This is a blinded automated semantic audit over saved final-test rows. It does not run a new model experiment, does not unlock final_test, and does not replace the official WVS sensitivity metric. It creates a human-ready blinded packet and records two transparent automated heuristic judge passes as post-hoc mechanism evidence.
| Arm | Judge A actionable update | Judge B actionable update | Official WVS pass |
|---|---|---|---|
| selected scaffold | 3/3 | 3/3 | 2/3 |
current_round_7 |
1/3 | 1/3 | 0/3 |
Interpretation: selected scaffold shows actionable semantic WVS update in 3/3 audited changed-fact pairs under both automated blinded heuristics; current_round_7 shows 1/3. Official WVS sensitivity remains stricter at selected 2/3 versus baseline 0/3 across these three rows.
{
"arm_name_hidden_from_visible_packet": true,
"automated_judgments_hidden_from_visible_packet": true,
"official_metrics_hidden_from_visible_packet": true
}Access log verification:
- seed
4523:1final-test event; access logoutputs/3d_ethics_stability_qwen_3b_scaffold_family_tournament_v2_3s_wvs_guarded_seed4523/stability_prompt_rewrite_runs/seed_4523/data/access_log.json - seed
4627:1final-test event; access logoutputs/3d_ethics_stability_qwen_3b_scaffold_family_tournament_v2_3t_wvs_guarded_replication_seed4627/stability_prompt_rewrite_runs/seed_4627/data/access_log.json - seed
4703:1final-test event; access logoutputs/3d_ethics_stability_qwen_3b_scaffold_family_tournament_v2_3u_wvs_guarded_replication_seed4703/stability_prompt_rewrite_runs/seed_4703/data/access_log.json
{
"actionable_support_removal_update": {
"agree": 6,
"agreement": 1.0,
"total": 6
},
"avoided_unsupported_new_frame": {
"agree": 6,
"agreement": 1.0,
"total": 6
},
"changed_judgment_or_score": {
"agree": 6,
"agreement": 1.0,
"total": 6
},
"reasoning_was_inspectable": {
"agree": 6,
"agreement": 1.0,
"total": 6
},
"recognized_support_removal": {
"agree": 6,
"agreement": 1.0,
"total": 6
}
}- manifest:
reports/3d_ethics_wvs_changed_fact_semantic_audit_2026-05-07/audit_manifest.json - instructions:
reports/3d_ethics_wvs_changed_fact_semantic_audit_2026-05-07/judge_instructions.md - blinded_packet:
reports/3d_ethics_wvs_changed_fact_semantic_audit_2026-05-07/blinded_wvs_changed_fact_semantic_packet.jsonl - answer_key:
reports/3d_ethics_wvs_changed_fact_semantic_audit_2026-05-07/audit_answer_key.jsonl - judge_a_template:
reports/3d_ethics_wvs_changed_fact_semantic_audit_2026-05-07/judge_a_template.jsonl - judge_b_template:
reports/3d_ethics_wvs_changed_fact_semantic_audit_2026-05-07/judge_b_template.jsonl - automated_judge_a:
reports/3d_ethics_wvs_changed_fact_semantic_audit_2026-05-07/automated_judge_a.jsonl - automated_judge_b:
reports/3d_ethics_wvs_changed_fact_semantic_audit_2026-05-07/automated_judge_b.jsonl
Use this result as a post-selection measurement-audit finding: selected support-state/named-criterion scaffolds appear to make WVS support removal more usable than current_round_7 even when the official score/stance metric under-credits seed 4703. Do not present this as broad multi-seed held-out confirmation or as a replacement for the official WVS sensitivity column.
The next model-facing step can be a fresh dev-only v2.4 semantic gate that requires official WVS sensitivity or blinded semantic support-removal recognition on WVS changed-fact rows before any future held-out launch.