{ "generated_at_utc": "2026-04-19T20:05:58+00:00", "runs_total": 34, "runs": [ { "run_id": "run-20260412T223658Z-rung_449_frontier_multib-qwen35_semparse_9b-42904", "scenario": "rung_449_frontier_multibind_uncle_query", "kb_name": "frontierv6_r1_rung_449_frontier_multibind_uncle_query", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 4, "validation_passed": 4, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:37:16+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_449_frontier_multibind_uncle_query_frontier_language_width_v6_20260412_222723.json", "report_json_rel": "kb_runs_published/rung_449_frontier_multibind_uncle_query_frontier_language_width_v6_20260412_222723.json", "report_html_rel": "reports/rung_449_frontier_multibind_uncle_query_frontier_language_width_v6_20260412_222723.html" }, { "run_id": "run-20260412T223612Z-rung_446_frontier_policy-qwen35_semparse_9b-23932", "scenario": "rung_446_frontier_policy_noisy_rebind_loop", "kb_name": "frontierv6_r1_rung_446_frontier_policy_noisy_rebind_loop", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 7, "validation_passed": 7, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:36:57+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_446_frontier_policy_noisy_rebind_loop_frontier_language_width_v6_20260412_222723.json", "report_json_rel": "kb_runs_published/rung_446_frontier_policy_noisy_rebind_loop_frontier_language_width_v6_20260412_222723.json", "report_html_rel": "reports/rung_446_frontier_policy_noisy_rebind_loop_frontier_language_width_v6_20260412_222723.html" }, { "run_id": "run-20260412T223545Z-rung_445_frontier_compou-qwen35_semparse_9b-46064", "scenario": "rung_445_frontier_compound_write_query_braid", "kb_name": "frontierv6_r1_rung_445_frontier_compound_write_query_braid", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 6, "validation_passed": 6, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:36:12+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_445_frontier_compound_write_query_braid_frontier_language_width_v6_20260412_222723.json", "report_json_rel": "kb_runs_published/rung_445_frontier_compound_write_query_braid_frontier_language_width_v6_20260412_222723.json", "report_html_rel": "reports/rung_445_frontier_compound_write_query_braid_frontier_language_width_v6_20260412_222723.html" }, { "run_id": "run-20260412T223515Z-rung_444_frontier_unpunc-qwen35_semparse_9b-40800", "scenario": "rung_444_frontier_unpunctuated_coref_sweep", "kb_name": "frontierv6_r1_rung_444_frontier_unpunctuated_coref_sweep", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 6, "validation_passed": 6, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:35:45+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_444_frontier_unpunctuated_coref_sweep_frontier_language_width_v6_20260412_222723.json", "report_json_rel": "kb_runs_published/rung_444_frontier_unpunctuated_coref_sweep_frontier_language_width_v6_20260412_222723.json", "report_html_rel": "reports/rung_444_frontier_unpunctuated_coref_sweep_frontier_language_width_v6_20260412_222723.html" }, { "run_id": "run-20260412T223423Z-rung_443_frontier_dual_i-qwen35_semparse_9b-37848", "scenario": "rung_443_frontier_dual_item_handoff_coref", "kb_name": "frontierv6_r1_rung_443_frontier_dual_item_handoff_coref", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 7, "validation_passed": 7, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:35:15+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_443_frontier_dual_item_handoff_coref_frontier_language_width_v6_20260412_222723.json", "report_json_rel": "kb_runs_published/rung_443_frontier_dual_item_handoff_coref_frontier_language_width_v6_20260412_222723.json", "report_html_rel": "reports/rung_443_frontier_dual_item_handoff_coref_frontier_language_width_v6_20260412_222723.html" }, { "run_id": "run-20260412T223328Z-rung_442_frontier_policy-qwen35_semparse_9b-45268", "scenario": "rung_442_frontier_policy_multirevision_guard", "kb_name": "frontierv6_r1_rung_442_frontier_policy_multirevision_guard", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 9, "validation_passed": 9, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:34:23+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_442_frontier_policy_multirevision_guard_frontier_language_width_v6_20260412_222723.json", "report_json_rel": "kb_runs_published/rung_442_frontier_policy_multirevision_guard_frontier_language_width_v6_20260412_222723.json", "report_html_rel": "reports/rung_442_frontier_policy_multirevision_guard_frontier_language_width_v6_20260412_222723.html" }, { "run_id": "run-20260412T223249Z-rung_441_frontier_pronou-qwen35_semparse_9b-45632", "scenario": "rung_441_frontier_pronoun_bucket_shuffle", "kb_name": "frontierv6_r1_rung_441_frontier_pronoun_bucket_shuffle", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 7, "validation_passed": 7, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:33:28+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_441_frontier_pronoun_bucket_shuffle_frontier_language_width_v6_20260412_222723.json", "report_json_rel": "kb_runs_published/rung_441_frontier_pronoun_bucket_shuffle_frontier_language_width_v6_20260412_222723.json", "report_html_rel": "reports/rung_441_frontier_pronoun_bucket_shuffle_frontier_language_width_v6_20260412_222723.html" }, { "run_id": "run-20260412T223205Z-rung_440_frontier_policy-qwen35_semparse_9b-46020", "scenario": "rung_440_frontier_policy_revision_loop", "kb_name": "frontierv6_r1_rung_440_frontier_policy_revision_loop", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 7, "validation_passed": 7, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:32:49+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_440_frontier_policy_revision_loop_frontier_language_width_v6_20260412_222723.json", "report_json_rel": "kb_runs_published/rung_440_frontier_policy_revision_loop_frontier_language_width_v6_20260412_222723.json", "report_html_rel": "reports/rung_440_frontier_policy_revision_loop_frontier_language_width_v6_20260412_222723.html" }, { "run_id": "run-20260412T223121Z-rung_439_frontier_plural-qwen35_semparse_9b-29424", "scenario": "rung_439_frontier_plural_coref_exception_guard", "kb_name": "frontierv6_r1_rung_439_frontier_plural_coref_exception_guard", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 5, "validation_passed": 5, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:32:04+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_439_frontier_plural_coref_exception_guard_frontier_language_width_v6_20260412_222723.json", "report_json_rel": "kb_runs_published/rung_439_frontier_plural_coref_exception_guard_frontier_language_width_v6_20260412_222723.json", "report_html_rel": "reports/rung_439_frontier_plural_coref_exception_guard_frontier_language_width_v6_20260412_222723.html" }, { "run_id": "run-20260412T223041Z-rung_438_frontier_multib-qwen35_semparse_9b-46244", "scenario": "rung_438_frontier_multibind_query_pressure", "kb_name": "frontierv6_r1_rung_438_frontier_multibind_query_pressure", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 6, "validation_passed": 6, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:31:21+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_438_frontier_multibind_query_pressure_frontier_language_width_v6_20260412_222723.json", "report_json_rel": "kb_runs_published/rung_438_frontier_multibind_query_pressure_frontier_language_width_v6_20260412_222723.json", "report_html_rel": "reports/rung_438_frontier_multibind_query_pressure_frontier_language_width_v6_20260412_222723.html" }, { "run_id": "run-20260412T222957Z-rung_437_frontier_policy-qwen35_semparse_9b-42932", "scenario": "rung_437_frontier_policy_override_flow", "kb_name": "frontierv6_r1_rung_437_frontier_policy_override_flow", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 7, "validation_passed": 7, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:30:41+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_437_frontier_policy_override_flow_frontier_language_width_v6_20260412_222723.json", "report_json_rel": "kb_runs_published/rung_437_frontier_policy_override_flow_frontier_language_width_v6_20260412_222723.json", "report_html_rel": "reports/rung_437_frontier_policy_override_flow_frontier_language_width_v6_20260412_222723.html" }, { "run_id": "run-20260412T222919Z-rung_436_frontier_noise_-qwen35_semparse_9b-46196", "scenario": "rung_436_frontier_noise_typo_coref", "kb_name": "frontierv6_r1_rung_436_frontier_noise_typo_coref", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 6, "validation_passed": 6, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:29:57+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_436_frontier_noise_typo_coref_frontier_language_width_v6_20260412_222723.json", "report_json_rel": "kb_runs_published/rung_436_frontier_noise_typo_coref_frontier_language_width_v6_20260412_222723.json", "report_html_rel": "reports/rung_436_frontier_noise_typo_coref_frontier_language_width_v6_20260412_222723.html" }, { "run_id": "run-20260412T222837Z-rung_435_frontier_checkp-qwen35_semparse_9b-34112", "scenario": "rung_435_frontier_checkpoint_compound_turns", "kb_name": "frontierv6_r1_rung_435_frontier_checkpoint_compound_turns", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 6, "validation_passed": 6, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:29:19+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_435_frontier_checkpoint_compound_turns_frontier_language_width_v6_20260412_222723.json", "report_json_rel": "kb_runs_published/rung_435_frontier_checkpoint_compound_turns_frontier_language_width_v6_20260412_222723.json", "report_html_rel": "reports/rung_435_frontier_checkpoint_compound_turns_frontier_language_width_v6_20260412_222723.html" }, { "run_id": "run-20260412T222813Z-rung_434_dual_pronoun_fl-qwen35_semparse_9b-46892", "scenario": "rung_434_dual_pronoun_flip_guard", "kb_name": "frontierv6_r1_rung_434_dual_pronoun_flip_guard", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 7, "validation_passed": 7, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:28:37+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_434_dual_pronoun_flip_guard_frontier_language_width_v6_20260412_222723.json", "report_json_rel": "kb_runs_published/rung_434_dual_pronoun_flip_guard_frontier_language_width_v6_20260412_222723.json", "report_html_rel": "reports/rung_434_dual_pronoun_flip_guard_frontier_language_width_v6_20260412_222723.html" }, { "run_id": "run-20260412T222745Z-rung_433_noisy_inverse_r-qwen35_semparse_9b-31104", "scenario": "rung_433_noisy_inverse_retarget_repair", "kb_name": "frontierv6_r1_rung_433_noisy_inverse_retarget_repair", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 6, "validation_passed": 6, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:28:13+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_433_noisy_inverse_retarget_repair_frontier_language_width_v6_20260412_222723.json", "report_json_rel": "kb_runs_published/rung_433_noisy_inverse_retarget_repair_frontier_language_width_v6_20260412_222723.json", "report_html_rel": "reports/rung_433_noisy_inverse_retarget_repair_frontier_language_width_v6_20260412_222723.html" }, { "run_id": "run-20260412T222724Z-rung_432_noise_pronoun_i-qwen35_semparse_9b-39528", "scenario": "rung_432_noise_pronoun_inversion_chain", "kb_name": "frontierv6_r1_rung_432_noise_pronoun_inversion_chain", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 6, "validation_passed": 6, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:27:45+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_432_noise_pronoun_inversion_chain_frontier_language_width_v6_20260412_222723.json", "report_json_rel": "kb_runs_published/rung_432_noise_pronoun_inversion_chain_frontier_language_width_v6_20260412_222723.json", "report_html_rel": "reports/rung_432_noise_pronoun_inversion_chain_frontier_language_width_v6_20260412_222723.html" }, { "run_id": "run-20260412T222457Z-rung_448_confirmation_ga-qwen35_semparse_9b-27556", "scenario": "rung_448_confirmation_gate_no_then_yes", "kb_name": "confprobe_r4_rung_448_confirmation_gate_no_then_yes", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 3, "validation_passed": 3, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:25:10+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_448_confirmation_gate_no_then_yes_frontier_confirmation_probe_v1_20260412_222449.json", "report_json_rel": "kb_runs_published/rung_448_confirmation_gate_no_then_yes_frontier_confirmation_probe_v1_20260412_222449.json", "report_html_rel": "reports/rung_448_confirmation_gate_no_then_yes_frontier_confirmation_probe_v1_20260412_222449.html" }, { "run_id": "run-20260412T222449Z-rung_447_confirmation_ga-qwen35_semparse_9b-36828", "scenario": "rung_447_confirmation_gate_single_yes", "kb_name": "confprobe_r4_rung_447_confirmation_gate_single_yes", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 2, "validation_passed": 2, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:24:57+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_447_confirmation_gate_single_yes_frontier_confirmation_probe_v1_20260412_222449.json", "report_json_rel": "kb_runs_published/rung_447_confirmation_gate_single_yes_frontier_confirmation_probe_v1_20260412_222449.json", "report_html_rel": "reports/rung_447_confirmation_gate_single_yes_frontier_confirmation_probe_v1_20260412_222449.html" }, { "run_id": "run-20260412T221940Z-rung_446_frontier_policy-qwen35_semparse_9b-13876", "scenario": "rung_446_frontier_policy_noisy_rebind_loop", "kb_name": "ceprobe_r1_rung_446_frontier_policy_noisy_rebind_loop", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 7, "validation_passed": 7, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:20:25+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_446_frontier_policy_noisy_rebind_loop_frontier_clarification_probe_v1_20260412_221839.json", "report_json_rel": "kb_runs_published/rung_446_frontier_policy_noisy_rebind_loop_frontier_clarification_probe_v1_20260412_221839.json", "report_html_rel": "reports/rung_446_frontier_policy_noisy_rebind_loop_frontier_clarification_probe_v1_20260412_221839.html" }, { "run_id": "run-20260412T221913Z-rung_445_frontier_compou-qwen35_semparse_9b-31308", "scenario": "rung_445_frontier_compound_write_query_braid", "kb_name": "ceprobe_r1_rung_445_frontier_compound_write_query_braid", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 6, "validation_passed": 6, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:19:39+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_445_frontier_compound_write_query_braid_frontier_clarification_probe_v1_20260412_221839.json", "report_json_rel": "kb_runs_published/rung_445_frontier_compound_write_query_braid_frontier_clarification_probe_v1_20260412_221839.json", "report_html_rel": "reports/rung_445_frontier_compound_write_query_braid_frontier_clarification_probe_v1_20260412_221839.html" }, { "run_id": "run-20260412T221839Z-rung_444_frontier_unpunc-qwen35_semparse_9b-29988", "scenario": "rung_444_frontier_unpunctuated_coref_sweep", "kb_name": "ceprobe_r1_rung_444_frontier_unpunctuated_coref_sweep", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 6, "validation_passed": 6, "validation_rate": 1.0, "finished_utc": "2026-04-12T22:19:13+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_444_frontier_unpunctuated_coref_sweep_frontier_clarification_probe_v1_20260412_221839.json", "report_json_rel": "kb_runs_published/rung_444_frontier_unpunctuated_coref_sweep_frontier_clarification_probe_v1_20260412_221839.json", "report_html_rel": "reports/rung_444_frontier_unpunctuated_coref_sweep_frontier_clarification_probe_v1_20260412_221839.html" }, { "run_id": "run-20260412T130439Z-demo_04_reimbursement_vi-qwen35_semparse_9b-44172", "scenario": "demo_04_reimbursement_violation_check", "kb_name": "demo_04_reimbursement_violation_check", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 3, "validation_passed": 3, "validation_rate": 1.0, "finished_utc": "2026-04-12T13:05:10+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/demo_04_reimbursement_violation_check.json", "report_json_rel": "kb_runs_published/demo_04_reimbursement_violation_check.json", "report_html_rel": "reports/demo_04_reimbursement_violation_check.html" }, { "run_id": "run-20260412T130421Z-demo_03_story_world_inte-qwen35_semparse_9b-19956", "scenario": "demo_03_story_world_interrogator", "kb_name": "demo_03_story_world_interrogator", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "failed", "validation_total": 3, "validation_passed": 0, "validation_rate": 0.0, "finished_utc": "2026-04-12T13:04:39+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/demo_03_story_world_interrogator.json", "report_json_rel": "kb_runs_published/demo_03_story_world_interrogator.json", "report_html_rel": "reports/demo_03_story_world_interrogator.html" }, { "run_id": "run-20260412T130357Z-demo_02_policy_stress_te-qwen35_semparse_9b-26940", "scenario": "demo_02_policy_stress_test_machine", "kb_name": "demo_02_policy_stress_test_machine", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 2, "validation_passed": 2, "validation_rate": 1.0, "finished_utc": "2026-04-12T13:04:21+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/demo_02_policy_stress_test_machine.json", "report_json_rel": "kb_runs_published/demo_02_policy_stress_test_machine.json", "report_html_rel": "reports/demo_02_policy_stress_test_machine.html" }, { "run_id": "run-20260412T102202Z-rung_360_ce_story_branch-qwen35_semparse_9b-41352", "scenario": "rung_360_ce_story_branch_merge_noise", "kb_name": "rung_360_ce_story_branch_merge_noise", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 12, "validation_passed": 12, "validation_rate": 1.0, "finished_utc": "2026-04-12T10:22:39+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_360_ce_story_branch_merge_noise_frontier_ce_340_360_refined3_check_20260412.json", "report_json_rel": "kb_runs_published/rung_360_ce_story_branch_merge_noise_frontier_ce_340_360_refined3_check_20260412.json", "report_html_rel": "reports/rung_360_ce_story_branch_merge_noise_frontier_ce_340_360_refined3_check_20260412.html" }, { "run_id": "run-20260412T102113Z-rung_350_ce_story_multi_-qwen35_semparse_9b-35144", "scenario": "rung_350_ce_story_multi_round_revision", "kb_name": "rung_350_ce_story_multi_round_revision", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 11, "validation_passed": 11, "validation_rate": 1.0, "finished_utc": "2026-04-12T10:22:02+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_350_ce_story_multi_round_revision_frontier_ce_340_360_refined3_check_20260412.json", "report_json_rel": "kb_runs_published/rung_350_ce_story_multi_round_revision_frontier_ce_340_360_refined3_check_20260412.json", "report_html_rel": "reports/rung_350_ce_story_multi_round_revision_frontier_ce_340_360_refined3_check_20260412.html" }, { "run_id": "run-20260412T102015Z-rung_340_ce_story_pronou-qwen35_semparse_9b-42296", "scenario": "rung_340_ce_story_pronoun_transfer", "kb_name": "rung_340_ce_story_pronoun_transfer", "backend": "ollama", "model": "qwen35-semparse:9b", "status": "passed", "validation_total": 11, "validation_passed": 11, "validation_rate": 1.0, "finished_utc": "2026-04-12T10:21:13+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_340_ce_story_pronoun_transfer_frontier_ce_340_360_refined3_check_20260412.json", "report_json_rel": "kb_runs_published/rung_340_ce_story_pronoun_transfer_frontier_ce_340_360_refined3_check_20260412.json", "report_html_rel": "reports/rung_340_ce_story_pronoun_transfer_frontier_ce_340_360_refined3_check_20260412.html" }, { "run_id": "run-20260412T022335Z-rung_230_fuzzy_ce_branch-qwen3_5_9b-44308", "scenario": "rung_230_fuzzy_ce_branch_exclusion_language", "kb_name": "rung_230_fuzzy_ce_branch_exclusion_language", "backend": "ollama", "model": "qwen3.5:9b", "status": "passed", "validation_total": 6, "validation_passed": 6, "validation_rate": 1.0, "finished_utc": "2026-04-12T02:23:53+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/rung_230_fuzzy_ce_branch_exclusion_language_stability_230_after_staysfix_20260412.json", "report_json_rel": "kb_runs_published/rung_230_fuzzy_ce_branch_exclusion_language_stability_230_after_staysfix_20260412.json", "report_html_rel": "reports/rung_230_fuzzy_ce_branch_exclusion_language_stability_230_after_staysfix_20260412.html" }, { "run_id": "run-20260411T235210Z-acid_05_long_context_lin-qwen3_5_9b-14228", "scenario": "acid_05_long_context_lineage", "kb_name": "acid_lineage", "backend": "ollama", "model": "qwen3.5:9b", "status": "passed", "validation_total": 5, "validation_passed": 5, "validation_rate": 1.0, "finished_utc": "2026-04-11T23:52:39+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/acid_05_long_context_lineage_auto_cycle_stage1_to_rung200_clean_20260411.json", "report_json_rel": "kb_runs_published/acid_05_long_context_lineage_auto_cycle_stage1_to_rung200_clean_20260411.json", "report_html_rel": "reports/acid_05_long_context_lineage_auto_cycle_stage1_to_rung200_clean_20260411.html" }, { "run_id": "run-20260411T235155Z-acid_04_alias_pressure-qwen3_5_9b-12656", "scenario": "acid_04_alias_pressure", "kb_name": "acid_alias", "backend": "ollama", "model": "qwen3.5:9b", "status": "passed", "validation_total": 3, "validation_passed": 3, "validation_rate": 1.0, "finished_utc": "2026-04-11T23:52:09+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/acid_04_alias_pressure_auto_cycle_stage1_to_rung200_clean_20260411.json", "report_json_rel": "kb_runs_published/acid_04_alias_pressure_auto_cycle_stage1_to_rung200_clean_20260411.json", "report_html_rel": "reports/acid_04_alias_pressure_auto_cycle_stage1_to_rung200_clean_20260411.html" }, { "run_id": "run-20260411T235139Z-acid_03_temporal_overrid-qwen3_5_9b-10212", "scenario": "acid_03_temporal_override", "kb_name": "acid_temporal", "backend": "ollama", "model": "qwen3.5:9b", "status": "passed", "validation_total": 3, "validation_passed": 3, "validation_rate": 1.0, "finished_utc": "2026-04-11T23:51:54+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/acid_03_temporal_override_auto_cycle_stage1_to_rung200_clean_20260411.json", "report_json_rel": "kb_runs_published/acid_03_temporal_override_auto_cycle_stage1_to_rung200_clean_20260411.json", "report_html_rel": "reports/acid_03_temporal_override_auto_cycle_stage1_to_rung200_clean_20260411.html" }, { "run_id": "run-20260411T235133Z-stage_03_transitive_chai-qwen3_5_9b-29012", "scenario": "stage_03_transitive_chain", "kb_name": "people_stage", "backend": "ollama", "model": "qwen3.5:9b", "status": "passed", "validation_total": 1, "validation_passed": 1, "validation_rate": 1.0, "finished_utc": "2026-04-11T23:51:39+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/stage_03_transitive_chain_auto_cycle_stage1_to_rung200_clean_20260411.json", "report_json_rel": "kb_runs_published/stage_03_transitive_chain_auto_cycle_stage1_to_rung200_clean_20260411.json", "report_html_rel": "reports/stage_03_transitive_chain_auto_cycle_stage1_to_rung200_clean_20260411.html" }, { "run_id": "run-20260411T235129Z-stage_02_rule_ingest-qwen3_5_9b-39620", "scenario": "stage_02_rule_ingest", "kb_name": "people_stage", "backend": "ollama", "model": "qwen3.5:9b", "status": "passed", "validation_total": 1, "validation_passed": 1, "validation_rate": 1.0, "finished_utc": "2026-04-11T23:51:33+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/stage_02_rule_ingest_auto_cycle_stage1_to_rung200_clean_20260411.json", "report_json_rel": "kb_runs_published/stage_02_rule_ingest_auto_cycle_stage1_to_rung200_clean_20260411.json", "report_html_rel": "reports/stage_02_rule_ingest_auto_cycle_stage1_to_rung200_clean_20260411.html" }, { "run_id": "run-20260411T235126Z-stage_01_facts_only-qwen3_5_9b-26908", "scenario": "stage_01_facts_only", "kb_name": "people_stage", "backend": "ollama", "model": "qwen3.5:9b", "status": "passed", "validation_total": 2, "validation_passed": 2, "validation_rate": 1.0, "finished_utc": "2026-04-11T23:51:29+00:00", "prompt_id": "sp-1e43c641b01b", "prompt_sha256": "1e43c641b01b7c845b82331b521d58c1993e8010bd9283f5085dac687520159e", "prompt_snapshot_rel": "prompts/sp-1e43c641b01b.md", "run_json_source_rel": "kb_runs_published/stage_01_facts_only_auto_cycle_stage1_to_rung200_clean_20260411.json", "report_json_rel": "kb_runs_published/stage_01_facts_only_auto_cycle_stage1_to_rung200_clean_20260411.json", "report_html_rel": "reports/stage_01_facts_only_auto_cycle_stage1_to_rung200_clean_20260411.html" } ] }