |
| "gemini-4-argon": { |
| "delta": 0.10899280004942724, |
| "ci95": [ |
| 0.04467793849992695, |
| 0.17696848361181572 |
| ] |
| }, |
| "claude-opus-5-5-medium": { |
| "delta": 0.1730305063027215, |
| "ci95": [ |
| 0.14083593515714832, |
| 0.20652628574165327 |
| ] |
| }, |
| "gpt-6.1-sol": { |
| "delta": 0.1808314426852562, |
| "ci95": [ |
| 0.13080445467071503, |
| 0.22946292326254902 |
| ] |
| }, |
| "rolejudge-v7f-r128soup": { |
| "delta": 0.02514385754245596, |
| "ci95": [ |
| 0.005967370588310595, |
| 0.04313225869484571 |
| ] |
| }, |
| "rolejudge-v7f-2500-cut": { |
| "delta": 0.07576445358264244, |
| "ci95": [ |
| 0.03961699314203121, |
| 0.10928092511103336 |
| ] |
| }, |
| "qwen3.5-9b-base": { |
| "delta": 0.4375702111145328, |
| "ci95": [ |
| 0.3703972864951111, |
| 0.4854255803273521 |
| ] |
| } |
| } |
| } |
| |
| ## Public reproduction |
| |
| Run `python robustness.py --data-dir . --output recomputed.json` (Python + NumPy). `records.jsonl` contains per-record sufficient statistics for frozen predictions, opaque cluster IDs and exclusion flags. Original release IDs (`ct-*`) are preserved. The 106 `excluded-*` records have counts only; their screenshots are not added to the original dataset. This reproduces statistical reanalysis, not model inference. The original Parquet splits and their labels remain unchanged. |
| |
| The historical implementation recorded in results.json belongs to the paper workspace; the portable command above requires only files in this release. No independent blind human validation or new training-data ablation is claimed. |
| |
