Datasets:

Biogod
/

mas_reasoning

Adds deterministic raw-output shards, analysis-ready inputs, provenance-labelled code snapshots, and rebuilt global integrity ledgers for the historical medical MAS evidence.

This view is limited to 50 files because it contains too many changes. See raw diff

Files changed (50) hide show

CHECKSUMS.sha256 +172 -11
DATA_MANIFEST.csv +170 -9
RELEASE_METADATA.json +8 -3
data/external/results/formal/early_mas_architecture_motivation/README.md +38 -2
data/external/results/formal/early_mas_architecture_motivation/analysis_ready/configs/med_board/real_llm_full_seed42.yaml +33 -0
data/external/results/formal/early_mas_architecture_motivation/analysis_ready/configs/med_board/real_llm_full_seed43.yaml +33 -0
data/external/results/formal/early_mas_architecture_motivation/analysis_ready/configs/med_board/real_llm_full_seed44.yaml +33 -0
data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/baseline/all_results.csv +0 -0
data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/baseline/summary.json +89 -0
data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/fixed6/all_results.csv +0 -0
data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/fixed6/summary.json +82 -0
data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/tuned/all_results.csv +0 -0
data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/tuned/summary.json +90 -0
data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_board/gain_bottleneck_report.json +681 -0
data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_board/legacy_3seed_summary.md +79 -0
data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_board/oracle_trials.csv +0 -0
data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_board/protocol_results.csv +0 -0
data/external/results/formal/early_mas_architecture_motivation/audits/code_security_gate.md +27 -0
data/external/results/formal/early_mas_architecture_motivation/audits/code_snapshot_security_scan.json +13 -0
data/external/results/formal/early_mas_architecture_motivation/audits/code_snapshot_security_scan.metadata.json +17 -0
data/external/results/formal/early_mas_architecture_motivation/audits/payload_security_scan.json +18 -0
data/external/results/formal/early_mas_architecture_motivation/audits/payload_security_scan.metadata.json +17 -0
data/external/results/formal/early_mas_architecture_motivation/audits/private_release_verification.json +13 -0
data/external/results/formal/early_mas_architecture_motivation/audits/private_release_verification.metadata.json +19 -0
data/external/results/formal/early_mas_architecture_motivation/audits/provenance_and_qa.md +19 -21
data/external/results/formal/early_mas_architecture_motivation/audits/provenance_and_qa.metadata.json +29 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/pyproject.toml +15 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/pytest.ini +5 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/__init__.py +1 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/__main__.py +194 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/agent.py +422 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/config.py +52 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/data/__init__.py +1 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/data/loader.py +51 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/metrics/__init__.py +1 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/metrics/analyzer.py +89 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/metrics/collector.py +38 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/runner.py +404 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/__init__.py +17 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/base.py +9 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/complete.py +12 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/hierarchical.py +36 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/ring.py +24 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/single.py +8 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/star.py +23 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/types.py +92 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/__init__.py +0 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/conftest.py +46 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_agent.py +218 -0
data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_analyze_ab_details.py +184 -0

CHECKSUMS.sha256 CHANGED Viewed

@@ -1,7 +1,7 @@
-24528d4cf006fe67f8ec6c2864480eb592c55335c4beae535a2e3f1b43d84674  DATA_MANIFEST.csv
 40b0edb03973b3e422e88359d0e5c5c61cd1cb40310c2a96e58467d7b01ce796  EXCLUDED_FILES.csv
 6ed02104f3215c59e03dcba1bb0c87a2e684b881d82acf6a1fafd1dc1ecaab69  README.md
-551a59a0ff293560c1361ae894f4dff2d16c8a23b93f1ceac767824205bc40f8  RELEASE_METADATA.json
 b8677fbe2da2f2dcd8d4d941b4e953e05caf6e27338db67e65aa2615aebd3429  SOURCE_SYMLINKS.csv
 3edbd31d6821ad7f499519d5516ab28264572352b404fbbdae9cfc6cbe72276a  THIRD_PARTY_NOTICES.md
 f92e97647aa426aec2b2fa08c830cb771aa840f0b6b6eafbffc37c146e2d7cf9  data/derived/pool_setting_k16_vote_structure_summary.csv
@@ -103,18 +103,179 @@ e66c021a3c18857ec51c7c35602a8d81ef363ee9e618095b8452d788b970d3f2  data/external/
 0998c2022173143eb97b05445d44b4e7f514ab95ebf603362551f5a3159b54e8  data/external/evidence/formal/prospective_diffusion_validation/sample_plan.metadata.json
 4a06d9e8fbad3d96b29d6274aafd621bd7283f8846a1c53cc269ed4eb0a596fe  data/external/evidence/formal/prospective_diffusion_validation/source_paths.md
 0ef16575b487cfe58527a1dcd2e78fcff1e0e6dda8241985d9f63ced392b895b  data/external/results/formal/README.md
-da10f260bbef39a5a5f6f34491852dc18dcd7b9fa2c3f5a4f07a914158270b60  data/external/results/formal/early_mas_architecture_motivation/README.md
-53bf2d7b4ab9c562099640bd74899c01dbc006241c17b4bb5d952ebdbb253941  data/external/results/formal/early_mas_architecture_motivation/audits/provenance_and_qa.md
-727e2600b0f8b1dc5bd8cd72849657097dc339641e5fb5920cf3cee28b388bed  data/external/results/formal/early_mas_architecture_motivation/figure_sources/extended_data_fig1_source.csv
-271a9c53f41ceffe49d2b59b4ea40fafb274be2fda4896d9f154d688ba449e71  data/external/results/formal/early_mas_architecture_motivation/source_manifest.csv
 1ea792219dde6450744dbb7ee4b6b10d2881f392d3401f3a3ec109ca5d9b9a79  data/external/results/formal/early_mas_architecture_motivation/summaries/med_agent_consensus_bypass_summary.csv
-d6961f0d2688b5193edea4184db9ff6a5c84979aafad92a077729bdce0c780e4  data/external/results/formal/early_mas_architecture_motivation/summaries/med_agent_round_exit_distribution.csv
-032b83a23659f5572a5faddf004a8f51297deb4ea108309c9bb77885326988d6  data/external/results/formal/early_mas_architecture_motivation/summaries/med_agent_topology_summary.csv
 dd2119db18e7d83e9e8efd5b945aa75c9ec2022210b43cb1f03d6a36bea5cdf8  data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_benchmark_discussion_controls.csv
-19f42ed4e2cf8e726c2e73b724b3f10b0ec4dfff0515263e9f780334c36e0abe  data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_judge_behavior_summary.csv
 7dd4aff44b730a7295ab79107164868f884ff9c9b2bd1353b62d9c1a258afe78  data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_mechanism_pilot_summary.csv
-f7842a093ae110ebfbd1ef477ac261d9db092e3164c3a6fe0fc983fab853316c  data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_oracle_gap_summary.csv
-9496bc56850aeb1b5d1bcd8520f6c613c1902843e3c3e9b043151fba1cfdae44  data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_protocol_cost_accuracy_summary.csv
 d871b3cabeb8a103203b5cf66be2e8a34f4d2829652a2498b6b8a4c4c5436c51  data/external/results/formal/five_benchmark_cost_efficiency/README.md
 f0c8b7ee3fb57c984b9caee42d68f58f126f2faf7b4d661a443cb1354de3227d  data/external/results/formal/five_benchmark_cost_efficiency/archive/source_inputs/mimo_k16_topup_bank/mimo_k16_topup.launchd.log
 adde7549dc8ef3f9954df91107df5cff37d630b99ab6a2fa4e332a101a5be0a1  data/external/results/formal/five_benchmark_cost_efficiency/archive/source_inputs/mimo_k16_topup_bank/mimo_k16_topup.launchd.status

+dd5d3af8f027b9cbd8f098586ebdff865ac816745b9432fe942ed5d9361390ac  DATA_MANIFEST.csv
 40b0edb03973b3e422e88359d0e5c5c61cd1cb40310c2a96e58467d7b01ce796  EXCLUDED_FILES.csv
 6ed02104f3215c59e03dcba1bb0c87a2e684b881d82acf6a1fafd1dc1ecaab69  README.md
+cf5aaea1850ddb23238eb3c3eeb2386f3f736ef903e65e4b3c64d23d05d84633  RELEASE_METADATA.json
 b8677fbe2da2f2dcd8d4d941b4e953e05caf6e27338db67e65aa2615aebd3429  SOURCE_SYMLINKS.csv
 3edbd31d6821ad7f499519d5516ab28264572352b404fbbdae9cfc6cbe72276a  THIRD_PARTY_NOTICES.md
 f92e97647aa426aec2b2fa08c830cb771aa840f0b6b6eafbffc37c146e2d7cf9  data/derived/pool_setting_k16_vote_structure_summary.csv
 0998c2022173143eb97b05445d44b4e7f514ab95ebf603362551f5a3159b54e8  data/external/evidence/formal/prospective_diffusion_validation/sample_plan.metadata.json
 4a06d9e8fbad3d96b29d6274aafd621bd7283f8846a1c53cc269ed4eb0a596fe  data/external/evidence/formal/prospective_diffusion_validation/source_paths.md
 0ef16575b487cfe58527a1dcd2e78fcff1e0e6dda8241985d9f63ced392b895b  data/external/results/formal/README.md
+21ce9218073aee47e04594de2386e8a3fd0e4a6f48ca5b7c97fd2d13a7beb0d1  data/external/results/formal/early_mas_architecture_motivation/README.md
+4aeb065fb9fb7de6521c557b73a7e8aef4cf360736d68e4c6a786ec7ddd6607f  data/external/results/formal/early_mas_architecture_motivation/analysis_ready/configs/med_board/real_llm_full_seed42.yaml
+7ff5fc11538def5c077309281e4a544897b3d27ba39a33fffd403cd853b9765a  data/external/results/formal/early_mas_architecture_motivation/analysis_ready/configs/med_board/real_llm_full_seed43.yaml
+e839ac0c62c5a149d1b99d237c7d8c73b98037f1ea9e81931c1c4e998d237d19  data/external/results/formal/early_mas_architecture_motivation/analysis_ready/configs/med_board/real_llm_full_seed44.yaml
+1d62980ea5a0d6fee72bf94f408a60be27b99995985963395e51e18aeea7d675  data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/baseline/all_results.csv
+2c0bdec413c879684db2528c4eb37d1cb7f8f366d49aacadedcd5bbc931eef16  data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/baseline/summary.json
+047cab62f4831e320516d14a5554cf3bd7f595004fa4400da84955bd5627675e  data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/fixed6/all_results.csv
+c4013cb3ad5255f07555f99deb4b96bb0d2202e260071d4fc0a660f98cb33ebe  data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/fixed6/summary.json
+c8d553568735f5c1bbe3f79ca280312090d448594709bace7fa77a21d15d86a4  data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/tuned/all_results.csv
+ca4122852b29fb9549c17bb08a22f86358eda2e47d323579dd87c433ff368bbb  data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/tuned/summary.json
+054b69cb30f3eac151d40e29cda7e4970705f288b1ca6a4c586ad32e33719acd  data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_board/gain_bottleneck_report.json
+d64229bd5fe5475527b6f5cd6719a3da09ff1bf1e086513d2ae248b6dbe0f15a  data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_board/legacy_3seed_summary.md
+de275959a071640633f446a344608fb84e57a49294b2ab956179de129f7e1638  data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_board/oracle_trials.csv
+b9d074ef8eb51315ceee82bb63dbc32da4926f42d8b7c04e3d16bdc74c3298e4  data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_board/protocol_results.csv
+d382e8fda4a21bd54aeb7f975b3e2df690f09f5fbb3a73d90ef22145907fcee9  data/external/results/formal/early_mas_architecture_motivation/audits/code_security_gate.md
+62aeedae6f9c813a65a4de47f626196a4a0aa69d8ca6729601fd263192b8f970  data/external/results/formal/early_mas_architecture_motivation/audits/code_snapshot_security_scan.json
+c052ffdf2024b5024b78cb61fca415b8e52da72300407b9aac335e31ec1b85c8  data/external/results/formal/early_mas_architecture_motivation/audits/code_snapshot_security_scan.metadata.json
+d3e82e115abddff2d6151381f0d44e52484471579f9c44629e7000754ad4afe1  data/external/results/formal/early_mas_architecture_motivation/audits/payload_security_scan.json
+5c018829edb980f1e488e746fc680b1b9254d97f39f50f596e4dbf1fd98e6327  data/external/results/formal/early_mas_architecture_motivation/audits/payload_security_scan.metadata.json
+0e19cf04a8030280e26eb239651ba47dd5ccb1073835f0cee9566fcc8e214639  data/external/results/formal/early_mas_architecture_motivation/audits/private_release_verification.json
+94d6c3ad4f5f9b662a028cf0cec93c56a5b5070c5140f7c1dccbcb459b70d61d  data/external/results/formal/early_mas_architecture_motivation/audits/private_release_verification.metadata.json
+cea540e96b20147b86dd2791dacf60ba29504d66c55e0300bc975e284420ae22  data/external/results/formal/early_mas_architecture_motivation/audits/provenance_and_qa.md
+b0ba46b78b85d8e74fb7e47ed59ec655db9a40284dcc502c6e5eb5095a7e4c6f  data/external/results/formal/early_mas_architecture_motivation/audits/provenance_and_qa.metadata.json
+2a52f22d56d2bccb87573b9e349497ae5f057b8108ffa8558b19f29ec1342ec0  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/pyproject.toml
+b769e070eb76e69b821b68d31bcf20e850ec74eda6491c664d04184e9ce1691b  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/pytest.ini
+222a8f2c7f63d7e3d4a5165f515111016b552c563a94420adb3fe23cd4e961fd  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/__init__.py
+da3c833adaea421e012e1c6c57c3d919035430ef0a4e30e97e458fc84761788b  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/__main__.py
+428357d3921165e377dfafb5b88dc79899e2ae69a911fe5460a3af6b9e1cae18  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/agent.py
+f67c2acb0b1bf49b613e977a189c968a6d9d0aa6630d5576753bafc7557a9e70  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/config.py
+c0bf0688d03ec00e93160193b9cc13960c99c05b90167e72f12d16b5fd5477b3  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/data/__init__.py
+247dace738280caae0eea745349fa02eb10f92106147ae6b4d988bc9968fc2ba  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/data/loader.py
+cca37fb69099489cbaed5ba0d9fd71636ece6a14313ce9946d9d33cd562a84f7  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/metrics/__init__.py
+81285eeb497531b129a5ba38d15c8059cf911b3e8e27aa587bb2bbf69c1f7784  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/metrics/analyzer.py
+21096979d1ef7087f23011e71cd2b211b2a9584ee9c18b12d8faa929c71790c1  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/metrics/collector.py
+fd9b29bac159a713df19ebadba81eb52e165447ceb76d197b93b1c2edb91b15f  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/runner.py
+8124df00004e9016f10c5b65d01bf3cfaca4f1dc5c63828a6b25f79f406c711f  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/__init__.py
+8de0d4d54010d988ac00f13fafe74a50bd2ecffc5c5a2035a4f6569d0ba11cf3  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/base.py
+dc160e6f7425e7d4e330feca841a66e0b0f11400c089fe12adeba29338536be8  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/complete.py
+afc8676ed6467bbf02e0c9447f02a4e4f5592d5531815f6d5cd07644f43f9b33  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/hierarchical.py
+cec62f0d9c171296297c03e318177f1cbd6c4bf762a97be519132b73d5f30a23  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/ring.py
+42daf2df33fc1fb5a191f24c73ef872db98223924b6f382b7ad2580e411ccfbb  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/single.py
+b8e41545da7f21f71274d6f4d9ff1c457e5bf6d50f7c40d67ce6529711599871  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/star.py
+f04094d8f37ef9893a9c5398c58dc26e860425eb0fd8f59d4f4a7d2c9f681b5d  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/types.py
+e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/__init__.py
+2e66206d6609983c96a99e1551fb4d49042e72b4f4ab7baf7fac19308e66d3a9  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/conftest.py
+96255965437e48df0b540ae557f078c9b60d2210deec8567bd7bb844d0550312  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_agent.py
+73e145e0f3a616a6e961fe203bc4146920b8528fb7edb06f68f526dafe984d47  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_analyze_ab_details.py
+33215c31b6cb99bd7dd1a844176232c5017aafc74f10212943f89ae91a446ea4  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_compare_ab.py
+b83565f4a1d361c89138c484a4d6c3f25e6db7b018772ffa17c11e7928d03395  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_grid_search.py
+be3f3fc823f899664241bf3a2eedc7ad722a64ce0923d4ae952add613c088ff7  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_loader.py
+3baa2d91a8a3562741189983bf5082f63883aa5091af75b8121e8fbff4e37065  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_random_round_0.py
+f4b794a0a34f8fe5de2af3c831fdd758d2dfb77ab7485f95ac11db38875ad22c  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_results_analysis.py
+cbe37aca0c9f15f40d7268255fb3cdfd55aece30c132de8352f7dbec3ac3ee97  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_runner_convergence.py
+e78f60ff58fd2d48e5938bd6cc327fa3af3ea70f22fc2b5aff0ee93d50fb2ba9  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_simulate_thresholds.py
+41458fd0ac7fb29bc2458aef6b206ba1337218bc28f279a76197bc18811c42fd  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_topologies.py
+f4ff5a0c8a4282dc6ef813f23c6d8a874d60bf988fca9f41ab46f4dfadd8a67e  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_types.py
+9277de18ac2cd23db4ecbdaf831e55c6ec640c3756accb856f9c6a9f14d840f2  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/uv.lock
+cd0cf81d44cb6a3985608901310fbc11a1298304dfb7c83b201fd22b1738e531  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_analysis/scripts/analyze_gain_bottlenecks.py
+a9813fb57029da6bc64658b7005098e09951a02f349f2ef00133ddabf667ad5c  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_analysis/scripts/analyze_multiseed.py
+d548941432ec8dcb8d66f80076d0f1f34e3a6a45e9bca7c78d9eec5e899b7783  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_analysis/tests/test_analyze_gain_bottlenecks.py
+c179f69926f41525c6bf35225dbfb23b7cffcc2fc04166121e12a0f028d41b70  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_analysis/tests/test_analyze_multiseed.py
+4aeb065fb9fb7de6521c557b73a7e8aef4cf360736d68e4c6a786ec7ddd6607f  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/configs/experiments/real_llm_full_seed42.yaml
+7ff5fc11538def5c077309281e4a544897b3d27ba39a33fffd403cd853b9765a  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/configs/experiments/real_llm_full_seed43.yaml
+e839ac0c62c5a149d1b99d237c7d8c73b98037f1ea9e81931c1c4e998d237d19  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/configs/experiments/real_llm_full_seed44.yaml
+822aa288d18bb4fce8ceb162dcdf560686c8aa6fb9d88a1f9e17e2d3be870478  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/__init__.py
+030602002390b85618417a22724e4c8ea40c72b873447936d8bcbeba218443ff  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/adapters/__init__.py
+8bb35ecc88047544edf2d2aac0551a07c0020bf158aa6c3c3c865ef00c7cb039  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/adapters/medxpertqa.py
+a3622474d26a02b4161fbb413be6b2b41f9d7ec0819563bf5845ecf64ae5689b  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/__init__.py
+662a6cc4de2dde928186990176e0f5c6f58f064a9dee81d4ab4f383334fa1f77  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/anonymization.py
+026be8acd756e50b7178bfb5f5e3879b7f13e7650821c01f5fec25cc912d4781  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/decision_rules.py
+d26b3cd8a7171ea7ef7cdb4c14ae182c6c0bdb59bb6ced165c8dad033cec67d9  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/evidence_board.py
+5cfa68101eac15b950f73da25b83c8eef86c590dc1aee089323975f6cffa1cb1  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/evidence_extraction.py
+648565efab2808056e1b40396fc34f586ac7fbf98f22f23a11da4967e4ccc810  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/metrics.py
+9288ebb5c1370e8ccf6f0f058d8451ee6b88e7de9d8ac497665dabe51dd4e76c  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/perturbations.py
+29768080386d5d50a8b466ca3ad362dca48f7b40343400b00c0e5dea225d05d0  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/visibility.py
+241213586aaeac569e3e6507582153539735b04005d2bd33c7b3548ce99066bb  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/graphs/__init__.py
+b3bb99806e3f83246dcfee505a8c7cd95f5b2da1643b1840c98122bc38207c3d  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/graphs/nodes.py
+2b6545f90ee6bbb99ac4c0189ee4559ab892cd2ad2fff89cbd7dac5e36589171  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/graphs/protocol_graph.py
+8182daba48ab765b8d7e8b4d2b30a5ae4848d6b3bcd63668c7695a22885d0c53  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/graphs/router.py
+514e44696c22e99231c983268a613483ffb46ebaa85aaf61fca6808c81faadc8  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/llm/__init__.py
+1b4e7350c4579c8057f79b1a3aa0d132898b7684f427cb0247c5e6f44eb2a605  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/llm/client.py
+098520d7041ac5e163806925a3b4dcda82ce2e36e40068ea2abf25436798a275  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/llm/cost.py
+1db2f28b4098511443e0b255d4242e6ce39e55063a218bc7e0284bda420b79b2  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/llm/structured_output.py
+822aa288d18bb4fce8ceb162dcdf560686c8aa6fb9d88a1f9e17e2d3be870478  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/__init__.py
+8ad1bf3951384e9d72434825c8147cc49af062ecbdc887192246a0f07089baab  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/devil_advocate.py
+ce424e14df88fcea6d9aa3d6d9c724fb3cafc98ed2377e4537522b5d5ef0cd2e  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/evidence_board.py
+9d8d4796f843bdbedc3aaaed407ef4dc497ecc0fcadc8580a0b0d450b001e3f4  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/hierarchical_mdt.py
+a46dd0823e75fd536281159b4e5105ef52cbabf7b0b3af0649762805d77bfa70  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/judge.py
+3f2bf931196601f88bf0e874adf228d3fac0e6d82536013eb5363a44ebf5d969  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/judge_advisor.py
+979876434f83c5aff1cb79be9447f625476b73c1ce9676f3b976fc9accd3936f  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/open_committee.py
+132c6eb858205a1cde40d6fe79a63978c12a9d758f42e85f2b78fa2a9dba5a8f  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/round0.py
+a9d603aa2dd87bd1080638478f9dca910261aa334aff430e43119ec700250c11  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/__init__.py
+9c3099de541902a300dbb9e5d5fcabd54a8c29e3a6a94fff923b2d8f73414fb9  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/anonymous_evidence_board.py
+01b0d987933d0c9dc537c07f8eb9088c72e6095bfe06560c362ebbb46b8a0b1e  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/common.py
+da25bf3bc078e495276c0cf5209ff01b639fd47587a41d5cd75a37de16688478  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/hierarchical_mdt.py
+4998c64f7271bcb80292e77a07a46b8053ed09933e6a6429576aaf332a13438f  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/judge_advisor.py
+621987ec89a0d5de89533c4ae608bd4cf5bce5c1862013957ce54d7f94049469  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/majority_vote.py
+f730e1c8428315f1aa9456af482862ecd0e60f21ca57313058261c921dd057c3  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/medboard_devil.py
+17ff93d835eb4a3c05ae998dd0a9c9cdc4b2d2ecd79a034a2f25e14eb20bf2c0  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/open_committee.py
+3702aab04fd06ecf51b8064ef8ac47a33b3d94252b49335e99da7d6fdedb69bb  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/single.py
+f4bfc445b119a6f65450850baacbb5c84573569714d6ced353af7848bf8fedbf  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/runners/__init__.py
+16df23f9a7274605a75266a0dd1ec6dc0d4af2e073bf9b6924e7a821be907d64  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/runners/check_data.py
+12094f4a0f6495035ca87f28ea146cdd918e9be40b7688c28480041f8b908ce4  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/runners/run_batch.py
+dc3746938f4ca6123d71e00a335f1022d9db8b8a092631dbf9bac8b3a28e4e9a  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/runners/run_case.py
+ab09ee5937cf163c06fe80d18d6d4af6665c0a20cdf82d9f6cb186fd425db258  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/runners/run_sweep.py
+d403bd9b6a20ef0d7e5bb0fc8f33dd8a60ac50617a835002d6e81e6d98c7c9ae  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/__init__.py
+2a31fb311700f4eedf7b3a2c87ffdf81f251cae480079ca57aa21247d5d145e1  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/agent.py
+42771cdbef4bad91f57cba1ca805ad63ae3b1bda1de29b1be1dd074e960ee208  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/case.py
+68300fa5e8f14c42958152d8adaaaa13f1bc5240c60ef6bd1aae65dc4219797d  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/config.py
+428b4b0f76ba43475a6471297dc4350d6934b1af029137bdfdb9868881ce5b7e  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/evidence.py
+47e36570ff84134ca95b9f78e1c68549efaf3260d97ee1e4c00c8aa9eb6a6808  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/result.py
+deead83015f8fa8a05600eacff06c8adf0ed550798d0ff148a43764a314b18e6  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/state.py
+3716f286a04634aff18a51fe64a5b91d318f758bc4e6ac27921c173c538a00ce  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/storage/__init__.py
+7bf92a914f25fc44a5fbd8d01bf519d24b22345d0ebf8847507ec7ba72a6ac43  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/storage/jsonl_logger.py
+55d400d3f6ab5bdf7ed9b4d97df0cf12e1f1cd97f0a4c47c3c95f59d6f9283ac  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/storage/parquet_writer.py
+75ad241729a476a459a79a1db327c4edbfe943a086ca0ca520940ca3c0b788bd  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/pyproject.toml
+381c6da25108d483bb1e12ef9266135c4bd10a04a007dbef6157865e2f97abee  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_analyze_pilot.py
+332ef306b88e34a28da68f0238b0883d74babad5b95f40ded9c4c6ca06144b8c  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_anonymization.py
+2f0384a233bba10277c3667a1990834def5ab5de0e9b01ebc5195c01956be6c1  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_answer_normalization.py
+003395d811aa11c8d9b1cacf85e8cca8a4bec722c57694ff73812bdde4973d91  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_batch_data_loading.py
+e0802cf6219ac7f83ba763cc9e3e7492dcbc073786087d900d66fe3005a73f52  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_checkpoint_resume.py
+dd429da9a8e9244288b35d13eb2fd4becb2ac610ed105ee09fef1b5b1636e4e2  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_data_sanity.py
+f8077f7d0c8462dc1f43166136a1eb28bce77738c4a55d05a52ebcc583add5fc  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_decision_rules.py
+a2e7455b2a516f66909f0e33895296b996f94789b2f4dbb2716bb1e227118d7d  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_medxpertqa_adapter.py
+a2f0424d4c00e8b275ff8ed3d2882481a2a8e300744ffb2538df048a4dc97988  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_metrics.py
+048bcc867d21faa424de93e7d27a8f7b367bb0b33e79fb2fb360e3ce3c7043cb  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_mock_conformity.py
+27d20443aa5e9461fc453ad0762b28f1e72dded860f47a9dcfd3041b81a16b7a  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_perturbations.py
+9c38e20a55a848b97dceb12e3c69290ea3d05d6c64bb7deff513d1a01cce58e9  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_qa_real_smoke.py
+ca6fe370c9af87b95c6948d4ca0d370b0e323f5e2fbe2eedb51e6b00781c9d86  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_real_protocol_smoke_path.py
+fd357bf78a67831caaff8cb1644b14c53af0c1aefd3514c1f73e7c034e353eb7  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_sweep_runner.py
+f1865ad8322d5d6dbf630d4f43fe877e4eedf8d70c91b910108d159adfd3bbe7  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_visibility_masking.py
+5d59773bbd91b3c0b4039beb42a6546377cd7f9b5459fa12fa1bf6fc5b9bad40  data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/uv.lock
+e2f75642e1cd6bd8869b38067b015ffc6fc4ae2704bb215b956c7cc777b470fa  data/external/results/formal/early_mas_architecture_motivation/figure_sources/extended_data_fig1_source.csv
+4367bc3e301feb0220643646af920f24c2f449a9e9ab42346d9b6f31c29554da  data/external/results/formal/early_mas_architecture_motivation/figure_sources/extended_data_fig1_source.metadata.json
+c02926fbbdda9340e5c2d47e936ec7d95b28f1d59cd7f8008513dde25333db22  data/external/results/formal/early_mas_architecture_motivation/manifests/analysis_ready_manifest.csv
+bb07b6c0fa96bf73e13c8ee37f40bf173b6a6bc1f31f6c20979fc4bebba1a147  data/external/results/formal/early_mas_architecture_motivation/manifests/analysis_ready_manifest.metadata.json
+0a3c22dece8c870c6cab1be0a996bec9a806d72e05ff4f2d580dec966ba91e17  data/external/results/formal/early_mas_architecture_motivation/manifests/code_snapshot_manifest.csv
+01c86cb869c03b75d0d437f1738c22fb3c2c3244a91ad22804feb09b05ba50bc  data/external/results/formal/early_mas_architecture_motivation/manifests/code_snapshot_manifest.metadata.json
+2aa848b944feaa5ad9bfef2fd6cb901bd3445709a4a898026dd757a3000fa041  data/external/results/formal/early_mas_architecture_motivation/manifests/private_release_shards.csv
+a89cfa6873a045481d938f447ac109e931f7085c8ee7551ede61de506c716c3e  data/external/results/formal/early_mas_architecture_motivation/manifests/private_release_shards.metadata.json
+f6243099b9c19e1596ff428e873827948671760f8c6627b07e2a97f3303b25ef  data/external/results/formal/early_mas_architecture_motivation/manifests/source_payload_inventory.csv.gz
+1b7a6ea5972195563cce0ed63b3cbbe1b0e4a7a252c9a7f3653a550c51ecdbad  data/external/results/formal/early_mas_architecture_motivation/manifests/source_payload_inventory.csv.metadata.json
+c598f711d960e2655c04324ef3d063d994623d5506cecf0b2b6c1edf6de8c620  data/external/results/formal/early_mas_architecture_motivation/manifests/source_repositories.yaml
+211fd4a1ca34f3d70923aea27ea2e894d9f462cdee49c4cb4a232386e26df922  data/external/results/formal/early_mas_architecture_motivation/private_release/README.md
+5a1c9c8772a6b2bbc36561619c9834ee774f5cffd2610a9134cdaa65d610374a  data/external/results/formal/early_mas_architecture_motivation/private_release/data/medical_mas_med_agent_baseline_01.tar.gz
+433c0697ed5fc8e91436095fff5e3477568da6e712beebd182428805f1b25703  data/external/results/formal/early_mas_architecture_motivation/private_release/data/medical_mas_med_agent_fixed6_01.tar.gz
+ed6f2c86ef775849195264307cc1d403daafbdecd65fdb620ddd38e71b943c9b  data/external/results/formal/early_mas_architecture_motivation/private_release/data/medical_mas_med_agent_fixed6_02.tar.gz
+d7a4e9874d6be924f4e95c823c1efe1fa4f5c4bb92e938e5f114fb5e9e09b5f1  data/external/results/formal/early_mas_architecture_motivation/private_release/data/medical_mas_med_agent_tuned_01.tar.gz
+45cb17d04e2203a8b47a76f17c45b40d3f0566be2b9826af49c69059d3f2b0bb  data/external/results/formal/early_mas_architecture_motivation/private_release/data/medical_mas_med_board_seed42_01.tar.gz
+4cae1b0c30e3b5ad3cfaa78851e50f00673b6403b08d5e6880df55c66f7e6a4a  data/external/results/formal/early_mas_architecture_motivation/private_release/data/medical_mas_med_board_seed43_01.tar.gz
+e459f187c56bcc9c9706294a0a4c3da9a36123116f726d4043161e6a0ff43fbe  data/external/results/formal/early_mas_architecture_motivation/private_release/data/medical_mas_med_board_seed44_01.tar.gz
+0a3c22dece8c870c6cab1be0a996bec9a806d72e05ff4f2d580dec966ba91e17  data/external/results/formal/early_mas_architecture_motivation/private_release/metadata/code_snapshot_manifest.csv
+2aa848b944feaa5ad9bfef2fd6cb901bd3445709a4a898026dd757a3000fa041  data/external/results/formal/early_mas_architecture_motivation/private_release/metadata/private_release_shards.csv
+0e19cf04a8030280e26eb239651ba47dd5ccb1073835f0cee9566fcc8e214639  data/external/results/formal/early_mas_architecture_motivation/private_release/metadata/private_release_verification.json
+f6243099b9c19e1596ff428e873827948671760f8c6627b07e2a97f3303b25ef  data/external/results/formal/early_mas_architecture_motivation/private_release/metadata/source_payload_inventory.csv.gz
+c598f711d960e2655c04324ef3d063d994623d5506cecf0b2b6c1edf6de8c620  data/external/results/formal/early_mas_architecture_motivation/private_release/metadata/source_repositories.yaml
+3ee0abf92de855eacae55e86f8a7e70eb1ad3edf4dc4e993f2602d71a263d88a  data/external/results/formal/early_mas_architecture_motivation/source_manifest.csv
+3cfdebaa8320b4465094863354e89f1e1c56b10390351ccd2dd755479a7c1323  data/external/results/formal/early_mas_architecture_motivation/source_manifest.metadata.json
 1ea792219dde6450744dbb7ee4b6b10d2881f392d3401f3a3ec109ca5d9b9a79  data/external/results/formal/early_mas_architecture_motivation/summaries/med_agent_consensus_bypass_summary.csv
+ac2b66e58614ca49a8d6a9bd441b0dcd1e1042f7fd73021f608998d52f32e6a5  data/external/results/formal/early_mas_architecture_motivation/summaries/med_agent_round_exit_distribution.csv
+19a98c43b313f0b51a5f6cbac1f9c59ee49f2c9f08c7af5382bf7fba13c191e8  data/external/results/formal/early_mas_architecture_motivation/summaries/med_agent_round_exit_distribution.metadata.json
+e222df3b80b0ac56327bdee3fcf73bf310f07b9b7e4bf66fa8d5803b2b46b186  data/external/results/formal/early_mas_architecture_motivation/summaries/med_agent_topology_summary.csv
+7795a6b7b35ab48fb594e0ff899c05ad3e27f8e51b479f0d7d2ac4fca8e7afd5  data/external/results/formal/early_mas_architecture_motivation/summaries/med_agent_topology_summary.metadata.json
 dd2119db18e7d83e9e8efd5b945aa75c9ec2022210b43cb1f03d6a36bea5cdf8  data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_benchmark_discussion_controls.csv
+9b9e2b43580ea0ffccc6909b39690cefc0b9b3f0a615b3d7a4eebed4ea68c3ee  data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_judge_behavior_summary.csv
+0033e87951b1ef1a25a332acabcba8a1ab11ca0d6d26081a651f892c7cb8673d  data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_judge_behavior_summary.metadata.json
 7dd4aff44b730a7295ab79107164868f884ff9c9b2bd1353b62d9c1a258afe78  data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_mechanism_pilot_summary.csv
+6e870422bd41de783c30a3fae87857f926e341428c9ac1628db73392b58fb8ca  data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_oracle_gap_summary.csv
+8385be80a95ddaa805dcea288822a4039535cea154184ea9e78af62c5600b379  data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_oracle_gap_summary.metadata.json
+75dd67cd93d37a80ccac2dd3b3309b35971730e06060f486eef03ea337f1002f  data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_protocol_bootstrap_ci.csv
+478bacea921a2f73bad6114eb4616cca329ff68bf19534b22870212fd2718ba7  data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_protocol_bootstrap_ci.metadata.json
+44f47a20ab6246861bbe20b224492cb040cdc5137c65253bb2eba6dd53844310  data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_protocol_cost_accuracy_summary.csv
+6e0bbaf6c4aa5b04821134c1360d6fc61298e0a4ac93bab97bb2f1b5e64cf781  data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_protocol_cost_accuracy_summary.metadata.json
 d871b3cabeb8a103203b5cf66be2e8a34f4d2829652a2498b6b8a4c4c5436c51  data/external/results/formal/five_benchmark_cost_efficiency/README.md
 f0c8b7ee3fb57c984b9caee42d68f58f126f2faf7b4d661a443cb1354de3227d  data/external/results/formal/five_benchmark_cost_efficiency/archive/source_inputs/mimo_k16_topup_bank/mimo_k16_topup.launchd.log
 adde7549dc8ef3f9954df91107df5cff37d630b99ab6a2fa4e332a101a5be0a1  data/external/results/formal/five_benchmark_cost_efficiency/archive/source_inputs/mimo_k16_topup_bank/mimo_k16_topup.launchd.status

DATA_MANIFEST.csv CHANGED Viewed

@@ -102,18 +102,179 @@ data/external/evidence/formal/prospective_diffusion_validation/sample_plan.csv,a
 data/external/evidence/formal/prospective_diffusion_validation/sample_plan.metadata.json,archive:manuscript/evidence/formal/prospective_diffusion_validation/sample_plan.metadata.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",711,0998c2022173143eb97b05445d44b4e7f514ab95ebf603362551f5a3159b54e8
 data/external/evidence/formal/prospective_diffusion_validation/source_paths.md,archive:manuscript/evidence/formal/prospective_diffusion_validation/source_paths.md,complete formal research artifact,"raw, intermediate, audit, or summary artifact",771,4a06d9e8fbad3d96b29d6274aafd621bd7283f8846a1c53cc269ed4eb0a596fe
 data/external/results/formal/README.md,archive:results/formal/README.md,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1196,0ef16575b487cfe58527a1dcd2e78fcff1e0e6dda8241985d9f63ced392b895b
-data/external/results/formal/early_mas_architecture_motivation/README.md,archive:results/formal/early_mas_architecture_motivation/README.md,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2340,da10f260bbef39a5a5f6f34491852dc18dcd7b9fa2c3f5a4f07a914158270b60
-data/external/results/formal/early_mas_architecture_motivation/audits/provenance_and_qa.md,archive:results/formal/early_mas_architecture_motivation/audits/provenance_and_qa.md,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1265,53bf2d7b4ab9c562099640bd74899c01dbc006241c17b4bb5d952ebdbb253941
-data/external/results/formal/early_mas_architecture_motivation/figure_sources/extended_data_fig1_source.csv,archive:results/formal/early_mas_architecture_motivation/figure_sources/extended_data_fig1_source.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",24607,727e2600b0f8b1dc5bd8cd72849657097dc339641e5fb5920cf3cee28b388bed
-data/external/results/formal/early_mas_architecture_motivation/source_manifest.csv,archive:results/formal/early_mas_architecture_motivation/source_manifest.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",861,271a9c53f41ceffe49d2b59b4ea40fafb274be2fda4896d9f154d688ba449e71
 data/external/results/formal/early_mas_architecture_motivation/summaries/med_agent_consensus_bypass_summary.csv,archive:results/formal/early_mas_architecture_motivation/summaries/med_agent_consensus_bypass_summary.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2600,1ea792219dde6450744dbb7ee4b6b10d2881f392d3401f3a3ec109ca5d9b9a79
-data/external/results/formal/early_mas_architecture_motivation/summaries/med_agent_round_exit_distribution.csv,archive:results/formal/early_mas_architecture_motivation/summaries/med_agent_round_exit_distribution.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",11534,d6961f0d2688b5193edea4184db9ff6a5c84979aafad92a077729bdce0c780e4
-data/external/results/formal/early_mas_architecture_motivation/summaries/med_agent_topology_summary.csv,archive:results/formal/early_mas_architecture_motivation/summaries/med_agent_topology_summary.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",3645,032b83a23659f5572a5faddf004a8f51297deb4ea108309c9bb77885326988d6
 data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_benchmark_discussion_controls.csv,archive:results/formal/early_mas_architecture_motivation/summaries/med_board_benchmark_discussion_controls.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1850,dd2119db18e7d83e9e8efd5b945aa75c9ec2022210b43cb1f03d6a36bea5cdf8
-data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_judge_behavior_summary.csv,archive:results/formal/early_mas_architecture_motivation/summaries/med_board_judge_behavior_summary.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",885,19f42ed4e2cf8e726c2e73b724b3f10b0ec4dfff0515263e9f780334c36e0abe
 data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_mechanism_pilot_summary.csv,archive:results/formal/early_mas_architecture_motivation/summaries/med_board_mechanism_pilot_summary.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2038,7dd4aff44b730a7295ab79107164868f884ff9c9b2bd1353b62d9c1a258afe78
-data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_oracle_gap_summary.csv,archive:results/formal/early_mas_architecture_motivation/summaries/med_board_oracle_gap_summary.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",889,f7842a093ae110ebfbd1ef477ac261d9db092e3164c3a6fe0fc983fab853316c
-data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_protocol_cost_accuracy_summary.csv,archive:results/formal/early_mas_architecture_motivation/summaries/med_board_protocol_cost_accuracy_summary.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",934,9496bc56850aeb1b5d1bcd8520f6c613c1902843e3c3e9b043151fba1cfdae44
 data/external/results/formal/five_benchmark_cost_efficiency/README.md,archive:results/formal/five_benchmark_cost_efficiency/README.md,complete formal research artifact,"raw, intermediate, audit, or summary artifact",4758,d871b3cabeb8a103203b5cf66be2e8a34f4d2829652a2498b6b8a4c4c5436c51
 data/external/results/formal/five_benchmark_cost_efficiency/archive/source_inputs/mimo_k16_topup_bank/mimo_k16_topup.launchd.log,archive:results/formal/five_benchmark_cost_efficiency/archive/source_inputs/mimo_k16_topup_bank/mimo_k16_topup.launchd.log,complete formal research artifact,"raw, intermediate, audit, or summary artifact",897196,f0c8b7ee3fb57c984b9caee42d68f58f126f2faf7b4d661a443cb1354de3227d
 data/external/results/formal/five_benchmark_cost_efficiency/archive/source_inputs/mimo_k16_topup_bank/mimo_k16_topup.launchd.status,archive:results/formal/five_benchmark_cost_efficiency/archive/source_inputs/mimo_k16_topup_bank/mimo_k16_topup.launchd.status,complete formal research artifact,"raw, intermediate, audit, or summary artifact",25,adde7549dc8ef3f9954df91107df5cff37d630b99ab6a2fa4e332a101a5be0a1

 data/external/evidence/formal/prospective_diffusion_validation/sample_plan.metadata.json,archive:manuscript/evidence/formal/prospective_diffusion_validation/sample_plan.metadata.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",711,0998c2022173143eb97b05445d44b4e7f514ab95ebf603362551f5a3159b54e8
 data/external/evidence/formal/prospective_diffusion_validation/source_paths.md,archive:manuscript/evidence/formal/prospective_diffusion_validation/source_paths.md,complete formal research artifact,"raw, intermediate, audit, or summary artifact",771,4a06d9e8fbad3d96b29d6274aafd621bd7283f8846a1c53cc269ed4eb0a596fe
 data/external/results/formal/README.md,archive:results/formal/README.md,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1196,0ef16575b487cfe58527a1dcd2e78fcff1e0e6dda8241985d9f63ced392b895b
+data/external/results/formal/early_mas_architecture_motivation/README.md,archive:results/formal/early_mas_architecture_motivation/README.md,complete formal research artifact,"raw, intermediate, audit, or summary artifact",4801,21ce9218073aee47e04594de2386e8a3fd0e4a6f48ca5b7c97fd2d13a7beb0d1
+data/external/results/formal/early_mas_architecture_motivation/analysis_ready/configs/med_board/real_llm_full_seed42.yaml,archive:results/formal/early_mas_architecture_motivation/analysis_ready/configs/med_board/real_llm_full_seed42.yaml,complete formal research artifact,"raw, intermediate, audit, or summary artifact",674,4aeb065fb9fb7de6521c557b73a7e8aef4cf360736d68e4c6a786ec7ddd6607f
+data/external/results/formal/early_mas_architecture_motivation/analysis_ready/configs/med_board/real_llm_full_seed43.yaml,archive:results/formal/early_mas_architecture_motivation/analysis_ready/configs/med_board/real_llm_full_seed43.yaml,complete formal research artifact,"raw, intermediate, audit, or summary artifact",674,7ff5fc11538def5c077309281e4a544897b3d27ba39a33fffd403cd853b9765a
+data/external/results/formal/early_mas_architecture_motivation/analysis_ready/configs/med_board/real_llm_full_seed44.yaml,archive:results/formal/early_mas_architecture_motivation/analysis_ready/configs/med_board/real_llm_full_seed44.yaml,complete formal research artifact,"raw, intermediate, audit, or summary artifact",674,e839ac0c62c5a149d1b99d237c7d8c73b98037f1ea9e81931c1c4e998d237d19
+data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/baseline/all_results.csv,archive:results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/baseline/all_results.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",927658,1d62980ea5a0d6fee72bf94f408a60be27b99995985963395e51e18aeea7d675
+data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/baseline/summary.json,archive:results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/baseline/summary.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2577,2c0bdec413c879684db2528c4eb37d1cb7f8f366d49aacadedcd5bbc931eef16
+data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/fixed6/all_results.csv,archive:results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/fixed6/all_results.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",945846,047cab62f4831e320516d14a5554cf3bd7f595004fa4400da84955bd5627675e
+data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/fixed6/summary.json,archive:results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/fixed6/summary.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2213,c4013cb3ad5255f07555f99deb4b96bb0d2202e260071d4fc0a660f98cb33ebe
+data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/tuned/all_results.csv,archive:results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/tuned/all_results.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1049233,c8d553568735f5c1bbe3f79ca280312090d448594709bace7fa77a21d15d86a4
+data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/tuned/summary.json,archive:results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/tuned/summary.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2649,ca4122852b29fb9549c17bb08a22f86358eda2e47d323579dd87c433ff368bbb
+data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_board/gain_bottleneck_report.json,archive:results/formal/early_mas_architecture_motivation/analysis_ready/med_board/gain_bottleneck_report.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",19990,054b69cb30f3eac151d40e29cda7e4970705f288b1ca6a4c586ad32e33719acd
+data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_board/legacy_3seed_summary.md,archive:results/formal/early_mas_architecture_motivation/analysis_ready/med_board/legacy_3seed_summary.md,complete formal research artifact,"raw, intermediate, audit, or summary artifact",3900,d64229bd5fe5475527b6f5cd6719a3da09ff1bf1e086513d2ae248b6dbe0f15a
+data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_board/oracle_trials.csv,archive:results/formal/early_mas_architecture_motivation/analysis_ready/med_board/oracle_trials.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2183571,de275959a071640633f446a344608fb84e57a49294b2ab956179de129f7e1638
+data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_board/protocol_results.csv,archive:results/formal/early_mas_architecture_motivation/analysis_ready/med_board/protocol_results.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2346246,b9d074ef8eb51315ceee82bb63dbc32da4926f42d8b7c04e3d16bdc74c3298e4
+data/external/results/formal/early_mas_architecture_motivation/audits/code_security_gate.md,archive:results/formal/early_mas_architecture_motivation/audits/code_security_gate.md,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1720,d382e8fda4a21bd54aeb7f975b3e2df690f09f5fbb3a73d90ef22145907fcee9
+data/external/results/formal/early_mas_architecture_motivation/audits/code_snapshot_security_scan.json,archive:results/formal/early_mas_architecture_motivation/audits/code_snapshot_security_scan.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",255,62aeedae6f9c813a65a4de47f626196a4a0aa69d8ca6729601fd263192b8f970
+data/external/results/formal/early_mas_architecture_motivation/audits/code_snapshot_security_scan.metadata.json,archive:results/formal/early_mas_architecture_motivation/audits/code_snapshot_security_scan.metadata.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",695,c052ffdf2024b5024b78cb61fca415b8e52da72300407b9aac335e31ec1b85c8
+data/external/results/formal/early_mas_architecture_motivation/audits/payload_security_scan.json,archive:results/formal/early_mas_architecture_motivation/audits/payload_security_scan.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",454,d3e82e115abddff2d6151381f0d44e52484471579f9c44629e7000754ad4afe1
+data/external/results/formal/early_mas_architecture_motivation/audits/payload_security_scan.metadata.json,archive:results/formal/early_mas_architecture_motivation/audits/payload_security_scan.metadata.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",661,5c018829edb980f1e488e746fc680b1b9254d97f39f50f596e4dbf1fd98e6327
+data/external/results/formal/early_mas_architecture_motivation/audits/private_release_verification.json,archive:results/formal/early_mas_architecture_motivation/audits/private_release_verification.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",444,0e19cf04a8030280e26eb239651ba47dd5ccb1073835f0cee9566fcc8e214639
+data/external/results/formal/early_mas_architecture_motivation/audits/private_release_verification.metadata.json,archive:results/formal/early_mas_architecture_motivation/audits/private_release_verification.metadata.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",891,94d6c3ad4f5f9b662a028cf0cec93c56a5b5070c5140f7c1dccbcb459b70d61d
+data/external/results/formal/early_mas_architecture_motivation/audits/provenance_and_qa.md,archive:results/formal/early_mas_architecture_motivation/audits/provenance_and_qa.md,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1273,cea540e96b20147b86dd2791dacf60ba29504d66c55e0300bc975e284420ae22
+data/external/results/formal/early_mas_architecture_motivation/audits/provenance_and_qa.metadata.json,archive:results/formal/early_mas_architecture_motivation/audits/provenance_and_qa.metadata.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1931,b0ba46b78b85d8e74fb7e47ed59ec655db9a40284dcc502c6e5eb5095a7e4c6f
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/pyproject.toml,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/pyproject.toml,complete formal research artifact,"raw, intermediate, audit, or summary artifact",255,2a52f22d56d2bccb87573b9e349497ae5f057b8108ffa8558b19f29ec1342ec0
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/pytest.ini,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/pytest.ini,complete formal research artifact,"raw, intermediate, audit, or summary artifact",141,b769e070eb76e69b821b68d31bcf20e850ec74eda6491c664d04184e9ce1691b
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/__init__.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/__init__.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",42,222a8f2c7f63d7e3d4a5165f515111016b552c563a94420adb3fe23cd4e961fd
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/__main__.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/__main__.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",8683,da3c833adaea421e012e1c6c57c3d919035430ef0a4e30e97e458fc84761788b
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/agent.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/agent.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",17801,428357d3921165e377dfafb5b88dc79899e2ae69a911fe5460a3af6b9e1cae18
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/config.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/config.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1825,f67c2acb0b1bf49b613e977a189c968a6d9d0aa6630d5576753bafc7557a9e70
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/data/__init__.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/data/__init__.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",33,c0bf0688d03ec00e93160193b9cc13960c99c05b90167e72f12d16b5fd5477b3
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/data/loader.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/data/loader.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1504,247dace738280caae0eea745349fa02eb10f92106147ae6b4d988bc9968fc2ba
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/metrics/__init__.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/metrics/__init__.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",36,cca37fb69099489cbaed5ba0d9fd71636ece6a14313ce9946d9d33cd562a84f7
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/metrics/analyzer.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/metrics/analyzer.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",3161,81285eeb497531b129a5ba38d15c8059cf911b3e8e27aa587bb2bbf69c1f7784
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/metrics/collector.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/metrics/collector.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1225,21096979d1ef7087f23011e71cd2b211b2a9584ee9c18b12d8faa929c71790c1
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/runner.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/runner.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",16046,fd9b29bac159a713df19ebadba81eb52e165447ceb76d197b93b1c2edb91b15f
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/__init__.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/__init__.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",483,8124df00004e9016f10c5b65d01bf3cfaca4f1dc5c63828a6b25f79f406c711f
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/base.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/base.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",311,8de0d4d54010d988ac00f13fafe74a50bd2ecffc5c5a2035a4f6569d0ba11cf3
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/complete.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/complete.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",436,dc160e6f7425e7d4e330feca841a66e0b0f11400c089fe12adeba29338536be8
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/hierarchical.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/hierarchical.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1395,afc8676ed6467bbf02e0c9447f02a4e4f5592d5531815f6d5cd07644f43f9b33
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/ring.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/ring.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",733,cec62f0d9c171296297c03e318177f1cbd6c4bf762a97be519132b73d5f30a23
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/single.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/single.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",293,42daf2df33fc1fb5a191f24c73ef872db98223924b6f382b7ad2580e411ccfbb
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/star.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/star.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",841,b8e41545da7f21f71274d6f4d9ff1c457e5bf6d50f7c40d67ce6529711599871
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/types.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/types.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",3092,f04094d8f37ef9893a9c5398c58dc26e860425eb0fd8f59d4f4a7d2c9f681b5d
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/__init__.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/__init__.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",0,e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/conftest.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/conftest.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1276,2e66206d6609983c96a99e1551fb4d49042e72b4f4ab7baf7fac19308e66d3a9
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_agent.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_agent.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",7609,96255965437e48df0b540ae557f078c9b60d2210deec8567bd7bb844d0550312
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_analyze_ab_details.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_analyze_ab_details.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",9216,73e145e0f3a616a6e961fe203bc4146920b8528fb7edb06f68f526dafe984d47
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_compare_ab.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_compare_ab.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",4352,33215c31b6cb99bd7dd1a844176232c5017aafc74f10212943f89ae91a446ea4
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_grid_search.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_grid_search.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",8101,b83565f4a1d361c89138c484a4d6c3f25e6db7b018772ffa17c11e7928d03395
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_loader.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_loader.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",484,be3f3fc823f899664241bf3a2eedc7ad722a64ce0923d4ae952add613c088ff7
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_random_round_0.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_random_round_0.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",3962,3baa2d91a8a3562741189983bf5082f63883aa5091af75b8121e8fbff4e37065
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_results_analysis.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_results_analysis.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",6494,f4b794a0a34f8fe5de2af3c831fdd758d2dfb77ab7485f95ac11db38875ad22c
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_runner_convergence.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_runner_convergence.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",7110,cbe37aca0c9f15f40d7268255fb3cdfd55aece30c132de8352f7dbec3ac3ee97
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_simulate_thresholds.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_simulate_thresholds.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",7480,e78f60ff58fd2d48e5938bd6cc327fa3af3ea70f22fc2b5aff0ee93d50fb2ba9
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_topologies.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_topologies.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2788,41458fd0ac7fb29bc2458aef6b206ba1337218bc28f279a76197bc18811c42fd
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_types.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_types.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1927,f4ff5a0c8a4282dc6ef813f23c6d8a874d60bf988fca9f41ab46f4dfadd8a67e
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/uv.lock,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/uv.lock,complete formal research artifact,"raw, intermediate, audit, or summary artifact",311193,9277de18ac2cd23db4ecbdaf831e55c6ec640c3756accb856f9c6a9f14d840f2
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_analysis/scripts/analyze_gain_bottlenecks.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_analysis/scripts/analyze_gain_bottlenecks.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",40503,cd0cf81d44cb6a3985608901310fbc11a1298304dfb7c83b201fd22b1738e531
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_analysis/scripts/analyze_multiseed.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_analysis/scripts/analyze_multiseed.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",20528,a9813fb57029da6bc64658b7005098e09951a02f349f2ef00133ddabf667ad5c
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_analysis/tests/test_analyze_gain_bottlenecks.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_analysis/tests/test_analyze_gain_bottlenecks.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",11704,d548941432ec8dcb8d66f80076d0f1f34e3a6a45e9bca7c78d9eec5e899b7783
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_analysis/tests/test_analyze_multiseed.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_analysis/tests/test_analyze_multiseed.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",6774,c179f69926f41525c6bf35225dbfb23b7cffcc2fc04166121e12a0f028d41b70
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/configs/experiments/real_llm_full_seed42.yaml,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/configs/experiments/real_llm_full_seed42.yaml,complete formal research artifact,"raw, intermediate, audit, or summary artifact",674,4aeb065fb9fb7de6521c557b73a7e8aef4cf360736d68e4c6a786ec7ddd6607f
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/configs/experiments/real_llm_full_seed43.yaml,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/configs/experiments/real_llm_full_seed43.yaml,complete formal research artifact,"raw, intermediate, audit, or summary artifact",674,7ff5fc11538def5c077309281e4a544897b3d27ba39a33fffd403cd853b9765a
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/configs/experiments/real_llm_full_seed44.yaml,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/configs/experiments/real_llm_full_seed44.yaml,complete formal research artifact,"raw, intermediate, audit, or summary artifact",674,e839ac0c62c5a149d1b99d237c7d8c73b98037f1ea9e81931c1c4e998d237d19
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/__init__.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/__init__.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",31,822aa288d18bb4fce8ceb162dcdf560686c8aa6fb9d88a1f9e17e2d3be870478
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/adapters/__init__.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/adapters/__init__.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",250,030602002390b85618417a22724e4c8ea40c72b873447936d8bcbeba218443ff
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/adapters/medxpertqa.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/adapters/medxpertqa.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",5970,8bb35ecc88047544edf2d2aac0551a07c0020bf158aa6c3c3c865ef00c7cb039
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/__init__.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/__init__.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",870,a3622474d26a02b4161fbb413be6b2b41f9d7ec0819563bf5845ecf64ae5689b
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/anonymization.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/anonymization.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1080,662a6cc4de2dde928186990176e0f5c6f58f064a9dee81d4ab4f383334fa1f77
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/decision_rules.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/decision_rules.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2420,026be8acd756e50b7178bfb5f5e3879b7f13e7650821c01f5fec25cc912d4781
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/evidence_board.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/evidence_board.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",707,d26b3cd8a7171ea7ef7cdb4c14ae182c6c0bdb59bb6ced165c8dad033cec67d9
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/evidence_extraction.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/evidence_extraction.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1462,5cfa68101eac15b950f73da25b83c8eef86c590dc1aee089323975f6cffa1cb1
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/metrics.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/metrics.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",3042,648565efab2808056e1b40396fc34f586ac7fbf98f22f23a11da4967e4ccc810
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/perturbations.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/perturbations.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2516,9288ebb5c1370e8ccf6f0f058d8451ee6b88e7de9d8ac497665dabe51dd4e76c
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/visibility.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/core/visibility.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2516,29768080386d5d50a8b466ca3ad362dca48f7b40343400b00c0e5dea225d05d0
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/graphs/__init__.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/graphs/__init__.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",252,241213586aaeac569e3e6507582153539735b04005d2bd33c7b3548ce99066bb
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/graphs/nodes.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/graphs/nodes.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",4032,b3bb99806e3f83246dcfee505a8c7cd95f5b2da1643b1840c98122bc38207c3d
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/graphs/protocol_graph.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/graphs/protocol_graph.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2507,2b6545f90ee6bbb99ac4c0189ee4559ab892cd2ad2fff89cbd7dac5e36589171
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/graphs/router.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/graphs/router.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1625,8182daba48ab765b8d7e8b4d2b30a5ae4848d6b3bcd63668c7695a22885d0c53
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/llm/__init__.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/llm/__init__.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",353,514e44696c22e99231c983268a613483ffb46ebaa85aaf61fca6808c81faadc8
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/llm/client.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/llm/client.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2839,1b4e7350c4579c8057f79b1a3aa0d132898b7684f427cb0247c5e6f44eb2a605
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/llm/cost.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/llm/cost.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",494,098520d7041ac5e163806925a3b4dcda82ce2e36e40068ea2abf25436798a275
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/llm/structured_output.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/llm/structured_output.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",496,1db2f28b4098511443e0b255d4242e6ce39e55063a218bc7e0284bda420b79b2
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/__init__.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/__init__.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",31,822aa288d18bb4fce8ceb162dcdf560686c8aa6fb9d88a1f9e17e2d3be870478
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/devil_advocate.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/devil_advocate.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",104,8ad1bf3951384e9d72434825c8147cc49af062ecbdc887192246a0f07089baab
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/evidence_board.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/evidence_board.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",91,ce424e14df88fcea6d9aa3d6d9c724fb3cafc98ed2377e4537522b5d5ef0cd2e
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/hierarchical_mdt.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/hierarchical_mdt.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",91,9d8d4796f843bdbedc3aaaed407ef4dc497ecc0fcadc8580a0b0d450b001e3f4
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/judge.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/judge.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",5288,a46dd0823e75fd536281159b4e5105ef52cbabf7b0b3af0649762805d77bfa70
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/judge_advisor.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/judge_advisor.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",83,3f2bf931196601f88bf0e874adf228d3fac0e6d82536013eb5363a44ebf5d969
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/open_committee.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/open_committee.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",4160,979876434f83c5aff1cb79be9447f625476b73c1ce9676f3b976fc9accd3936f
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/round0.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/prompts/round0.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2814,132c6eb858205a1cde40d6fe79a63978c12a9d758f42e85f2b78fa2a9dba5a8f
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/__init__.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/__init__.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",744,a9d603aa2dd87bd1080638478f9dca910261aa334aff430e43119ec700250c11
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/anonymous_evidence_board.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/anonymous_evidence_board.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",475,9c3099de541902a300dbb9e5d5fcabd54a8c29e3a6a94fff923b2d8f73414fb9
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/common.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/common.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",31618,01b0d987933d0c9dc537c07f8eb9088c72e6095bfe06560c362ebbb46b8a0b1e
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/hierarchical_mdt.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/hierarchical_mdt.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",452,da25bf3bc078e495276c0cf5209ff01b639fd47587a41d5cd75a37de16688478
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/judge_advisor.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/judge_advisor.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",443,4998c64f7271bcb80292e77a07a46b8053ed09933e6a6429576aaf332a13438f
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/majority_vote.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/majority_vote.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",443,621987ec89a0d5de89533c4ae608bd4cf5bce5c1862013957ce54d7f94049469
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/medboard_devil.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/medboard_devil.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",455,f730e1c8428315f1aa9456af482862ecd0e60f21ca57313058261c921dd057c3
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/open_committee.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/open_committee.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",446,17ff93d835eb4a3c05ae998dd0a9c9cdc4b2d2ecd79a034a2f25e14eb20bf2c0
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/single.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/protocols/single.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",440,3702aab04fd06ecf51b8064ef8ac47a33b3d94252b49335e99da7d6fdedb69bb
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/runners/__init__.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/runners/__init__.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",39,f4bfc445b119a6f65450850baacbb5c84573569714d6ced353af7848bf8fedbf
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/runners/check_data.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/runners/check_data.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2989,16df23f9a7274605a75266a0dd1ec6dc0d4af2e073bf9b6924e7a821be907d64
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/runners/run_batch.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/runners/run_batch.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",5599,12094f4a0f6495035ca87f28ea146cdd918e9be40b7688c28480041f8b908ce4
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/runners/run_case.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/runners/run_case.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1117,dc3746938f4ca6123d71e00a335f1022d9db8b8a092631dbf9bac8b3a28e4e9a
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/runners/run_sweep.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/runners/run_sweep.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",4679,ab09ee5937cf163c06fe80d18d6d4af6665c0a20cdf82d9f6cb186fd425db258
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/__init__.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/__init__.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",529,d403bd9b6a20ef0d7e5bb0fc8f33dd8a60ac50617a835002d6e81e6d98c7c9ae
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/agent.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/agent.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",766,2a31fb311700f4eedf7b3a2c87ffdf81f251cae480079ca57aa21247d5d145e1
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/case.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/case.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",303,42771cdbef4bad91f57cba1ca805ad63ae3b1bda1de29b1be1dd074e960ee208
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/config.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/config.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2551,68300fa5e8f14c42958152d8adaaaa13f1bc5240c60ef6bd1aae65dc4219797d
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/evidence.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/evidence.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",625,428b4b0f76ba43475a6471297dc4350d6934b1af029137bdfdb9868881ce5b7e
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/result.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/result.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1178,47e36570ff84134ca95b9f78e1c68549efaf3260d97ee1e4c00c8aa9eb6a6808
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/state.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/schemas/state.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",825,deead83015f8fa8a05600eacff06c8adf0ed550798d0ff148a43764a314b18e6
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/storage/__init__.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/storage/__init__.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",205,3716f286a04634aff18a51fe64a5b91d318f758bc4e6ac27921c173c538a00ce
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/storage/jsonl_logger.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/storage/jsonl_logger.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1163,7bf92a914f25fc44a5fbd8d01bf519d24b22345d0ebf8847507ec7ba72a6ac43
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/storage/parquet_writer.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/medboard/storage/parquet_writer.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1159,55d400d3f6ab5bdf7ed9b4d97df0cf12e1f1cd97f0a4c47c3c95f59d6f9283ac
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/pyproject.toml,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/pyproject.toml,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1012,75ad241729a476a459a79a1db327c4edbfe943a086ca0ca520940ca3c0b788bd
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_analyze_pilot.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_analyze_pilot.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",4834,381c6da25108d483bb1e12ef9266135c4bd10a04a007dbef6157865e2f97abee
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_anonymization.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_anonymization.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1014,332ef306b88e34a28da68f0238b0883d74babad5b95f40ded9c4c6ca06144b8c
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_answer_normalization.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_answer_normalization.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2467,2f0384a233bba10277c3667a1990834def5ab5de0e9b01ebc5195c01956be6c1
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_batch_data_loading.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_batch_data_loading.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",3566,003395d811aa11c8d9b1cacf85e8cca8a4bec722c57694ff73812bdde4973d91
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_checkpoint_resume.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_checkpoint_resume.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",6763,e0802cf6219ac7f83ba763cc9e3e7492dcbc073786087d900d66fe3005a73f52
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_data_sanity.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_data_sanity.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2915,dd429da9a8e9244288b35d13eb2fd4becb2ac610ed105ee09fef1b5b1636e4e2
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_decision_rules.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_decision_rules.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",3784,f8077f7d0c8462dc1f43166136a1eb28bce77738c4a55d05a52ebcc583add5fc
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_medxpertqa_adapter.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_medxpertqa_adapter.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",5263,a2e7455b2a516f66909f0e33895296b996f94789b2f4dbb2716bb1e227118d7d
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_metrics.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_metrics.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",7338,a2f0424d4c00e8b275ff8ed3d2882481a2a8e300744ffb2538df048a4dc97988
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_mock_conformity.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_mock_conformity.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",4000,048bcc867d21faa424de93e7d27a8f7b367bb0b33e79fb2fb360e3ce3c7043cb
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_perturbations.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_perturbations.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2657,27d20443aa5e9461fc453ad0762b28f1e72dded860f47a9dcfd3041b81a16b7a
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_qa_real_smoke.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_qa_real_smoke.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",4076,9c38e20a55a848b97dceb12e3c69290ea3d05d6c64bb7deff513d1a01cce58e9
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_real_protocol_smoke_path.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_real_protocol_smoke_path.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",19221,ca6fe370c9af87b95c6948d4ca0d370b0e323f5e2fbe2eedb51e6b00781c9d86
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_sweep_runner.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_sweep_runner.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",6144,fd357bf78a67831caaff8cb1644b14c53af0c1aefd3514c1f73e7c034e353eb7
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_visibility_masking.py,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/tests/test_visibility_masking.py,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1054,f1865ad8322d5d6dbf630d4f43fe877e4eedf8d70c91b910108d159adfd3bbe7
+data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/uv.lock,archive:results/formal/early_mas_architecture_motivation/code_snapshot/med_board_execution/uv.lock,complete formal research artifact,"raw, intermediate, audit, or summary artifact",644766,5d59773bbd91b3c0b4039beb42a6546377cd7f9b5459fa12fa1bf6fc5b9bad40
+data/external/results/formal/early_mas_architecture_motivation/figure_sources/extended_data_fig1_source.csv,archive:results/formal/early_mas_architecture_motivation/figure_sources/extended_data_fig1_source.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",23993,e2f75642e1cd6bd8869b38067b015ffc6fc4ae2704bb215b956c7cc777b470fa
+data/external/results/formal/early_mas_architecture_motivation/figure_sources/extended_data_fig1_source.metadata.json,archive:results/formal/early_mas_architecture_motivation/figure_sources/extended_data_fig1_source.metadata.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1931,4367bc3e301feb0220643646af920f24c2f449a9e9ab42346d9b6f31c29554da
+data/external/results/formal/early_mas_architecture_motivation/manifests/analysis_ready_manifest.csv,archive:results/formal/early_mas_architecture_motivation/manifests/analysis_ready_manifest.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",8785,c02926fbbdda9340e5c2d47e936ec7d95b28f1d59cd7f8008513dde25333db22
+data/external/results/formal/early_mas_architecture_motivation/manifests/analysis_ready_manifest.metadata.json,archive:results/formal/early_mas_architecture_motivation/manifests/analysis_ready_manifest.metadata.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",621,bb07b6c0fa96bf73e13c8ee37f40bf173b6a6bc1f31f6c20979fc4bebba1a147
+data/external/results/formal/early_mas_architecture_motivation/manifests/code_snapshot_manifest.csv,archive:results/formal/early_mas_architecture_motivation/manifests/code_snapshot_manifest.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",35049,0a3c22dece8c870c6cab1be0a996bec9a806d72e05ff4f2d580dec966ba91e17
+data/external/results/formal/early_mas_architecture_motivation/manifests/code_snapshot_manifest.metadata.json,archive:results/formal/early_mas_architecture_motivation/manifests/code_snapshot_manifest.metadata.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",693,01c86cb869c03b75d0d437f1738c22fb3c2c3244a91ad22804feb09b05ba50bc
+data/external/results/formal/early_mas_architecture_motivation/manifests/private_release_shards.csv,archive:results/formal/early_mas_architecture_motivation/manifests/private_release_shards.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1206,2aa848b944feaa5ad9bfef2fd6cb901bd3445709a4a898026dd757a3000fa041
+data/external/results/formal/early_mas_architecture_motivation/manifests/private_release_shards.metadata.json,archive:results/formal/early_mas_architecture_motivation/manifests/private_release_shards.metadata.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",889,a89cfa6873a045481d938f447ac109e931f7085c8ee7551ede61de506c716c3e
+data/external/results/formal/early_mas_architecture_motivation/manifests/source_payload_inventory.csv.gz,archive:results/formal/early_mas_architecture_motivation/manifests/source_payload_inventory.csv.gz,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2389352,f6243099b9c19e1596ff428e873827948671760f8c6627b07e2a97f3303b25ef
+data/external/results/formal/early_mas_architecture_motivation/manifests/source_payload_inventory.csv.metadata.json,archive:results/formal/early_mas_architecture_motivation/manifests/source_payload_inventory.csv.metadata.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",659,1b7a6ea5972195563cce0ed63b3cbbe1b0e4a7a252c9a7f3653a550c51ecdbad
+data/external/results/formal/early_mas_architecture_motivation/manifests/source_repositories.yaml,archive:results/formal/early_mas_architecture_motivation/manifests/source_repositories.yaml,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1468,c598f711d960e2655c04324ef3d063d994623d5506cecf0b2b6c1edf6de8c620
+data/external/results/formal/early_mas_architecture_motivation/private_release/README.md,archive:results/tmp/medical_mas_private_release/README.md,controlled medical MAS raw archive,private deterministic archive shard or release metadata,863,211fd4a1ca34f3d70923aea27ea2e894d9f462cdee49c4cb4a232386e26df922
+data/external/results/formal/early_mas_architecture_motivation/private_release/data/medical_mas_med_agent_baseline_01.tar.gz,archive:results/tmp/medical_mas_private_release/data/medical_mas_med_agent_baseline_01.tar.gz,controlled medical MAS raw archive,private deterministic archive shard or release metadata,11072544,5a1c9c8772a6b2bbc36561619c9834ee774f5cffd2610a9134cdaa65d610374a
+data/external/results/formal/early_mas_architecture_motivation/private_release/data/medical_mas_med_agent_fixed6_01.tar.gz,archive:results/tmp/medical_mas_private_release/data/medical_mas_med_agent_fixed6_01.tar.gz,controlled medical MAS raw archive,private deterministic archive shard or release metadata,26893988,433c0697ed5fc8e91436095fff5e3477568da6e712beebd182428805f1b25703
+data/external/results/formal/early_mas_architecture_motivation/private_release/data/medical_mas_med_agent_fixed6_02.tar.gz,archive:results/tmp/medical_mas_private_release/data/medical_mas_med_agent_fixed6_02.tar.gz,controlled medical MAS raw archive,private deterministic archive shard or release metadata,9021076,ed6f2c86ef775849195264307cc1d403daafbdecd65fdb620ddd38e71b943c9b
+data/external/results/formal/early_mas_architecture_motivation/private_release/data/medical_mas_med_agent_tuned_01.tar.gz,archive:results/tmp/medical_mas_private_release/data/medical_mas_med_agent_tuned_01.tar.gz,controlled medical MAS raw archive,private deterministic archive shard or release metadata,18819201,d7a4e9874d6be924f4e95c823c1efe1fa4f5c4bb92e938e5f114fb5e9e09b5f1
+data/external/results/formal/early_mas_architecture_motivation/private_release/data/medical_mas_med_board_seed42_01.tar.gz,archive:results/tmp/medical_mas_private_release/data/medical_mas_med_board_seed42_01.tar.gz,controlled medical MAS raw archive,private deterministic archive shard or release metadata,19078912,45cb17d04e2203a8b47a76f17c45b40d3f0566be2b9826af49c69059d3f2b0bb
+data/external/results/formal/early_mas_architecture_motivation/private_release/data/medical_mas_med_board_seed43_01.tar.gz,archive:results/tmp/medical_mas_private_release/data/medical_mas_med_board_seed43_01.tar.gz,controlled medical MAS raw archive,private deterministic archive shard or release metadata,18989294,4cae1b0c30e3b5ad3cfaa78851e50f00673b6403b08d5e6880df55c66f7e6a4a
+data/external/results/formal/early_mas_architecture_motivation/private_release/data/medical_mas_med_board_seed44_01.tar.gz,archive:results/tmp/medical_mas_private_release/data/medical_mas_med_board_seed44_01.tar.gz,controlled medical MAS raw archive,private deterministic archive shard or release metadata,18940738,e459f187c56bcc9c9706294a0a4c3da9a36123116f726d4043161e6a0ff43fbe
+data/external/results/formal/early_mas_architecture_motivation/private_release/metadata/code_snapshot_manifest.csv,archive:results/tmp/medical_mas_private_release/metadata/code_snapshot_manifest.csv,controlled medical MAS raw archive,private deterministic archive shard or release metadata,35049,0a3c22dece8c870c6cab1be0a996bec9a806d72e05ff4f2d580dec966ba91e17
+data/external/results/formal/early_mas_architecture_motivation/private_release/metadata/private_release_shards.csv,archive:results/tmp/medical_mas_private_release/metadata/private_release_shards.csv,controlled medical MAS raw archive,private deterministic archive shard or release metadata,1206,2aa848b944feaa5ad9bfef2fd6cb901bd3445709a4a898026dd757a3000fa041
+data/external/results/formal/early_mas_architecture_motivation/private_release/metadata/private_release_verification.json,archive:results/tmp/medical_mas_private_release/metadata/private_release_verification.json,controlled medical MAS raw archive,private deterministic archive shard or release metadata,444,0e19cf04a8030280e26eb239651ba47dd5ccb1073835f0cee9566fcc8e214639
+data/external/results/formal/early_mas_architecture_motivation/private_release/metadata/source_payload_inventory.csv.gz,archive:results/tmp/medical_mas_private_release/metadata/source_payload_inventory.csv.gz,controlled medical MAS raw archive,private deterministic archive shard or release metadata,2389352,f6243099b9c19e1596ff428e873827948671760f8c6627b07e2a97f3303b25ef
+data/external/results/formal/early_mas_architecture_motivation/private_release/metadata/source_repositories.yaml,archive:results/tmp/medical_mas_private_release/metadata/source_repositories.yaml,controlled medical MAS raw archive,private deterministic archive shard or release metadata,1468,c598f711d960e2655c04324ef3d063d994623d5506cecf0b2b6c1edf6de8c620
+data/external/results/formal/early_mas_architecture_motivation/source_manifest.csv,archive:results/formal/early_mas_architecture_motivation/source_manifest.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2406,3ee0abf92de855eacae55e86f8a7e70eb1ad3edf4dc4e993f2602d71a263d88a
+data/external/results/formal/early_mas_architecture_motivation/source_manifest.metadata.json,archive:results/formal/early_mas_architecture_motivation/source_manifest.metadata.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1931,3cfdebaa8320b4465094863354e89f1e1c56b10390351ccd2dd755479a7c1323
 data/external/results/formal/early_mas_architecture_motivation/summaries/med_agent_consensus_bypass_summary.csv,archive:results/formal/early_mas_architecture_motivation/summaries/med_agent_consensus_bypass_summary.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2600,1ea792219dde6450744dbb7ee4b6b10d2881f392d3401f3a3ec109ca5d9b9a79
+data/external/results/formal/early_mas_architecture_motivation/summaries/med_agent_round_exit_distribution.csv,archive:results/formal/early_mas_architecture_motivation/summaries/med_agent_round_exit_distribution.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",12491,ac2b66e58614ca49a8d6a9bd441b0dcd1e1042f7fd73021f608998d52f32e6a5
+data/external/results/formal/early_mas_architecture_motivation/summaries/med_agent_round_exit_distribution.metadata.json,archive:results/formal/early_mas_architecture_motivation/summaries/med_agent_round_exit_distribution.metadata.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1931,19a98c43b313f0b51a5f6cbac1f9c59ee49f2c9f08c7af5382bf7fba13c191e8
+data/external/results/formal/early_mas_architecture_motivation/summaries/med_agent_topology_summary.csv,archive:results/formal/early_mas_architecture_motivation/summaries/med_agent_topology_summary.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",3795,e222df3b80b0ac56327bdee3fcf73bf310f07b9b7e4bf66fa8d5803b2b46b186
+data/external/results/formal/early_mas_architecture_motivation/summaries/med_agent_topology_summary.metadata.json,archive:results/formal/early_mas_architecture_motivation/summaries/med_agent_topology_summary.metadata.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1931,7795a6b7b35ab48fb594e0ff899c05ad3e27f8e51b479f0d7d2ac4fca8e7afd5
 data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_benchmark_discussion_controls.csv,archive:results/formal/early_mas_architecture_motivation/summaries/med_board_benchmark_discussion_controls.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1850,dd2119db18e7d83e9e8efd5b945aa75c9ec2022210b43cb1f03d6a36bea5cdf8
+data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_judge_behavior_summary.csv,archive:results/formal/early_mas_architecture_motivation/summaries/med_board_judge_behavior_summary.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",776,9b9e2b43580ea0ffccc6909b39690cefc0b9b3f0a615b3d7a4eebed4ea68c3ee
+data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_judge_behavior_summary.metadata.json,archive:results/formal/early_mas_architecture_motivation/summaries/med_board_judge_behavior_summary.metadata.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1931,0033e87951b1ef1a25a332acabcba8a1ab11ca0d6d26081a651f892c7cb8673d
 data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_mechanism_pilot_summary.csv,archive:results/formal/early_mas_architecture_motivation/summaries/med_board_mechanism_pilot_summary.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",2038,7dd4aff44b730a7295ab79107164868f884ff9c9b2bd1353b62d9c1a258afe78
+data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_oracle_gap_summary.csv,archive:results/formal/early_mas_architecture_motivation/summaries/med_board_oracle_gap_summary.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",780,6e870422bd41de783c30a3fae87857f926e341428c9ac1628db73392b58fb8ca
+data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_oracle_gap_summary.metadata.json,archive:results/formal/early_mas_architecture_motivation/summaries/med_board_oracle_gap_summary.metadata.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1931,8385be80a95ddaa805dcea288822a4039535cea154184ea9e78af62c5600b379
+data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_protocol_bootstrap_ci.csv,archive:results/formal/early_mas_architecture_motivation/summaries/med_board_protocol_bootstrap_ci.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",414,75dd67cd93d37a80ccac2dd3b3309b35971730e06060f486eef03ea337f1002f
+data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_protocol_bootstrap_ci.metadata.json,archive:results/formal/early_mas_architecture_motivation/summaries/med_board_protocol_bootstrap_ci.metadata.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1931,478bacea921a2f73bad6114eb4616cca329ff68bf19534b22870212fd2718ba7
+data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_protocol_cost_accuracy_summary.csv,archive:results/formal/early_mas_architecture_motivation/summaries/med_board_protocol_cost_accuracy_summary.csv,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1005,44f47a20ab6246861bbe20b224492cb040cdc5137c65253bb2eba6dd53844310
+data/external/results/formal/early_mas_architecture_motivation/summaries/med_board_protocol_cost_accuracy_summary.metadata.json,archive:results/formal/early_mas_architecture_motivation/summaries/med_board_protocol_cost_accuracy_summary.metadata.json,complete formal research artifact,"raw, intermediate, audit, or summary artifact",1929,6e0bbaf6c4aa5b04821134c1360d6fc61298e0a4ac93bab97bb2f1b5e64cf781
 data/external/results/formal/five_benchmark_cost_efficiency/README.md,archive:results/formal/five_benchmark_cost_efficiency/README.md,complete formal research artifact,"raw, intermediate, audit, or summary artifact",4758,d871b3cabeb8a103203b5cf66be2e8a34f4d2829652a2498b6b8a4c4c5436c51
 data/external/results/formal/five_benchmark_cost_efficiency/archive/source_inputs/mimo_k16_topup_bank/mimo_k16_topup.launchd.log,archive:results/formal/five_benchmark_cost_efficiency/archive/source_inputs/mimo_k16_topup_bank/mimo_k16_topup.launchd.log,complete formal research artifact,"raw, intermediate, audit, or summary artifact",897196,f0c8b7ee3fb57c984b9caee42d68f58f126f2faf7b4d661a443cb1354de3227d
 data/external/results/formal/five_benchmark_cost_efficiency/archive/source_inputs/mimo_k16_topup_bank/mimo_k16_topup.launchd.status,archive:results/formal/five_benchmark_cost_efficiency/archive/source_inputs/mimo_k16_topup_bank/mimo_k16_topup.launchd.status,complete formal research artifact,"raw, intermediate, audit, or summary artifact",25,adde7549dc8ef3f9954df91107df5cff37d630b99ab6a2fa4e332a101a5be0a1

RELEASE_METADATA.json CHANGED Viewed

@@ -2,12 +2,17 @@
   "contains_benchmark_material": true,
   "contains_candidate_text": true,
   "contains_rationales": true,
-  "created_utc": "2026-07-15T13:58:11+00:00",
   "dataset_repository": "Biogod/mas_reasoning",
   "excluded_file_count": 2,
   "expected_repository_visibility": "private",
-  "file_count": 1407,
-  "payload_bytes": 12816725415,
   "release_scope": "full_private",
   "source_symlink_count": 242
 }

   "contains_benchmark_material": true,
   "contains_candidate_text": true,
   "contains_rationales": true,
+  "created_utc": "2026-08-24T09:39:37+00:00",
   "dataset_repository": "Biogod/mas_reasoning",
   "excluded_file_count": 2,
   "expected_repository_visibility": "private",
+  "file_count": 1568,
+  "medical_mas_increment": {
+    "base_revision": "963debcaed4fc46cc7de50f0507ecb296d404a79",
+    "remote_prefix": "data/external/results/formal/early_mas_architecture_motivation",
+    "uploaded_payload_files": 170
+  },
+  "payload_bytes": 12953259548,
   "release_scope": "full_private",
   "source_symlink_count": 242
 }

data/external/results/formal/early_mas_architecture_motivation/README.md CHANGED Viewed

@@ -7,21 +7,57 @@ from black-box communication architectures to a measurable selection-bottleneck
 ## Scope
 - `med-agent` topology experiments compare star, complete, ring and hierarchical communication topologies on MedXpertQA-style medical diagnosis tasks.
-- `med_board` protocol experiments compare single-agent, majority vote, open committee and anonymous evidence board protocols, plus mechanism variants.
 - The central interpretation is historical motivation: more communication and richer architectures produced modest gains at substantial token cost, while oracle readouts showed that correct traces were often already present but not reliably selected.
 ## Generated Files
 - `summaries/med_agent_topology_summary.csv`: topology-level accuracy, round and token summaries for baseline, tuned and fixed-six-round runs.
 - `summaries/med_agent_consensus_bypass_summary.csv`: one-round exits and weighted-majority behavior in the topology experiments.
 - `summaries/med_agent_round_exit_distribution.csv`: exact stopping-round counts and shares by topology and stopping condition.
 - `summaries/med_board_protocol_cost_accuracy_summary.csv`: full 3-seed protocol accuracy and token-cost summary.
 - `summaries/med_board_benchmark_discussion_controls.csv`: MMLU-Pro and GPQA discussion-vs-majority paired controls.
 - `summaries/med_board_oracle_gap_summary.csv`: round-0 oracle versus final accuracy decomposition.
 - `summaries/med_board_judge_behavior_summary.csv`: follow-majority/follow-minority behavior, including majority-wrong/minority-correct cases.
 - `summaries/med_board_mechanism_pilot_summary.csv`: mechanism-pilot condition summaries.
 - `figure_sources/extended_data_fig1_source.csv`: direct source table used by Extended Data Fig. 1.
-- `audits/provenance_and_qa.md`: source paths, row counts and interpretation limits.
 ## Interpretation Boundary

 ## Scope
 - `med-agent` topology experiments compare star, complete, ring and hierarchical communication topologies on MedXpertQA-style medical diagnosis tasks.
+- `med_board` protocol experiments compare single-agent, majority vote, open committee and anonymous evidence board protocols.
 - The central interpretation is historical motivation: more communication and richer architectures produced modest gains at substantial token cost, while oracle readouts showed that correct traces were often already present but not reliably selected.
+## Reproduction
+The formal analysis runs only from repository-local, question-text-free inputs:
+```bash
+python scripts/run_analysis.py --config configs/analyses/early_mas_architecture_motivation.yaml
+```
+The one-time source freeze and analysis-ready import require the protected archive root explicitly:
+```bash
+MEDICAL_MAS_SOURCE_ROOT=/path/to/med_repos_20260604 \
+  python scripts/run_analysis.py --config configs/analyses/medical_mas_source_inventory.yaml
+MEDICAL_MAS_SOURCE_ROOT=/path/to/med_repos_20260604 \
+  python scripts/run_analysis.py --config configs/analyses/medical_mas_analysis_ready_import.yaml
+MEDICAL_MAS_SOURCE_ROOT=/path/to/med_repos_20260604 \
+  python scripts/run_analysis.py --config configs/analyses/medical_mas_code_snapshot.yaml
+MEDICAL_MAS_SOURCE_ROOT=/path/to/med_repos_20260604 \
+  python scripts/run_analysis.py --config configs/analyses/medical_mas_private_release.yaml
+```
+No source path is hard-coded in the implementation or formal outputs.
 ## Generated Files
 - `summaries/med_agent_topology_summary.csv`: topology-level accuracy, round and token summaries for baseline, tuned and fixed-six-round runs.
 - `summaries/med_agent_consensus_bypass_summary.csv`: one-round exits and weighted-majority behavior in the topology experiments.
 - `summaries/med_agent_round_exit_distribution.csv`: exact stopping-round counts and shares by topology and stopping condition.
 - `summaries/med_board_protocol_cost_accuracy_summary.csv`: full 3-seed protocol accuracy and token-cost summary.
+- `summaries/med_board_protocol_bootstrap_ci.csv`: paired question-cluster confidence intervals for the two manuscript comparisons.
 - `summaries/med_board_benchmark_discussion_controls.csv`: MMLU-Pro and GPQA discussion-vs-majority paired controls.
 - `summaries/med_board_oracle_gap_summary.csv`: round-0 oracle versus final accuracy decomposition.
 - `summaries/med_board_judge_behavior_summary.csv`: follow-majority/follow-minority behavior, including majority-wrong/minority-correct cases.
 - `summaries/med_board_mechanism_pilot_summary.csv`: mechanism-pilot condition summaries.
 - `figure_sources/extended_data_fig1_source.csv`: direct source table used by Extended Data Fig. 1.
+- `analysis_ready/`: compact trial tables and execution configs needed for numerical reproduction; it contains no benchmark question text or raw model rationales.
+- `manifests/analysis_ready_manifest.csv`: source and destination hashes for every imported or normalized analysis input.
+- `manifests/source_payload_inventory.csv.gz`: hashes and release paths for the full private raw payload.
+- `manifests/code_snapshot_manifest.csv`: source revisions, source/output hashes and transformations for the sanitized historical code snapshot.
+- `manifests/private_release_shards.csv`: deterministic private-release shard hashes, member counts and compressed sizes.
+- `code_snapshot/`: the sanitized `med-agent` reconstruction and byte-identical `med_board` Git snapshots.
+- `audits/provenance_and_qa.md`: row counts, headline checks and interpretation limits.
+- `audits/code_security_gate.md`: the passed payload and code-snapshot security gates.
+- `audits/private_release_verification.json`: complete member-hash and repeat-build verification for the Git-ignored private payload.
+The raw release shards are written to the Git-ignored `results/tmp/medical_mas_private_release/` directory. They must not be committed to this repository.
+The cross-benchmark discussion-control and mechanism-pilot summaries are retained as secondary historical artifacts. They are not inputs to the manuscript-facing analysis and do not expand the private raw-data migration scope.
 ## Interpretation Boundary

data/external/results/formal/early_mas_architecture_motivation/analysis_ready/configs/med_board/real_llm_full_seed42.yaml ADDED Viewed

	@@ -0,0 +1,33 @@

+experiment:
+  name: real_llm_full_seed42
+  n_cases: 2450
+  seed: 42
+protocol:
+  name: single_agent
+  protocols:
+    - single_agent
+    - majority_vote
+    - open_committee
+    - anonymous_evidence_board
+  n_agents: 5
+  max_rounds: 2
+  visibility_mode: answer_rationale
+  anonymous: false
+  answer_masking: false
+  use_devil_advocate: false
+model:
+  provider: openai
+  model: openai/deepseek-v4-flash-free
+  api_base: http://127.0.0.1:8000/v1
+  api_key_env: MYAPI_DUMMY_KEY
+  temperature: 0.0
+  max_tokens: 60000
+  json_mode: true
+runtime:
+  mock_llm: false
+  concurrency: 200
+data:
+  source: medxpertqa_jsonl
+  path: data/medxpertqa_text_test.jsonl
+output:
+  root_dir: runs

data/external/results/formal/early_mas_architecture_motivation/analysis_ready/configs/med_board/real_llm_full_seed43.yaml ADDED Viewed

	@@ -0,0 +1,33 @@

+experiment:
+  name: real_llm_full_seed43
+  n_cases: 2450
+  seed: 43
+protocol:
+  name: single_agent
+  protocols:
+    - single_agent
+    - majority_vote
+    - open_committee
+    - anonymous_evidence_board
+  n_agents: 5
+  max_rounds: 2
+  visibility_mode: answer_rationale
+  anonymous: false
+  answer_masking: false
+  use_devil_advocate: false
+model:
+  provider: openai
+  model: openai/deepseek-v4-flash-free
+  api_base: http://127.0.0.1:8000/v1
+  api_key_env: MYAPI_DUMMY_KEY
+  temperature: 0.0
+  max_tokens: 60000
+  json_mode: true
+runtime:
+  mock_llm: false
+  concurrency: 200
+data:
+  source: medxpertqa_jsonl
+  path: data/medxpertqa_text_test.jsonl
+output:
+  root_dir: runs

data/external/results/formal/early_mas_architecture_motivation/analysis_ready/configs/med_board/real_llm_full_seed44.yaml ADDED Viewed

	@@ -0,0 +1,33 @@

+experiment:
+  name: real_llm_full_seed44
+  n_cases: 2450
+  seed: 44
+protocol:
+  name: single_agent
+  protocols:
+    - single_agent
+    - majority_vote
+    - open_committee
+    - anonymous_evidence_board
+  n_agents: 5
+  max_rounds: 2
+  visibility_mode: answer_rationale
+  anonymous: false
+  answer_masking: false
+  use_devil_advocate: false
+model:
+  provider: openai
+  model: openai/deepseek-v4-flash-free
+  api_base: http://127.0.0.1:8000/v1
+  api_key_env: MYAPI_DUMMY_KEY
+  temperature: 0.0
+  max_tokens: 60000
+  json_mode: true
+runtime:
+  mock_llm: false
+  concurrency: 200
+data:
+  source: medxpertqa_jsonl
+  path: data/medxpertqa_text_test.jsonl
+output:
+  root_dir: runs

data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/baseline/all_results.csv ADDED Viewed

The diff for this file is too large to render. See raw diff

data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/baseline/summary.json ADDED Viewed

	@@ -0,0 +1,89 @@

+{
+  "single": {
+    "n_cases": 3150,
+    "n_errors": 0,
+    "accuracy_mean": 0.4523809523809524,
+    "accuracy_std": 0.4978062841037845,
+    "rounds_mean": 0.0,
+    "rounds_std": 0.0,
+    "tokens_mean": 5217.271428571428,
+    "tokens_std": 4905.742235504171,
+    "confidence_mean": 0.7631873015873015,
+    "confidence_std": 0.18517969909391874,
+    "convergence_rate": 1.0,
+    "convergence_type_counts": {
+      "single": 3150
+    }
+  },
+  "star": {
+    "n_cases": 3150,
+    "n_errors": 0,
+    "accuracy_mean": 0.4726984126984127,
+    "accuracy_std": 0.4993333324651835,
+    "rounds_mean": 1.126984126984127,
+    "rounds_std": 0.5870428903401118,
+    "tokens_mean": 43825.049523809525,
+    "tokens_std": 42008.32999775721,
+    "confidence_mean": 0.8631968253968254,
+    "confidence_std": 0.07059756726047633,
+    "convergence_rate": 0.9990476190476191,
+    "convergence_type_counts": {
+      "unanimous": 2488,
+      "weighted_majority": 659,
+      "not_converged": 3
+    }
+  },
+  "complete": {
+    "n_cases": 3150,
+    "n_errors": 0,
+    "accuracy_mean": 0.4758730158730159,
+    "accuracy_std": 0.4994968408945377,
+    "rounds_mean": 1.0733333333333333,
+    "rounds_std": 0.4372537641226672,
+    "tokens_mean": 43636.473333333335,
+    "tokens_std": 45277.658958650725,
+    "confidence_mean": 0.8757473015873015,
+    "confidence_std": 0.06699027778165821,
+    "convergence_rate": 1.0,
+    "convergence_type_counts": {
+      "unanimous": 2795,
+      "weighted_majority": 355
+    }
+  },
+  "ring": {
+    "n_cases": 3150,
+    "n_errors": 0,
+    "accuracy_mean": 0.4777777777777778,
+    "accuracy_std": 0.49958523427121443,
+    "rounds_mean": 1.1838095238095239,
+    "rounds_std": 0.8085314943933478,
+    "tokens_mean": 45207.77968253968,
+    "tokens_std": 47501.1089303294,
+    "confidence_mean": 0.8669657142857143,
+    "confidence_std": 0.0694969318054578,
+    "convergence_rate": 0.9968253968253968,
+    "convergence_type_counts": {
+      "unanimous": 2465,
+      "weighted_majority": 675,
+      "not_converged": 10
+    }
+  },
+  "hierarchical": {
+    "n_cases": 3150,
+    "n_errors": 0,
+    "accuracy_mean": 0.473015873015873,
+    "accuracy_std": 0.4993505942157123,
+    "rounds_mean": 1.2368253968253968,
+    "rounds_std": 1.0257603356077607,
+    "tokens_mean": 46365.31904761905,
+    "tokens_std": 52039.323225782915,
+    "confidence_mean": 0.8659126984126985,
+    "confidence_std": 0.06985970378398493,
+    "convergence_rate": 0.9923809523809524,
+    "convergence_type_counts": {
+      "unanimous": 2394,
+      "weighted_majority": 732,
+      "not_converged": 24
+    }
+  }
+}

data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/fixed6/all_results.csv ADDED Viewed

The diff for this file is too large to render. See raw diff

data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/fixed6/summary.json ADDED Viewed

	@@ -0,0 +1,82 @@

+{
+  "single": {
+    "n_cases": 3150,
+    "n_errors": 0,
+    "accuracy_mean": 0.4473015873015873,
+    "accuracy_std": 0.497294063583583,
+    "rounds_mean": 0.0,
+    "rounds_std": 0.0,
+    "tokens_mean": 4954.85746031746,
+    "tokens_std": 4572.679194993217,
+    "confidence_mean": 0.4491650793650794,
+    "confidence_std": 0.030202577175845895,
+    "convergence_rate": 1.0,
+    "convergence_type_counts": {
+      "single": 3150
+    }
+  },
+  "star": {
+    "n_cases": 3150,
+    "n_errors": 0,
+    "accuracy_mean": 0.4822222222222222,
+    "accuracy_std": 0.4997631844533109,
+    "rounds_mean": 6.0,
+    "rounds_std": 0.0,
+    "tokens_mean": 170603.08952380953,
+    "tokens_std": 63394.452189804906,
+    "confidence_mean": 0.9692,
+    "confidence_std": 0.0412673474594445,
+    "convergence_rate": 0.0,
+    "convergence_type_counts": {
+      "not_converged": 3150
+    }
+  },
+  "complete": {
+    "n_cases": 3150,
+    "n_errors": 0,
+    "accuracy_mean": 0.48603174603174604,
+    "accuracy_std": 0.4998842027946899,
+    "rounds_mean": 6.0,
+    "rounds_std": 0.0,
+    "tokens_mean": 242564.97650793652,
+    "tokens_std": 68010.65136104719,
+    "confidence_mean": 0.9801371428571428,
+    "confidence_std": 0.03557990642147404,
+    "convergence_rate": 0.0,
+    "convergence_type_counts": {
+      "not_converged": 3150
+    }
+  },
+  "ring": {
+    "n_cases": 3150,
+    "n_errors": 0,
+    "accuracy_mean": 0.4857142857142857,
+    "accuracy_std": 0.4998752282735986,
+    "rounds_mean": 6.0,
+    "rounds_std": 0.0,
+    "tokens_mean": 175621.92698412697,
+    "tokens_std": 63411.70728232742,
+    "confidence_mean": 0.9749425396825396,
+    "confidence_std": 0.037597653520301325,
+    "convergence_rate": 0.0,
+    "convergence_type_counts": {
+      "not_converged": 3150
+    }
+  },
+  "hierarchical": {
+    "n_cases": 3150,
+    "n_errors": 0,
+    "accuracy_mean": 0.47174603174603175,
+    "accuracy_std": 0.49928033213386164,
+    "rounds_mean": 6.0,
+    "rounds_std": 0.0,
+    "tokens_mean": 164798.6346031746,
+    "tokens_std": 60460.81742642259,
+    "confidence_mean": 0.969691746031746,
+    "confidence_std": 0.039974625507567335,
+    "convergence_rate": 0.0,
+    "convergence_type_counts": {
+      "not_converged": 3150
+    }
+  }
+}

data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/tuned/all_results.csv ADDED Viewed

The diff for this file is too large to render. See raw diff

data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/tuned/summary.json ADDED Viewed

	@@ -0,0 +1,90 @@

+{
+  "single": {
+    "n_cases": 3150,
+    "n_errors": 0,
+    "accuracy_mean": 0.4473015873015873,
+    "accuracy_std": 0.497294063583583,
+    "rounds_mean": 0.0,
+    "rounds_std": 0.0,
+    "tokens_mean": 4880.5203174603175,
+    "tokens_std": 4537.68029257931,
+    "confidence_mean": 0.45020952380952384,
+    "confidence_std": 0.028907610708988347,
+    "convergence_rate": 1.0,
+    "convergence_type_counts": {
+      "single": 3150
+    }
+  },
+  "star": {
+    "n_cases": 3150,
+    "n_errors": 0,
+    "accuracy_mean": 0.4803174603174603,
+    "accuracy_std": 0.49969176988453656,
+    "rounds_mean": 2.0095238095238095,
+    "rounds_std": 1.0292197042688307,
+    "tokens_mean": 65312.631746031744,
+    "tokens_std": 59772.90713464083,
+    "confidence_mean": 0.8558736507936507,
+    "confidence_std": 0.04420131244563625,
+    "convergence_rate": 0.9965079365079365,
+    "convergence_type_counts": {
+      "unanimous_with_conf": 2670,
+      "weighted_majority": 469,
+      "not_converged": 11
+    }
+  },
+  "complete": {
+    "n_cases": 3150,
+    "n_errors": 0,
+    "accuracy_mean": 0.48063492063492064,
+    "accuracy_std": 0.49970417738537004,
+    "rounds_mean": 1.638095238095238,
+    "rounds_std": 0.9512453723787976,
+    "tokens_mean": 63638.094603174606,
+    "tokens_std": 75213.33908212742,
+    "confidence_mean": 0.8655428571428572,
+    "confidence_std": 0.04347797124495981,
+    "convergence_rate": 0.9980952380952381,
+    "convergence_type_counts": {
+      "unanimous_with_conf": 2920,
+      "weighted_majority": 224,
+      "not_converged": 6
+    }
+  },
+  "ring": {
+    "n_cases": 3150,
+    "n_errors": 0,
+    "accuracy_mean": 0.48,
+    "accuracy_std": 0.499679160320077,
+    "rounds_mean": 1.9247619047619047,
+    "rounds_std": 1.0426242283273774,
+    "tokens_mean": 64122.485396825396,
+    "tokens_std": 63833.248813473605,
+    "confidence_mean": 0.8559358730158729,
+    "confidence_std": 0.04147061249810893,
+    "convergence_rate": 0.9971428571428571,
+    "convergence_type_counts": {
+      "unanimous_with_conf": 2676,
+      "weighted_majority": 465,
+      "not_converged": 9
+    }
+  },
+  "hierarchical": {
+    "n_cases": 3150,
+    "n_errors": 0,
+    "accuracy_mean": 0.4774603174603175,
+    "accuracy_std": 0.4995710076262872,
+    "rounds_mean": 2.0714285714285716,
+    "rounds_std": 1.1910175847082758,
+    "tokens_mean": 66152.03047619047,
+    "tokens_std": 65717.94584206482,
+    "confidence_mean": 0.8548107936507936,
+    "confidence_std": 0.04186784252691231,
+    "convergence_rate": 0.9949206349206349,
+    "convergence_type_counts": {
+      "unanimous_with_conf": 2582,
+      "weighted_majority": 552,
+      "not_converged": 16
+    }
+  }
+}

data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_board/gain_bottleneck_report.json ADDED Viewed

	@@ -0,0 +1,681 @@

+{
+  "aeb_compression": {
+    "aeb_win_rate": 0.4865497076023392,
+    "mean_retained_ids_losses": 8.186788154897494,
+    "mean_retained_ids_wins": 8.04326923076923,
+    "mean_retention_losses": 0.3399298375151417,
+    "mean_retention_wins": 0.33387560027370944,
+    "mean_units_losses": 25.11617312072893,
+    "mean_units_wins": 25.07451923076923,
+    "mwmc_sole_correct_rescue_n": 685,
+    "mwmc_sole_correct_rescue_rate": 0.22481751824817517,
+    "n_aeb_losses": 439,
+    "n_aeb_wins": 416,
+    "n_paired": 7350,
+    "n_ties": 6495,
+    "novel_answer_accuracy": 0.2916666666666667,
+    "novel_answer_n": 240,
+    "win_rate_by_retention_decile": [
+      {
+        "aeb_win_rate": 0.46875,
+        "decile": 0,
+        "n": 187,
+        "range": "0.00-0.15"
+      },
+      {
+        "aeb_win_rate": 0.5193798449612403,
+        "decile": 1,
+        "n": 933,
+        "range": "0.15-0.21"
+      },
+      {
+        "aeb_win_rate": 0.515,
+        "decile": 2,
+        "n": 1397,
+        "range": "0.21-0.27"
+      },
+      {
+        "aeb_win_rate": 0.4385026737967914,
+        "decile": 3,
+        "n": 1429,
+        "range": "0.27-0.34"
+      },
+      {
+        "aeb_win_rate": 0.5263157894736842,
+        "decile": 4,
+        "n": 891,
+        "range": "0.34-0.41"
+      },
+      {
+        "aeb_win_rate": 0.484375,
+        "decile": 5,
+        "n": 724,
+        "range": "0.41-0.50"
+      },
+      {
+        "aeb_win_rate": 0.39622641509433965,
+        "decile": 6,
+        "n": 686,
+        "range": "0.50-0.59"
+      },
+      {
+        "aeb_win_rate": 0.4,
+        "decile": 7,
+        "n": 425,
+        "range": "0.59-0.70"
+      },
+      {
+        "aeb_win_rate": 0.5454545454545454,
+        "decile": 8,
+        "n": 344,
+        "range": "0.70-0.83"
+      },
+      {
+        "aeb_win_rate": 0.5416666666666666,
+        "decile": 9,
+        "n": 334,
+        "range": "0.83-1.01"
+      }
+    ]
+  },
+  "confidence_calibration": {
+    "by_protocol": {
+      "anonymous_evidence_board": {
+        "binned_calibration": [
+          {
+            "accuracy": 0.24561403508771928,
+            "bin": "0.0-0.5",
+            "mean_confidence": 0.30000000000000004,
+            "n": 57
+          },
+          {
+            "accuracy": 0.32277745496804183,
+            "bin": "0.5-0.8",
+            "mean_confidence": 0.7074970947123765,
+            "n": 6884
+          },
+          {
+            "accuracy": 0.39944325286359694,
+            "bin": "0.8-0.95",
+            "mean_confidence": 0.8558344361794369,
+            "n": 21913
+          },
+          {
+            "accuracy": 0.47188449848024316,
+            "bin": "0.95-1.01",
+            "mean_confidence": 0.9539247720364741,
+            "n": 7896
+          }
+        ],
+        "mean_confidence_correct": 0.8601202854230378,
+        "mean_confidence_wrong": 0.8403417290673928,
+        "n_confidence_pairs": 36750
+      },
+      "majority_vote": {
+        "binned_calibration": [
+          {
+            "accuracy": 0.23076923076923078,
+            "bin": "0.0-0.5",
+            "mean_confidence": 0.3038461538461539,
+            "n": 52
+          },
+          {
+            "accuracy": 0.3165593914569924,
+            "bin": "0.5-0.8",
+            "mean_confidence": 0.7077091866588647,
+            "n": 6836
+          },
+          {
+            "accuracy": 0.39703119989004443,
+            "bin": "0.8-0.95",
+            "mean_confidence": 0.8555559627983691,
+            "n": 21827
+          },
+          {
+            "accuracy": 0.4760423148724331,
+            "bin": "0.95-1.01",
+            "mean_confidence": 0.9537685127566895,
+            "n": 8035
+          }
+        ],
+        "mean_confidence_correct": 0.8608767982545851,
+        "mean_confidence_wrong": 0.8406905764615316,
+        "n_confidence_pairs": 36750
+      },
+      "open_committee": {
+        "binned_calibration": [
+          {
+            "accuracy": 0.19696969696969696,
+            "bin": "0.0-0.5",
+            "mean_confidence": 0.31287878787878787,
+            "n": 66
+          },
+          {
+            "accuracy": 0.3172144201534675,
+            "bin": "0.5-0.8",
+            "mean_confidence": 0.7081583900390908,
+            "n": 6907
+          },
+          {
+            "accuracy": 0.39371483463702484,
+            "bin": "0.8-0.95",
+            "mean_confidence": 0.8563789396642423,
+            "n": 21861
+          },
+          {
+            "accuracy": 0.4729661445174331,
+            "bin": "0.95-1.01",
+            "mean_confidence": 0.9539274886306215,
+            "n": 7916
+          }
+        ],
+        "mean_confidence_correct": 0.8609467536928891,
+        "mean_confidence_wrong": 0.8404329804009912,
+        "n_confidence_pairs": 36750
+      },
+      "single_agent": {
+        "binned_calibration": [
+          {
+            "accuracy": 0.2,
+            "bin": "0.0-0.5",
+            "mean_confidence": 0.29,
+            "n": 10
+          },
+          {
+            "accuracy": 0.3105542900531511,
+            "bin": "0.5-0.8",
+            "mean_confidence": 0.7089749430523917,
+            "n": 1317
+          },
+          {
+            "accuracy": 0.3990439335306169,
+            "bin": "0.8-0.95",
+            "mean_confidence": 0.8580150239016617,
+            "n": 4393
+          },
+          {
+            "accuracy": 0.4705521472392638,
+            "bin": "0.95-1.01",
+            "mean_confidence": 0.954276073619632,
+            "n": 1630
+          }
+        ],
+        "mean_confidence_correct": 0.8647424087342204,
+        "mean_confidence_wrong": 0.8433559628875311,
+        "n_confidence_pairs": 7350
+      }
+    },
+    "majority_comparison": [
+      {
+        "confidence_weighted_majority_accuracy": 0.39877551020408164,
+        "n": 7350,
+        "protocol": "single_agent",
+        "simple_majority_accuracy": 0.39877551020408164
+      },
+      {
+        "confidence_weighted_majority_accuracy": 0.4089795918367347,
+        "n": 7350,
+        "protocol": "majority_vote",
+        "simple_majority_accuracy": 0.41034013605442177
+      },
+      {
+        "confidence_weighted_majority_accuracy": 0.4042176870748299,
+        "n": 7350,
+        "protocol": "open_committee",
+        "simple_majority_accuracy": 0.40530612244897957
+      },
+      {
+        "confidence_weighted_majority_accuracy": 0.41034013605442177,
+        "n": 7350,
+        "protocol": "anonymous_evidence_board",
+        "simple_majority_accuracy": 0.4107482993197279
+      }
+    ],
+    "mwmc_confidence_comparison": [
+      {
+        "mean_majority_confidence": 0.0,
+        "mean_minority_confidence": 0.0,
+        "n_cases": 7350,
+        "n_majority_agents": 0,
+        "n_minority_agents": 0,
+        "protocol": "single_agent"
+      },
+      {
+        "mean_majority_confidence": 0.8280207100591717,
+        "mean_minority_confidence": 0.8063287671232877,
+        "n_cases": 7350,
+        "n_majority_agents": 3380,
+        "n_minority_agents": 1460,
+        "protocol": "majority_vote"
+      },
+      {
+        "mean_majority_confidence": 0.8291143270622287,
+        "mean_minority_confidence": 0.8057752200406229,
+        "n_cases": 7350,
+        "n_majority_agents": 3455,
+        "n_minority_agents": 1477,
+        "protocol": "open_committee"
+      },
+      {
+        "mean_majority_confidence": 0.8268896713615024,
+        "mean_minority_confidence": 0.8034575569358179,
+        "n_cases": 7350,
+        "n_majority_agents": 3408,
+        "n_minority_agents": 1449,
+        "protocol": "anonymous_evidence_board"
+      }
+    ]
+  },
+  "difficulty_strata": [
+    {
+      "global_all_correct": 1.0,
+      "global_all_correct_n": 2217,
+      "global_all_wrong": 0.0,
+      "global_all_wrong_n": 3510,
+      "global_single_correct_multi_wrong": 1.0,
+      "global_single_correct_multi_wrong_n": 157,
+      "global_single_wrong_multi_correct": 0.0,
+      "global_single_wrong_multi_correct_n": 486,
+      "protocol": "single_agent",
+      "protocol_no_round0_agent_correct": 0.0,
+      "protocol_no_round0_agent_correct_n": 4419
+    },
+    {
+      "global_all_correct": 1.0,
+      "global_all_correct_n": 2217,
+      "global_all_wrong": 0.0,
+      "global_all_wrong_n": 3510,
+      "global_single_correct_multi_wrong": 0.0,
+      "global_single_correct_multi_wrong_n": 157,
+      "global_single_wrong_multi_correct": 0.7962962962962963,
+      "global_single_wrong_multi_correct_n": 486,
+      "protocol": "majority_vote",
+      "protocol_majority_correct": 1.0,
+      "protocol_majority_correct_n": 2918,
+      "protocol_majority_wrong_minority_correct": 0.04060434372049103,
+      "protocol_majority_wrong_minority_correct_n": 1059,
+      "protocol_no_round0_agent_correct": 0.0,
+      "protocol_no_round0_agent_correct_n": 3275
+    },
+    {
+      "global_all_correct": 1.0,
+      "global_all_correct_n": 2217,
+      "global_all_wrong": 0.0,
+      "global_all_wrong_n": 3510,
+      "global_single_correct_multi_wrong": 0.0,
+      "global_single_correct_multi_wrong_n": 157,
+      "global_single_wrong_multi_correct": 0.8888888888888888,
+      "global_single_wrong_multi_correct_n": 486,
+      "protocol": "open_committee",
+      "protocol_majority_correct": 0.9735006973500697,
+      "protocol_majority_correct_n": 2868,
+      "protocol_majority_wrong_minority_correct": 0.21691176470588236,
+      "protocol_majority_wrong_minority_correct_n": 1088,
+      "protocol_no_round0_agent_correct": 0.0018275967103259215,
+      "protocol_no_round0_agent_correct_n": 3283
+    },
+    {
+      "global_all_correct": 1.0,
+      "global_all_correct_n": 2217,
+      "global_all_wrong": 0.0,
+      "global_all_wrong_n": 3510,
+      "global_single_correct_multi_wrong": 0.0,
+      "global_single_correct_multi_wrong_n": 157,
+      "global_single_wrong_multi_correct": 0.808641975308642,
+      "global_single_wrong_multi_correct_n": 486,
+      "protocol": "anonymous_evidence_board",
+      "protocol_majority_correct": 0.907128169979438,
+      "protocol_majority_correct_n": 2918,
+      "protocol_majority_wrong_minority_correct": 0.29896907216494845,
+      "protocol_majority_wrong_minority_correct_n": 1067,
+      "protocol_no_round0_agent_correct": 0.02144607843137255,
+      "protocol_no_round0_agent_correct_n": 3264
+    }
+  ],
+  "discussion_transitions": {
+    "agent_transitions": {
+      "correct_to_correct": 12550,
+      "correct_to_wrong": 2005,
+      "wrong_to_correct": 2680,
+      "wrong_to_wrong": 19515
+    },
+    "case_damage_count": 755,
+    "case_rescue_count": 1106,
+    "final_accuracy_when_correct_agents_decrease": 0.03329065300896287,
+    "final_accuracy_when_correct_agents_increase": 0.9501718213058419,
+    "final_error_rate_when_correct_agents_decrease": 0.9667093469910372,
+    "n_cases": 7350,
+    "n_correct_agents_decrease": 781,
+    "n_correct_agents_increase": 1164
+  },
+  "error_correlation": {
+    "all_correct_rate": 0.3016326530612245,
+    "all_wrong_rate": 0.4775510204081633,
+    "n_correct_protocols_histogram": [
+      {
+        "count": 3510,
+        "n_correct": 0
+      },
+      {
+        "count": 580,
+        "n_correct": 1
+      },
+      {
+        "count": 447,
+        "n_correct": 2
+      },
+      {
+        "count": 596,
+        "n_correct": 3
+      },
+      {
+        "count": 2217,
+        "n_correct": 4
+      }
+    ],
+    "n_protocols": 4,
+    "pairwise": [
+      {
+        "agreement": 0.8801360544217687,
+        "both_correct": 2534,
+        "both_wrong": 3935,
+        "error_correlation": 0.8170681063122923,
+        "error_correlation_n": 4816,
+        "n": 7350,
+        "pair": "single_agent vs majority_vote",
+        "single_agent_correct_majority_vote_wrong": 397,
+        "single_agent_wrong_majority_vote_correct": 484
+      },
+      {
+        "agreement": 0.8755102040816326,
+        "both_correct": 2559,
+        "both_wrong": 3876,
+        "error_correlation": 0.8090169067000627,
+        "error_correlation_n": 4791,
+        "n": 7350,
+        "pair": "single_agent vs open_committee",
+        "single_agent_correct_open_committee_wrong": 372,
+        "single_agent_wrong_open_committee_correct": 543
+      },
+      {
+        "agreement": 0.8546938775510204,
+        "both_correct": 2471,
+        "both_wrong": 3811,
+        "error_correlation": 0.7811026849764297,
+        "error_correlation_n": 4879,
+        "n": 7350,
+        "pair": "single_agent vs anonymous_evidence_board",
+        "single_agent_correct_anonymous_evidence_board_wrong": 460,
+        "single_agent_wrong_anonymous_evidence_board_correct": 608
+      },
+      {
+        "agreement": 0.9096598639455782,
+        "both_correct": 2728,
+        "both_wrong": 3958,
+        "error_correlation": 0.8563392470791865,
+        "error_correlation_n": 4622,
+        "majority_vote_correct_open_committee_wrong": 290,
+        "majority_vote_wrong_open_committee_correct": 374,
+        "n": 7350,
+        "pair": "majority_vote vs open_committee"
+      },
+      {
+        "agreement": 0.873061224489796,
+        "both_correct": 2582,
+        "both_wrong": 3835,
+        "error_correlation": 0.8043204697986577,
+        "error_correlation_n": 4768,
+        "majority_vote_correct_anonymous_evidence_board_wrong": 436,
+        "majority_vote_wrong_anonymous_evidence_board_correct": 497,
+        "n": 7350,
+        "pair": "majority_vote vs anonymous_evidence_board"
+      },
+      {
+        "agreement": 0.8836734693877552,
+        "both_correct": 2663,
+        "both_wrong": 3832,
+        "error_correlation": 0.8175805419244719,
+        "error_correlation_n": 4687,
+        "n": 7350,
+        "open_committee_correct_anonymous_evidence_board_wrong": 439,
+        "open_committee_wrong_anonymous_evidence_board_correct": 416,
+        "pair": "open_committee vs anonymous_evidence_board"
+      }
+    ]
+  },
+  "judge_behavior": {
+    "anonymous_evidence_board": {
+      "full_dataset": {
+        "follows_highest_confidence": 0.8308843537414966,
+        "follows_majority": 0.8315646258503402,
+        "follows_minority": 0.13578231292517007,
+        "novel_answer": 0.0326530612244898
+      },
+      "majority_wrong_minority_correct_subset": {
+        "follows_highest_confidence": 0.5904404873477038,
+        "follows_majority": 0.5960637300843487,
+        "follows_minority": 0.37488284910965325,
+        "novel_answer": 0.029053420805998126
+      }
+    },
+    "open_committee": {
+      "full_dataset": {
+        "follows_highest_confidence": 0.9193197278911565,
+        "follows_majority": 0.9165986394557823,
+        "follows_minority": 0.0817687074829932,
+        "novel_answer": 0.0016326530612244899
+      },
+      "majority_wrong_minority_correct_subset": {
+        "follows_highest_confidence": 0.7389705882352942,
+        "follows_majority": 0.7325367647058824,
+        "follows_minority": 0.265625,
+        "novel_answer": 0.001838235294117647
+      }
+    }
+  },
+  "metadata": {
+    "data_path": "TsinghuaC3I/MedXpertQA@7e7c465a68eb2b866926bfa59c8c9d17a8daba65/Text/test.jsonl#sha256=07ea977f7e326909b51c883b65f6ad7f399fc6db07408ef3630f4e3048ec71b2",
+    "n_case_ids": 2450,
+    "n_indexed": 29400,
+    "n_raw_records": 29400,
+    "n_rows": 29400,
+    "run_dirs": [
+      "medical_mas/raw/med_board/seed42",
+      "medical_mas/raw/med_board/seed43",
+      "medical_mas/raw/med_board/seed44"
+    ],
+    "seeds": [
+      42,
+      43,
+      44
+    ]
+  },
+  "oracle_gap_decomposition": [
+    {
+      "final_accuracy": 0.39877551020408164,
+      "fraction_majority_correct": 0.39877551020408164,
+      "fraction_majority_wrong_minority_correct": 0.0,
+      "fraction_no_agent_correct": 0.6012244897959184,
+      "n": 7350,
+      "oracle_gap": 0.0,
+      "protocol": "single_agent",
+      "round0_oracle_accuracy": 0.39877551020408164
+    },
+    {
+      "final_accuracy": 0.41061224489795917,
+      "fraction_majority_correct": 0.41034013605442177,
+      "fraction_majority_wrong_minority_correct": 0.14408163265306123,
+      "fraction_no_agent_correct": 0.445578231292517,
+      "n": 7350,
+      "oracle_gap": 0.1438095238095238,
+      "protocol": "majority_vote",
+      "round0_oracle_accuracy": 0.5544217687074829
+    },
+    {
+      "final_accuracy": 0.4220408163265306,
+      "fraction_majority_correct": 0.40530612244897957,
+      "fraction_majority_wrong_minority_correct": 0.14802721088435375,
+      "fraction_no_agent_correct": 0.44666666666666666,
+      "n": 7350,
+      "oracle_gap": 0.13129251700680272,
+      "protocol": "open_committee",
+      "round0_oracle_accuracy": 0.5533333333333333
+    },
+    {
+      "final_accuracy": 0.41891156462585033,
+      "fraction_majority_correct": 0.4107482993197279,
+      "fraction_majority_wrong_minority_correct": 0.14517006802721089,
+      "fraction_no_agent_correct": 0.4440816326530612,
+      "n": 7350,
+      "oracle_gap": 0.13700680272108842,
+      "protocol": "anonymous_evidence_board",
+      "round0_oracle_accuracy": 0.5559183673469388
+    }
+  ],
+  "round0_correct_agent_count": [
+    {
+      "final_accuracy": 0.0,
+      "n": 4419,
+      "n_correct_agents": 0,
+      "protocol": "single_agent"
+    },
+    {
+      "final_accuracy": 1.0,
+      "n": 2931,
+      "n_correct_agents": 1,
+      "protocol": "single_agent"
+    },
+    {
+      "final_accuracy": 0.0,
+      "n": 0,
+      "n_correct_agents": 2,
+      "protocol": "single_agent"
+    },
+    {
+      "final_accuracy": 0.0,
+      "n": 0,
+      "n_correct_agents": 3,
+      "protocol": "single_agent"
+    },
+    {
+      "final_accuracy": 0.0,
+      "n": 0,
+      "n_correct_agents": 4,
+      "protocol": "single_agent"
+    },
+    {
+      "final_accuracy": 0.0,
+      "n": 0,
+      "n_correct_agents": 5,
+      "protocol": "single_agent"
+    },
+    {
+      "final_accuracy": 0.0,
+      "n": 3275,
+      "n_correct_agents": 0,
+      "protocol": "majority_vote"
+    },
+    {
+      "final_accuracy": 0.0015151515151515152,
+      "n": 660,
+      "n_correct_agents": 1,
+      "protocol": "majority_vote"
+    },
+    {
+      "final_accuracy": 0.19919517102615694,
+      "n": 497,
+      "n_correct_agents": 2,
+      "protocol": "majority_vote"
+    },
+    {
+      "final_accuracy": 1.0,
+      "n": 513,
+      "n_correct_agents": 3,
+      "protocol": "majority_vote"
+    },
+    {
+      "final_accuracy": 1.0,
+      "n": 551,
+      "n_correct_agents": 4,
+      "protocol": "majority_vote"
+    },
+    {
+      "final_accuracy": 1.0,
+      "n": 1854,
+      "n_correct_agents": 5,
+      "protocol": "majority_vote"
+    },
+    {
+      "final_accuracy": 0.0018275967103259215,
+      "n": 3283,
+      "n_correct_agents": 0,
+      "protocol": "open_committee"
+    },
+    {
+      "final_accuracy": 0.10413694721825963,
+      "n": 701,
+      "n_correct_agents": 1,
+      "protocol": "open_committee"
+    },
+    {
+      "final_accuracy": 0.463855421686747,
+      "n": 498,
+      "n_correct_agents": 2,
+      "protocol": "open_committee"
+    },
+    {
+      "final_accuracy": 0.8589743589743589,
+      "n": 468,
+      "n_correct_agents": 3,
+      "protocol": "open_committee"
+    },
+    {
+      "final_accuracy": 0.9816849816849816,
+      "n": 546,
+      "n_correct_agents": 4,
+      "protocol": "open_committee"
+    },
+    {
+      "final_accuracy": 1.0,
+      "n": 1854,
+      "n_correct_agents": 5,
+      "protocol": "open_committee"
+    },
+    {
+      "final_accuracy": 0.02144607843137255,
+      "n": 3264,
+      "n_correct_agents": 0,
+      "protocol": "anonymous_evidence_board"
+    },
+    {
+      "final_accuracy": 0.2259475218658892,
+      "n": 686,
+      "n_correct_agents": 1,
+      "protocol": "anonymous_evidence_board"
+    },
+    {
+      "final_accuracy": 0.42946058091286304,
+      "n": 482,
+      "n_correct_agents": 2,
+      "protocol": "anonymous_evidence_board"
+    },
+    {
+      "final_accuracy": 0.6993603411513859,
+      "n": 469,
+      "n_correct_agents": 3,
+      "protocol": "anonymous_evidence_board"
+    },
+    {
+      "final_accuracy": 0.8330494037478705,
+      "n": 587,
+      "n_correct_agents": 4,
+      "protocol": "anonymous_evidence_board"
+    },
+    {
+      "final_accuracy": 0.9828141783029001,
+      "n": 1862,
+      "n_correct_agents": 5,
+      "protocol": "anonymous_evidence_board"
+    }
+  ]
+}

data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_board/legacy_3seed_summary.md ADDED Viewed

	@@ -0,0 +1,79 @@

+# Real LLM Full-Dataset 3-Seed Summary
+## Run metadata
+- Total cases: `2450`
+- Seeds: `42`, `43`, `44`
+- Protocols: `single_agent`, `majority_vote`, `open_committee`, `anonymous_evidence_board`
+- Rows per seed: `9800` (`2450` cases × 4 protocols)
+- Total analyzed rows: `29400`
+- Full-run configs:
+  - `configs/experiments/real_llm_full_seed42.yaml`
+  - `configs/experiments/real_llm_full_seed43.yaml`
+  - `configs/experiments/real_llm_full_seed44.yaml`
+## QA status
+All three full-dataset seeds passed `scripts/qa_real_smoke.py --n-cases 2450`.
+| seed | QA | rows | failed cases | invalid final | invalid agent outputs | silent fallback | retry total | retry failure | normalization warnings | AEB leakage |
+|---:|---|---:|---:|---:|---:|---:|---:|---:|---:|---|
+| 42 | PASS | 9800 | 0 | 0 | 0 | 0 | 46 | 0 | 4 | PASS |
+| 43 | PASS | 9800 | 0 | 0 | 0 | 0 | 59 | 0 | 0 | PASS |
+| 44 | PASS | 9800 | 0 | 0 | 0 | 0 | 47 | 0 | 2 | PASS |
+Aggregate QA totals:
+| metric | value |
+|---|---:|
+| final decisions valid | 29400 / 29400 |
+| agent outputs valid | 161700 / 161700 |
+| failed cases | 0 |
+| invalid final decisions | 0 |
+| invalid agent outputs | 0 |
+| silent fallbacks | 0 |
+| retry total | 152 |
+| retry success | 152 |
+| retry failure | 0 |
+| normalization warnings | 6 |
+AEB leakage checks passed for persisted evidence boards, retained evidence IDs, and reconstructed judge prompts in all three seeds. The QA checks found no `source_agent`, visible source, `supports_option`, `contradicts_option`, agent identity in evidence IDs, agent diagnoses section, answer field leakage, or supports-label leakage.
+## Accuracy by seed and protocol
+| seed | single_agent | majority_vote | open_committee | anonymous_evidence_board |
+|---:|---:|---:|---:|---:|
+| 42 | 0.4008 | 0.4086 | 0.4220 | 0.4196 |
+| 43 | 0.3943 | 0.4131 | 0.4245 | 0.4208 |
+| 44 | 0.4012 | 0.4102 | 0.4196 | 0.4163 |
+## Mean ± SD across seeds
+| protocol | mean accuracy | SD | mean gain vs single_agent |
+|---|---:|---:|---:|
+| single_agent | 0.3988 | 0.0039 | 0.0000 |
+| majority_vote | 0.4106 | 0.0023 | +0.0118 |
+| open_committee | 0.4220 | 0.0024 | +0.0233 |
+| anonymous_evidence_board | 0.4189 | 0.0023 | +0.0201 |
+## Token usage and cost tradeoff
+Token counts are used as the cost proxy because exact upstream dollar pricing is not encoded in these run artifacts.
+| protocol | calls | prompt tokens | completion tokens | total tokens | token ratio vs single | accuracy gain vs single | gain per extra 1M tokens |
+|---|---:|---:|---:|---:|---:|---:|---:|
+| single_agent | 7359 | 5260686 | 17650380 | 22911066 | 1.00× | 0.0000 | N/A |
+| majority_vote | 36782 | 26298327 | 87302201 | 113600528 | 4.96× | +0.0118 | 0.000131 |
+| open_committee | 80921 | 102389632 | 143362727 | 245752359 | 10.73× | +0.0233 | 0.000104 |
+| anonymous_evidence_board | 44140 | 36113815 | 102915171 | 139028986 | 6.07× | +0.0201 | 0.000173 |
+Cost interpretation:
+- `open_committee` has the best raw accuracy but uses the most calls and tokens: about `10.73×` single-agent token usage.
+- `anonymous_evidence_board` is very close to `open_committee` in raw accuracy while using about `56.6%` of open-committee tokens (`139.0M` vs `245.8M`).
+- `anonymous_evidence_board` has the strongest token-normalized gain among the multi-agent protocols in this run.
+- `majority_vote` is cheaper than AEB and open committee, but it also has the smallest accuracy lift among the multi-agent methods.
+## Conclusion
+`open_committee` has the highest raw accuracy on the full-dataset 3-seed run (`0.4220 ± 0.0024`). `anonymous_evidence_board` is very close (`0.4189 ± 0.0023`), lower-cost than open committee, and leakage-safe under the current QA checks. All multi-agent methods modestly outperform `single_agent`, with mean gains of `+0.0118` for `majority_vote`, `+0.0233` for `open_committee`, and `+0.0201` for `anonymous_evidence_board`.

data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_board/oracle_trials.csv ADDED Viewed

The diff for this file is too large to render. See raw diff

data/external/results/formal/early_mas_architecture_motivation/analysis_ready/med_board/protocol_results.csv ADDED Viewed

The diff for this file is too large to render. See raw diff

data/external/results/formal/early_mas_architecture_motivation/audits/code_security_gate.md ADDED Viewed

	@@ -0,0 +1,27 @@

+# Medical MAS Code Security Gate
+## Payload scan
+The selected experimental outputs passed the automated credential-pattern scan:
+- scanned files: 47,300
+- scanned bytes: 1,126,934,115
+- credential-pattern matches: 0
+- excluded files: three macOS `.DS_Store` files
+The machine-readable result is `payload_security_scan.json`. File hashes and release-relative paths are recorded in `manifests/source_payload_inventory.csv.gz`.
+## Code snapshot status
+The code-snapshot gate passed. The snapshot contains 109 files and 1,338,100 bytes, with zero credential-pattern matches:
+- 75 `med_board` files are byte-identical Git blobs from the provenance-labelled execution and analysis revisions.
+- 34 `med-agent` files preserve the archived worktree and are labelled `reconstructed_worktree`.
+- 32 `med-agent` files are byte-identical copies.
+- Two runtime configuration files were sanitized in the copied snapshot. Their embedded credential fallback was removed, and the reconstructed configuration now requires `OPENAI_API_KEY` from the environment.
+The original archive was not edited. Source and output hashes, revisions and transformations are recorded in `manifests/code_snapshot_manifest.csv`; the machine-readable scan is `code_snapshot_security_scan.json`. If the historical fallback was ever an active credential, rotation or revocation remains an operational security follow-up outside this repository.
+## Release effect
+The numerical archive and the code snapshot both pass their repository security gates. These checks do not change the provenance boundary: `med_board` remains tied to inferred historical commits, whereas `med-agent` remains a sanitized reconstruction of an incompletely tracked worktree.

data/external/results/formal/early_mas_architecture_motivation/audits/code_snapshot_security_scan.json ADDED Viewed

	@@ -0,0 +1,13 @@

+{
+  "match_file_count": 0,
+  "matched_paths": {},
+  "scanned_bytes": 1338100,
+  "scanned_file_count": 109,
+  "secret_pattern_names": [
+    "assigned_api_key",
+    "bearer_token",
+    "huggingface_token",
+    "openai_style_token"
+  ],
+  "status": "pass"
+}

data/external/results/formal/early_mas_architecture_motivation/audits/code_snapshot_security_scan.metadata.json ADDED Viewed

	@@ -0,0 +1,17 @@

+{
+  "artifact_kind": "analysis_auxiliary",
+  "command_mode": "medical_mas_code_snapshot",
+  "config_hash": "7ca1a29be00a",
+  "config_path": "configs/analyses/medical_mas_code_snapshot.yaml",
+  "evidence_status": "complete",
+  "generated_at_utc": "2026-08-24T09:18:43.517885+00:00",
+  "git_commit": "b9585011871af839e163a9a9f8612b30c2a6f391",
+  "input_paths": [
+    "results/formal/early_mas_architecture_motivation/manifests/source_repositories.yaml"
+  ],
+  "output_paths": [
+    "results/formal/early_mas_architecture_motivation/manifests/code_snapshot_manifest.csv",
+    "results/formal/early_mas_architecture_motivation/audits/code_snapshot_security_scan.json"
+  ],
+  "run_class": "formal"
+}

data/external/results/formal/early_mas_architecture_motivation/audits/payload_security_scan.json ADDED Viewed

	@@ -0,0 +1,18 @@

+{
+  "excluded_file_count": 3,
+  "excluded_paths": [
+    "med-agent/results_concurrency50/.DS_Store",
+    "med-agent/results_concurrency50_fixed6/.DS_Store",
+    "med-agent/results_concurrency50_tuned/.DS_Store"
+  ],
+  "match_file_count": 0,
+  "matched_paths": {},
+  "scanned_bytes": 1126934115,
+  "scanned_file_count": 47300,
+  "secret_pattern_names": [
+    "assigned_api_key",
+    "bearer_token",
+    "huggingface_token",
+    "openai_style_token"
+  ]
+}

data/external/results/formal/early_mas_architecture_motivation/audits/payload_security_scan.metadata.json ADDED Viewed

	@@ -0,0 +1,17 @@

+{
+  "artifact_kind": "analysis_auxiliary",
+  "command_mode": "medical_mas_source_inventory",
+  "config_hash": "ce88ed63583f",
+  "config_path": "configs/analyses/medical_mas_source_inventory.yaml",
+  "evidence_status": "complete",
+  "generated_at_utc": "2026-08-24T08:47:30.911805+00:00",
+  "git_commit": "e8971f3a5a5b5fb3f2d7277ae928520d401d1e62",
+  "input_paths": [
+    "docs/migration/medical_mas_migration_plan.md"
+  ],
+  "output_paths": [
+    "results/formal/early_mas_architecture_motivation/manifests/source_payload_inventory.csv.gz",
+    "results/formal/early_mas_architecture_motivation/audits/payload_security_scan.json"
+  ],
+  "run_class": "formal"
+}

data/external/results/formal/early_mas_architecture_motivation/audits/private_release_verification.json ADDED Viewed

	@@ -0,0 +1,13 @@

+{
+  "all_archive_hashes_recorded": true,
+  "all_member_hashes_match": true,
+  "archive_count": 7,
+  "inventory_sha256": "f6243099b9c19e1596ff428e873827948671760f8c6627b07e2a97f3303b25ef",
+  "member_count": 47300,
+  "previous_manifest_present": true,
+  "rebuild_hashes_match_previous": true,
+  "release_root": "results/tmp/medical_mas_private_release",
+  "shard_target_bytes": 268435456,
+  "status": "pass",
+  "uncompressed_bytes": 1126934115
+}

data/external/results/formal/early_mas_architecture_motivation/audits/private_release_verification.metadata.json ADDED Viewed

	@@ -0,0 +1,19 @@

+{
+  "artifact_kind": "analysis_auxiliary",
+  "command_mode": "medical_mas_private_release",
+  "config_hash": "f014d8afffa1",
+  "config_path": "configs/analyses/medical_mas_private_release.yaml",
+  "evidence_status": "complete",
+  "generated_at_utc": "2026-08-24T09:35:15.212556+00:00",
+  "git_commit": "adbdb91e927e374f7fcfc9c1c46c054a03fda95b",
+  "input_paths": [
+    "results/formal/early_mas_architecture_motivation/manifests/source_payload_inventory.csv.gz",
+    "results/formal/early_mas_architecture_motivation/manifests/source_repositories.yaml",
+    "results/formal/early_mas_architecture_motivation/manifests/code_snapshot_manifest.csv"
+  ],
+  "output_paths": [
+    "results/formal/early_mas_architecture_motivation/manifests/private_release_shards.csv",
+    "results/formal/early_mas_architecture_motivation/audits/private_release_verification.json"
+  ],
+  "run_class": "formal"
+}

data/external/results/formal/early_mas_architecture_motivation/audits/provenance_and_qa.md CHANGED Viewed

@@ -1,27 +1,25 @@
-# Provenance And QA
-## Source Archive
-- Archive root: `/Users/jay/research/evole-meme/med_repos_20260604`
-- Generated into: `results/formal/early_mas_architecture_motivation`
-## Row Counts
-- med-agent topology summary rows: `15`
-- med-agent baseline topology rows: `5`
-- med-board protocol summary rows: `4`
-- benchmark discussion-control rows: `8`
-- mechanism-pilot condition rows: `8`
-## Primary Source Families
-- `med-agent/results_concurrency50/`: baseline topology sweep, `15,750` rows in `all_results.csv`.
-- `med-agent/results_concurrency50_tuned/`: tuned consensus topology sweep, `15,750` rows.
-- `med-agent/results_concurrency50_fixed6/`: fixed-six-round topology sweep, `15,750` rows.
-- `med_board/reports/real_llm_full_3seed_summary.md`: full 3-seed protocol summary, `29,400` analyzed rows.
-- `med_board/runs/gain_bottleneck_analysis/gain_bottleneck_report.json`: oracle-gap and follow-majority decomposition.
-- `med_board/runs/mechanism_pilot_500_analysis/mechanism_pilot_report.json`: mechanism-pilot condition summary, `12,000` rows.
-## Use Limits
-These results motivate the selection-bottleneck question. They should not be mixed with the primary five-benchmark selector-boundary evidence as if they were generated under the same protocol.

+# Medical MAS Formal Analysis
+## Coverage
+- med-agent trial rows: 47,250 across three run families
+- med-agent topology summary rows: 15
+- med-board protocol rows: 29,400 across 2,450 questions and three seeds
+- med-board protocol summary rows: 4
+- oracle-gap rows: 4
+- judge-behaviour rows: 4
+- oracle and judge inputs: 29,400 text-free case-protocol-seed records derived from raw outputs
+- legacy cross-check: raw-derived oracle and judge summaries exactly match the archived gain-bottleneck report
+## Manuscript-facing checks
+- single agent: 0.398776 accuracy, 1.0000x token ratio
+- majority vote: 0.410612 accuracy, 4.9583x token ratio
+- open committee: 0.422041 accuracy, 10.7264x token ratio
+- anonymous evidence board: 0.418912 accuracy, 6.0682x token ratio
+- open committee versus majority vote: 1.1429 percentage points, 95% CI 0.4218 to 1.9048 (10,000 question-cluster resamples, seed 46)
+- anonymous evidence board versus majority vote: 0.8299 percentage points, 95% CI -0.0544 to 1.7415 (10,000 question-cluster resamples, seed 49)
+## Evidence boundary
+These archived experiments provide historical motivation for the selection bottleneck in Fig. 1a and Extended Data Fig. 1. They are not part of the later fixed-pool selector-boundary experiment family.

data/external/results/formal/early_mas_architecture_motivation/audits/provenance_and_qa.metadata.json ADDED Viewed

	@@ -0,0 +1,29 @@

+{
+  "artifact_kind": "analysis_auxiliary",
+  "command_mode": "early_mas_architecture_motivation",
+  "config_hash": "c27e24558091",
+  "config_path": "configs/analyses/early_mas_architecture_motivation.yaml",
+  "evidence_status": "complete",
+  "generated_at_utc": "2026-08-24T09:11:48.071341+00:00",
+  "git_commit": "67fc30b663ddd160845baa9a104ba9ca1a27c1f7",
+  "input_paths": [
+    "results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/baseline/all_results.csv",
+    "results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/tuned/all_results.csv",
+    "results/formal/early_mas_architecture_motivation/analysis_ready/med_agent/fixed6/all_results.csv",
+    "results/formal/early_mas_architecture_motivation/analysis_ready/med_board/protocol_results.csv",
+    "results/formal/early_mas_architecture_motivation/analysis_ready/med_board/oracle_trials.csv",
+    "results/formal/early_mas_architecture_motivation/analysis_ready/med_board/gain_bottleneck_report.json"
+  ],
+  "output_paths": [
+    "results/formal/early_mas_architecture_motivation/summaries/med_board_protocol_cost_accuracy_summary.csv",
+    "results/formal/early_mas_architecture_motivation/audits/provenance_and_qa.md",
+    "results/formal/early_mas_architecture_motivation/source_manifest.csv",
+    "results/formal/early_mas_architecture_motivation/summaries/med_agent_topology_summary.csv",
+    "results/formal/early_mas_architecture_motivation/summaries/med_agent_round_exit_distribution.csv",
+    "results/formal/early_mas_architecture_motivation/summaries/med_board_protocol_bootstrap_ci.csv",
+    "results/formal/early_mas_architecture_motivation/summaries/med_board_oracle_gap_summary.csv",
+    "results/formal/early_mas_architecture_motivation/summaries/med_board_judge_behavior_summary.csv",
+    "results/formal/early_mas_architecture_motivation/figure_sources/extended_data_fig1_source.csv"
+  ],
+  "run_class": "formal"
+}

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/pyproject.toml ADDED Viewed

	@@ -0,0 +1,15 @@

+[project]
+name = "med-agent"
+version = "0.1.0"
+requires-python = ">=3.12"
+dependencies = [
+    "httpx>=0.20.0",
+    "datasets>=3.0.0",
+    "pydantic>=2.13.4",
+]
+[project.optional-dependencies]
+dev = [
+    "pytest>=8.0.0",
+    "pytest-asyncio>=0.24.0",
+]

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/pytest.ini ADDED Viewed

	@@ -0,0 +1,5 @@

+[pytest]
+testpaths = tests
+markers =
+    integration: tests that require external services (LLM proxy, dataset download)
+asyncio_mode = auto

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/__init__.py ADDED Viewed

	@@ -0,0 +1 @@


1	+ """MAS topology comparison experiment."""

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/__main__.py ADDED Viewed

	@@ -0,0 +1,194 @@

+# src/med_agent/__main__.py
+import argparse
+import logging
+import os
+import sys
+from pathlib import Path
+from .config import ExperimentConfig
+from .runner import run_experiment
+from .metrics.collector import save_results_csv
+from .metrics.analyzer import aggregate_results, print_summary
+def main():
+    parser = argparse.ArgumentParser(description="MAS Topology Comparison Experiment")
+    parser.add_argument("--model", default="deepseek-v4-flash-free")
+    parser.add_argument("--api-base", default="http://localhost:8000/v1")
+    parser.add_argument("--api-key", default=os.environ.get("OPENAI_API_KEY", ""))
+    parser.add_argument("--temperature", type=float, default=1.0)
+    parser.add_argument("--n-cases", type=int, default=50, help="Number of cases to run (0 = all)")
+    parser.add_argument("--n-trials", type=int, default=3, help="Trials per case per topology")
+    parser.add_argument("--max-rounds", type=int, default=10)
+    parser.add_argument("--topologies", nargs="+",
+                        choices=["single", "star", "complete", "ring", "hierarchical"],
+                        default=["star", "complete", "ring", "hierarchical"],
+                        help="Topologies to test")
+    parser.add_argument("--seeds", nargs="+", type=int, default=[42, 123, 256])
+    parser.add_argument("--results-dir", default="results")
+    parser.add_argument("--log-level", default="INFO")
+    parser.add_argument("--resume", action="store_true", help="Skip already-completed cases")
+    parser.add_argument("--async", action="store_true", dest="use_async",
+                        help="Use async parallel agent execution")
+    parser.add_argument("--concurrency", type=int, default=50,
+                        help="Max concurrent cases/trials running in parallel")
+    parser.add_argument("--unanimous-mode", choices=["standard", "needs_confidence", "disabled"],
+                        default="standard", help="Unanimous convergence mode")
+    parser.add_argument("--weight-majority-ratio", type=float, default=0.6,
+                        help="Weight majority ratio threshold (0.0 to 1.0)")
+    parser.add_argument("--confidence-threshold", type=float, default=0.7,
+                        help="Average confidence threshold (0.0 to 1.0)")
+    parser.add_argument("--disable-convergence", action="store_true",
+                        help="Disable early convergence checking and force running all max rounds")
+    parser.add_argument("--randomize-round-0", action="store_true",
+                        help="Generate random answers for Round 0 instead of calling the LLM API")
+    args = parser.parse_args()
+    logging.basicConfig(
+        level=getattr(logging, args.log_level.upper()),
+        format="%(asctime)s [%(levelname)s] %(name)s: %(message)s",
+    )
+    # Suppress noisy httpx / datasets INFO logs
+    logging.getLogger("httpx").setLevel(logging.WARNING)
+    logging.getLogger("datasets").setLevel(logging.WARNING)
+    logger = logging.getLogger(__name__)
+    # Validate range limits
+    if not (0.0 <= args.weight_majority_ratio <= 1.0):
+        parser.error("--weight-majority-ratio must be between 0.0 and 1.0")
+    if not (0.0 <= args.confidence_threshold <= 1.0):
+        parser.error("--confidence-threshold must be between 0.0 and 1.0")
+    # Validate seeds vs n_trials
+    if args.n_trials > len(args.seeds):
+        logger.warning(f"n_trials={args.n_trials} > seeds={len(args.seeds)}, truncating to {len(args.seeds)}")
+        args.n_trials = len(args.seeds)
+    config = ExperimentConfig(
+        model=args.model,
+        api_base=args.api_base,
+        api_key=args.api_key,
+        temperature=args.temperature,
+        n_cases=args.n_cases if args.n_cases > 0 else None,
+        n_trials=args.n_trials,
+        max_rounds=args.max_rounds,
+        seeds=args.seeds,
+        results_dir=args.results_dir,
+        unanimous_mode=args.unanimous_mode,
+        weight_majority_ratio=args.weight_majority_ratio,
+        confidence_threshold=args.confidence_threshold,
+        disable_convergence=args.disable_convergence,
+        randomize_round_0=args.randomize_round_0,
+        resume=args.resume,
+    )
+    logger.info(f"Starting experiment: {args.n_cases} cases x {len(args.topologies)} topologies x {args.n_trials} trials")
+    logger.info(f"Model: {args.model} | Temp: {args.temperature}")
+    if args.use_async:
+        import asyncio
+        from .runner import run_single_case_async
+        from .data.loader import load_cases
+        from .types import TopologyResult
+        import json
+        cases = load_cases(config)
+        os.makedirs(config.results_dir, exist_ok=True)
+        sem = asyncio.Semaphore(args.concurrency)
+        all_results: dict[str, list] = {t: [] for t in args.topologies}
+        async def run_task(case, case_idx, topology, trial_idx, seed):
+            result_path = Path(config.results_dir) / topology / f"{case.id}_seed{seed}.json"
+            if args.resume and result_path.exists():
+                logger.info(f"  [{topology}] Case {case_idx+1}/{len(cases)} (seed={seed}): already done, skipping")
+                try:
+                    with open(result_path) as f:
+                        data = json.load(f)
+                    return topology, TopologyResult.from_dict(data)
+                except Exception as e:
+                    logger.warning(f"Failed to load checkpoint {result_path}: {e}. Re-running...")
+            async with sem:
+                logger.info(f"  [{topology}] Case {case_idx+1}/{len(cases)} (seed={seed}) starting...")
+                try:
+                    result = await run_single_case_async(case, topology, seed, config)
+                except Exception as e:
+                    logger.error(f"  [{topology}] Case {case.id} FAILED: {e}")
+                    result = TopologyResult(
+                        case_id=case.id, topology=topology, seed=seed,
+                        ground_truth=case.label, system_output="", accuracy=0.0,
+                        rounds_to_converge=config.max_rounds, convergence_type="error",
+                        total_tokens=0, avg_confidence=0.0,
+                        confidence_history={}, agent_final_outputs={},
+                        conversation_history=[], error=str(e),
+                    )
+                # Save checkpoint
+                try:
+                    result_path.parent.mkdir(parents=True, exist_ok=True)
+                    with open(result_path, "w") as f:
+                        json.dump(result.to_dict(), f, indent=2)
+                except Exception as e:
+                    logger.error(f"Failed to save checkpoint for {case.id}: {e}")
+                logger.info(f"  [{topology}] Case {case_idx+1}/{len(cases)} (seed={seed}) completed.")
+                return topology, result
+        # Build list of all coroutines
+        tasks = []
+        for case_idx, case in enumerate(cases):
+            for topology in args.topologies:
+                for trial_idx, seed in enumerate(config.seeds[:config.n_trials]):
+                    tasks.append(run_task(case, case_idx, topology, trial_idx, seed))
+        loop = asyncio.new_event_loop()
+        asyncio.set_event_loop(loop)
+        try:
+            # Run all tasks concurrently
+            results = loop.run_until_complete(asyncio.gather(*tasks))
+        finally:
+            # Explicitly close async clients inside the active loop
+            from .agent import Agent
+            async def close_async_clients():
+                for client in list(Agent._async_client_cache.values()):
+                    try:
+                        await client.aclose()
+                    except Exception:
+                        pass
+            try:
+                loop.run_until_complete(close_async_clients())
+            except Exception:
+                pass
+            loop.close()
+        # Group results by topology
+        for topology, result in results:
+            all_results[topology].append(result)
+    else:
+        all_results = run_experiment(config, topologies=args.topologies)
+    # Filter to requested topologies
+    filtered = {t: all_results.get(t, []) for t in args.topologies}
+    # Save CSV
+    csv_path = Path(args.results_dir) / "all_results.csv"
+    save_results_csv(filtered, csv_path)
+    logger.info(f"Results saved to {csv_path}")
+    # Print summary
+    summary = aggregate_results(filtered)
+    print_summary(summary)
+    # Save summary JSON
+    import json
+    summary_path = Path(args.results_dir) / "summary.json"
+    with open(summary_path, "w") as f:
+        json.dump(summary, f, indent=2)
+    logger.info(f"Summary saved to {summary_path}")
+if __name__ == "__main__":
+    main()

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/agent.py ADDED Viewed

	@@ -0,0 +1,422 @@

+# src/med_agent/agent.py
+import atexit
+import asyncio
+import json
+import logging
+import threading
+import httpx
+from pydantic import BaseModel, Field, field_validator, ValidationInfo
+from .config import ExperimentConfig
+from .types import AgentOutput, AgentMessage
+logger = logging.getLogger(__name__)
+MAX_RETRIES = 3
+SYSTEM_PROMPT = (
+    "You are a doctor. Analyze the clinical case, choose the best answer from the options, "
+    "and provide your confidence level and supporting evidence.\n\n"
+    "Respond in JSON format with exactly three fields:\n"
+    '  "answer": the letter of your chosen option (A-J)\n'
+    '  "confidence": a float indicating how certain you are. Note: in the initial round (Round 0), you must set your confidence to a value between 0.0 and 0.5 (inclusive) because you have not yet discussed the case with colleagues. In later rounds, your confidence can go up to 1.0.\n'
+    '  "evidence": a brief summary of the key clinical findings supporting your diagnosis\n\n'
+    "Example: {\"answer\": \"C\", \"confidence\": 0.45, \"evidence\": \"Patient presents with ...\"}"
+)
+RETRY_PROMPT = (
+    "Your previous response was not valid JSON. Respond with valid JSON containing exactly "
+    'three fields: "answer" (A-J), "confidence" (0.0-1.0), "evidence" (string). '
+    'Example: {"answer": "C", "confidence": 0.85, "evidence": "Key findings summary"}'
+)
+RETRY_PROMPT_ROUND_0 = (
+    "Your previous response was invalid. In the initial round (Round 0), you must set "
+    '"confidence" to a value between 0.0 and 0.5 (inclusive). '
+    'Respond with valid JSON containing exactly three fields: "answer" (A-J), "confidence" (0.0-0.5), "evidence" (string). '
+    'Example: {"answer": "C", "confidence": 0.45, "evidence": "Key findings summary"}'
+)
+class ResponseModel(BaseModel):
+    """Pydantic model for LLM output validation."""
+    answer: str = Field(...)
+    confidence: float = Field(..., ge=0.0, le=1.0)
+    evidence: str = Field(...)
+    @field_validator('answer')
+    @classmethod
+    def validate_answer(cls, v: str, info: ValidationInfo) -> str:
+        context = info.context
+        if context and "valid_options" in context:
+            valid = context["valid_options"]
+            if valid:
+                val = v.strip().upper()
+                if val in valid:
+                    return val
+                raise ValueError(f"Answer must be one of: {valid}. Got: '{v}'")
+        return v
+class ResponseModelRound0(BaseModel):
+    """Pydantic model for LLM output validation at round 0."""
+    answer: str = Field(...)
+    confidence: float = Field(..., ge=0.0, le=0.5)
+    evidence: str = Field(...)
+    @field_validator('answer')
+    @classmethod
+    def validate_answer(cls, v: str, info: ValidationInfo) -> str:
+        context = info.context
+        if context and "valid_options" in context:
+            valid = context["valid_options"]
+            if valid:
+                val = v.strip().upper()
+                if val in valid:
+                    return val
+                raise ValueError(f"Answer must be one of: {valid}. Got: '{v}'")
+        return v
+class Agent:
+    _sync_client_cache: dict[tuple[str, str], httpx.Client] = {}
+    _async_client_cache: dict[tuple[str, str], httpx.AsyncClient] = {}
+    _sync_lock = threading.Lock()
+    _async_lock = threading.Lock()
+    def __init__(self, agent_id: str, config: ExperimentConfig, seed: int):
+        self.agent_id = agent_id
+        self.config = config
+        self.seed = seed
+        self.valid_options: list[str] = []
+        self.messages: list[dict] = [
+            {"role": "system", "content": SYSTEM_PROMPT}
+        ]
+    @property
+    def client(self) -> httpx.Client:
+        cache_key = (self.config.api_base, self.config.api_key)
+        if cache_key not in self._sync_client_cache:
+            with self._sync_lock:
+                if cache_key not in self._sync_client_cache:
+                    self._sync_client_cache[cache_key] = httpx.Client(timeout=300.0)
+        return self._sync_client_cache[cache_key]
+    @property
+    def async_client(self) -> httpx.AsyncClient:
+        cache_key = (self.config.api_base, self.config.api_key)
+        if cache_key not in self._async_client_cache:
+            with self._async_lock:
+                if cache_key not in self._async_client_cache:
+                    limits = httpx.Limits(max_connections=500, max_keepalive_connections=100)
+                    self._async_client_cache[cache_key] = httpx.AsyncClient(limits=limits, timeout=300.0)
+        return self._async_client_cache[cache_key]
+    def add_case(self, question: str, options: dict[str, str]):
+        """Inject the clinical case at round 0."""
+        self.valid_options = sorted([k.strip().upper() for k in options.keys() if k.strip()])
+        options_text = "\n".join(f"{k}. {v}" for k, v in sorted(options.items()) if v.strip())
+        case_prompt = f"## Clinical Case\n\n{question}\n\n## Options\n\n{options_text}"
+        self.messages.append({"role": "user", "content": case_prompt})
+    def add_messages(self, new_messages: list[AgentMessage]):
+        """Add routed messages from other agents to conversation history."""
+        if not new_messages:
+            return
+        lines = []
+        for m in new_messages:
+            o = m.output
+            lines.append(
+                f"--- Dr. {m.agent_id} (Round {m.round}) ---\n"
+                f"Answer: {o.answer} | Confidence: {o.confidence:.2f}\n"
+                f"Evidence: {o.evidence}"
+            )
+        lines_text = "\n\n".join(lines)
+        content = (
+            "You see the following opinions from your colleagues:\n\n"
+            f"{lines_text}\n\n"
+            "Please critically evaluate these opinions based on objective clinical criteria. Pay attention to:\n"
+            "1. Clinical details or diagnostic logic presented by your colleagues.\n"
+            "2. The strengths and weaknesses of each argument.\n"
+            "CRITICAL RULES:\n"
+            "- Independently assess all arguments. Update your answer if you are convinced by stronger, more valid clinical evidence or reasoning.\n"
+            "- In your JSON 'evidence' field, briefly state your evaluation of the different viewpoints and justify your final choice."
+        )
+        self.messages.append({"role": "user", "content": content})
+    @staticmethod
+    def _extract_json(text: str) -> str | None:
+        """Find and extract a JSON object from text.
+        Handles thinking block, markdown code block formatting, or conversational text.
+        """
+        import re
+        import json
+        # Find all '{' indices in the text
+        brace_indices = [m.start() for m in re.finditer(r"\{", text)]
+        if not brace_indices:
+            return None
+        # Find the last '}'
+        end_idx = text.rfind("}")
+        if end_idx < 0:
+            return None
+        # Try candidates from left to right to find the outer valid JSON object
+        for start_idx in brace_indices:
+            if start_idx > end_idx:
+                break
+            candidate = text[start_idx : end_idx + 1]
+            try:
+                json.loads(candidate)
+                return candidate
+            except json.JSONDecodeError:
+                continue
+        return None
+    def _parse_response(self, content: str, round_idx: int = 0) -> AgentOutput | None:
+        """Parse LLM response with Pydantic validation. Returns None on failure."""
+        extracted = self._extract_json(content)
+        if extracted:
+            content = extracted
+        context = {"valid_options": self.valid_options} if self.valid_options else None
+        try:
+            if round_idx == 0:
+                parsed = ResponseModelRound0.model_validate_json(content, context=context)
+            else:
+                parsed = ResponseModel.model_validate_json(content, context=context)
+            return AgentOutput(
+                answer=parsed.answer,
+                confidence=max(0.0, min(1.0, parsed.confidence)),
+                evidence=parsed.evidence,
+            )
+        except Exception:
+            return None
+    def _llm_call(self) -> tuple[str, int]:
+        """Make a synchronous LLM call via HTTP POST."""
+        headers = {
+            "Authorization": f"Bearer {self.config.api_key}",
+            "Content-Type": "application/json",
+        }
+        payload = {
+            "model": self.config.model,
+            "messages": self.messages,
+            "temperature": self.config.temperature,
+            "max_tokens": self.config.max_tokens,
+            "seed": self.seed,
+        }
+        url = self.config.api_base
+        if not url.endswith("/"):
+            url += "/"
+        url += "chat/completions"
+        try:
+            response = self.client.post(url, headers=headers, json=payload)
+            response.raise_for_status()
+            data = response.json()
+            msg = data["choices"][0]["message"]
+            content = (msg.get("content") or "").strip()
+            tokens = data.get("usage", {}).get("total_tokens", 0)
+            # Reasoning model support
+            if not content:
+                rc = msg.get("reasoning_content") or ""
+                extracted = self._extract_json(rc)
+                if extracted:
+                    content = extracted
+            return content or "{}", tokens
+        except Exception as e:
+            logger.error(f"Agent {self.agent_id} API error (sync): {repr(e)}")
+            return "{}", 0
+    async def _llm_call_async(self) -> tuple[str, int]:
+        """Make an asynchronous LLM call via HTTP POST."""
+        headers = {
+            "Authorization": f"Bearer {self.config.api_key}",
+            "Content-Type": "application/json",
+        }
+        payload = {
+            "model": self.config.model,
+            "messages": self.messages,
+            "temperature": self.config.temperature,
+            "max_tokens": self.config.max_tokens,
+            "seed": self.seed,
+        }
+        url = self.config.api_base
+        if not url.endswith("/"):
+            url += "/"
+        url += "chat/completions"
+        try:
+            response = await self.async_client.post(url, headers=headers, json=payload)
+            response.raise_for_status()
+            data = response.json()
+            msg = data["choices"][0]["message"]
+            content = (msg.get("content") or "").strip()
+            tokens = data.get("usage", {}).get("total_tokens", 0)
+            # Reasoning model support
+            if not content:
+                rc = msg.get("reasoning_content") or ""
+                extracted = self._extract_json(rc)
+                if extracted:
+                    content = extracted
+            return content or "{}", tokens
+        except Exception as e:
+            logger.error(f"Agent {self.agent_id} API error (async): {repr(e)}")
+            return "{}", 0
+    def respond(self, round_idx: int = 0) -> AgentMessage:
+        """Call the LLM with retry on parse failure. Returns parsed response."""
+        content = ""
+        total_tokens = 0
+        original_msg_len = len(self.messages)
+        for attempt in range(MAX_RETRIES):
+            if attempt > 0:
+                # Maintain correct role alternation: user -> assistant (invalid response) -> user (retry prompt)
+                self.messages.append({"role": "assistant", "content": content or "{}"})
+                options_str = "/".join(self.valid_options) if self.valid_options else "A-J"
+                example_ans = self.valid_options[0] if self.valid_options else "C"
+                if round_idx == 0:
+                    retry_prompt = (
+                        "Your previous response was invalid. In the initial round (Round 0), you must set "
+                        '"confidence" to a value between 0.0 and 0.5 (inclusive). '
+                        f'Respond with valid JSON containing exactly three fields: "answer" ({options_str}), "confidence" (0.0-0.5), "evidence" (string). '
+                        f'Example: {{"answer": "{example_ans}", "confidence": 0.45, "evidence": "Key findings summary"}}'
+                    )
+                else:
+                    retry_prompt = (
+                        "Your previous response was not valid JSON or the answer was not one of the allowed options. "
+                        f'Respond with valid JSON containing exactly three fields: "answer" ({options_str}), "confidence" (0.0-1.0), "evidence" (string). '
+                        f'Example: {{"answer": "{example_ans}", "confidence": 0.85, "evidence": "Key findings summary"}}'
+                    )
+                self.messages.append({"role": "user", "content": retry_prompt})
+            content, tokens = self._llm_call()
+            total_tokens += tokens
+            output = self._parse_response(content, round_idx=round_idx)
+            if output is not None:
+                break
+            logger.warning(f"Agent {self.agent_id}: retry {attempt + 1}/{MAX_RETRIES} — parse failed")
+        else:
+            # Max retries exhausted — use fallback
+            logger.error(f"Agent {self.agent_id}: all {MAX_RETRIES} retries exhausted, using fallback")
+            fallback_ans = self.valid_options[0] if self.valid_options else "A"
+            output = AgentOutput(answer=fallback_ans, confidence=0.0, evidence="")
+        # Rollback temporary retry messages to avoid polluting the agent's long-term history
+        self.messages = self.messages[:original_msg_len]
+        # Add own response to history
+        self.messages.append({
+            "role": "assistant",
+            "content": json.dumps({
+                "answer": output.answer,
+                "confidence": output.confidence,
+                "evidence": output.evidence,
+            }),
+        })
+        return AgentMessage(
+            agent_id=self.agent_id,
+            round=round_idx,
+            output=output,
+            tokens=total_tokens,
+        )
+    async def respond_async(self, round_idx: int = 0) -> AgentMessage:
+        """Async version of respond() using non-blocking HTTP requests."""
+        content = ""
+        total_tokens = 0
+        original_msg_len = len(self.messages)
+        for attempt in range(MAX_RETRIES):
+            if attempt > 0:
+                # Maintain correct role alternation: user -> assistant (invalid response) -> user (retry prompt)
+                self.messages.append({"role": "assistant", "content": content or "{}"})
+                options_str = "/".join(self.valid_options) if self.valid_options else "A-J"
+                example_ans = self.valid_options[0] if self.valid_options else "C"
+                if round_idx == 0:
+                    retry_prompt = (
+                        "Your previous response was invalid. In the initial round (Round 0), you must set "
+                        '"confidence" to a value between 0.0 and 0.5 (inclusive). '
+                        f'Respond with valid JSON containing exactly three fields: "answer" ({options_str}), "confidence" (0.0-0.5), "evidence" (string). '
+                        f'Example: {{"answer": "{example_ans}", "confidence": 0.45, "evidence": "Key findings summary"}}'
+                    )
+                else:
+                    retry_prompt = (
+                        "Your previous response was not valid JSON or the answer was not one of the allowed options. "
+                        f'Respond with valid JSON containing exactly three fields: "answer" ({options_str}), "confidence" (0.0-1.0), "evidence" (string). '
+                        f'Example: {{"answer": "{example_ans}", "confidence": 0.85, "evidence": "Key findings summary"}}'
+                    )
+                self.messages.append({"role": "user", "content": retry_prompt})
+            content, tokens = await self._llm_call_async()
+            total_tokens += tokens
+            output = self._parse_response(content, round_idx=round_idx)
+            if output is not None:
+                break
+            logger.warning(f"Agent {self.agent_id}: retry {attempt + 1}/{MAX_RETRIES} — parse failed")
+        else:
+            # Max retries exhausted — use fallback
+            logger.error(f"Agent {self.agent_id}: all {MAX_RETRIES} retries exhausted, using fallback")
+            fallback_ans = self.valid_options[0] if self.valid_options else "A"
+            output = AgentOutput(answer=fallback_ans, confidence=0.0, evidence="")
+        # Rollback temporary retry messages to avoid polluting the agent's long-term history
+        self.messages = self.messages[:original_msg_len]
+        # Add own response to history
+        self.messages.append({
+            "role": "assistant",
+            "content": json.dumps({
+                "answer": output.answer,
+                "confidence": output.confidence,
+                "evidence": output.evidence,
+            }),
+        })
+        return AgentMessage(
+            agent_id=self.agent_id,
+            round=round_idx,
+            output=output,
+            tokens=total_tokens,
+        )
+@atexit.register
+def cleanup_clients():
+    """Close all cached HTTP clients on exit to release connection pools."""
+    # Close sync clients
+    for client in Agent._sync_client_cache.values():
+        try:
+            client.close()
+        except Exception:
+            pass
+    # Close async clients
+    for client in Agent._async_client_cache.values():
+        try:
+            try:
+                loop = asyncio.get_running_loop()
+                if loop.is_running():
+                    loop.create_task(client.aclose())
+                else:
+                    asyncio.run(client.aclose())
+            except RuntimeError:
+                asyncio.run(client.aclose())
+        except Exception:
+            pass

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/config.py ADDED Viewed

	@@ -0,0 +1,52 @@

+import os
+from dataclasses import dataclass, field
+@dataclass
+class ExperimentConfig:
+    # LLM
+    model: str = "deepseek-v4-flash-free"
+    api_base: str = "http://localhost:8000/v1"
+    api_key: str = field(default_factory=lambda: os.environ.get("OPENAI_API_KEY", ""))
+    temperature: float = 1.0
+    max_tokens: int = 65536  # effectively unlimited for reasoning models
+    # Experiment
+    n_agents: int = 5
+    max_rounds: int = 10
+    n_trials: int = 3
+    seeds: list[int] = field(default_factory=lambda: [42, 123, 256])
+    # Convergence
+    confidence_threshold: float = 0.7
+    weight_majority_ratio: float = 0.6
+    unanimous_mode: str = "standard"
+    disable_convergence: bool = False
+    randomize_round_0: bool = False
+    # Dataset
+    dataset_name: str = "TsinghuaC3I/MedXpertQA"
+    dataset_config: str = "Text"
+    dataset_split: str = "test"
+    medical_task_filter: str = "Diagnosis"
+    n_cases: int = 50
+    dataset_revision: str | None = "main"
+    # Paths
+    results_dir: str = "results"
+    resume: bool = False
+    def __post_init__(self):
+        if not self.api_key:
+            raise ValueError("OPENAI_API_KEY must be set")
+        if not (0.0 <= self.weight_majority_ratio <= 1.0):
+            raise ValueError("weight_majority_ratio must be between 0.0 and 1.0")
+        if not (0.0 <= self.confidence_threshold <= 1.0):
+            raise ValueError("confidence_threshold must be between 0.0 and 1.0")
+        allowed_modes = ["standard", "needs_confidence", "disabled"]
+        if self.unanimous_mode not in allowed_modes:
+            raise ValueError(f"unanimous_mode must be one of {allowed_modes}")
+        if self.n_trials > len(self.seeds):
+            raise ValueError(f"n_trials ({self.n_trials}) cannot be greater than the number of available seeds ({len(self.seeds)})")

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/data/__init__.py ADDED Viewed

	@@ -0,0 +1 @@


1	+ # src/med_agent/data/__init__.py

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/data/loader.py ADDED Viewed

	@@ -0,0 +1,51 @@

+# src/med_agent/data/loader.py
+import logging
+from dataclasses import dataclass
+from typing import Generator
+from datasets import load_dataset
+from ..config import ExperimentConfig
+logger = logging.getLogger(__name__)
+@dataclass
+class DiagnosisCase:
+    id: str
+    question: str
+    options: dict[str, str]
+    label: str
+    medical_task: str
+    body_system: str
+def load_cases(config: ExperimentConfig) -> list[DiagnosisCase]:
+    """Load MedXpertQA cases filtered by medical_task."""
+    logger.info(f"Loading {config.dataset_name} ({config.dataset_config})...")
+    ds = load_dataset(
+        config.dataset_name,
+        config.dataset_config,
+        split=config.dataset_split,
+        streaming=False,
+        revision=getattr(config, "dataset_revision", None),
+    )
+    cases: list[DiagnosisCase] = []
+    for example in ds:
+        if example["medical_task"] != config.medical_task_filter:
+            continue
+        # Clean options: remove empty ones
+        options = {k: v for k, v in example["options"].items() if v.strip()}
+        case = DiagnosisCase(
+            id=example["id"],
+            question=example["question"],
+            options=options,
+            label=example["label"],
+            medical_task=example["medical_task"],
+            body_system=example["body_system"],
+        )
+        cases.append(case)
+        if config.n_cases and len(cases) >= config.n_cases:
+            break
+    logger.info(f"Loaded {len(cases)} diagnosis cases")
+    return cases

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/metrics/__init__.py ADDED Viewed

	@@ -0,0 +1 @@


1	+ # src/med_agent/metrics/__init__.py

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/metrics/analyzer.py ADDED Viewed

	@@ -0,0 +1,89 @@

+# src/med_agent/metrics/analyzer.py
+import logging
+from collections import defaultdict
+from typing import Any
+from ..types import TopologyResult
+logger = logging.getLogger(__name__)
+def aggregate_results(results: dict[str, list[TopologyResult]]) -> dict[str, dict[str, Any]]:
+    """Aggregate per-topology results into summary statistics."""
+    summary: dict[str, dict[str, Any]] = {}
+    for topology, topology_results in results.items():
+        if not topology_results:
+            summary[topology] = {"error": "no results"}
+            continue
+        accuracies = [r.accuracy for r in topology_results]
+        rounds = [r.rounds_to_converge for r in topology_results]
+        tokens = [r.total_tokens for r in topology_results]
+        confs = [r.avg_confidence for r in topology_results]
+        conv_types = [r.convergence_type for r in topology_results]
+        errors = [r.error for r in topology_results if r.error]
+        # Convergence rate (excluding errors)
+        valid = [r for r in topology_results if r.error is None]
+        if valid:
+            n_converged = sum(1 for r in valid if r.convergence_type != "not_converged")
+            convergence_rate = n_converged / len(valid)
+        else:
+            convergence_rate = 0.0
+        # Count convergence types
+        type_counts = defaultdict(int)
+        for ct in conv_types:
+            type_counts[ct] += 1
+        summary[topology] = {
+            "n_cases": len(topology_results),
+            "n_errors": len(errors),
+            "accuracy_mean": _mean(accuracies),
+            "accuracy_std": _std(accuracies),
+            "rounds_mean": _mean(rounds),
+            "rounds_std": _std(rounds),
+            "tokens_mean": _mean(tokens),
+            "tokens_std": _std(tokens),
+            "confidence_mean": _mean(confs),
+            "confidence_std": _std(confs),
+            "convergence_rate": convergence_rate,
+            "convergence_type_counts": dict(type_counts),
+        }
+    return summary
+def print_summary(summary: dict[str, dict[str, Any]]):
+    """Pretty-print the summary table."""
+    header = f"{'Topology':<15} {'Accuracy':<12} {'Rounds':<12} {'Tokens':<16} {'Confidence':<12} {'Conv Rate':<12}"
+    sep = "-" * len(header)
+    lines = [sep, header, sep]
+    for topology, s in summary.items():
+        if "error" in s:
+            lines.append(f"{topology:<15} ERROR: {s['error']}")
+            continue
+        lines.append(
+            f"{topology:<15}"
+            f"{s['accuracy_mean']:.1%}±{s['accuracy_std']:.1%}  "
+            f"{s['rounds_mean']:.1f}±{s['rounds_std']:.1f}  "
+            f"{s['tokens_mean']:.0f}±{s['tokens_std']:.0f}  "
+            f"{s['confidence_mean']:.2f}±{s['confidence_std']:.2f}  "
+            f"{s['convergence_rate']:.1%}"
+        )
+    lines.append(sep)
+    logger.info("\n" + "\n".join(lines))
+    print("\n".join(lines))
+def _mean(values: list[float]) -> float:
+    return sum(values) / len(values) if values else 0.0
+def _std(values: list[float]) -> float:
+    if len(values) < 2:
+        return 0.0
+    m = _mean(values)
+    return (sum((v - m) ** 2 for v in values) / (len(values) - 1)) ** 0.5

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/metrics/collector.py ADDED Viewed

	@@ -0,0 +1,38 @@

+# src/med_agent/metrics/collector.py
+import csv
+import json
+import os
+from pathlib import Path
+from typing import Any
+from ..types import TopologyResult
+def save_results_csv(results: dict[str, list[TopologyResult]], path: str | Path):
+    """Save per-trial results to CSV for easy pandas analysis."""
+    path = Path(path)
+    path.parent.mkdir(parents=True, exist_ok=True)
+    rows = []
+    for topology, topology_results in results.items():
+        for r in topology_results:
+            rows.append({
+                "case_id": r.case_id,
+                "topology": r.topology,
+                "seed": r.seed,
+                "ground_truth": r.ground_truth,
+                "system_output": r.system_output,
+                "accuracy": r.accuracy,
+                "rounds_to_converge": r.rounds_to_converge,
+                "convergence_type": r.convergence_type,
+                "total_tokens": r.total_tokens,
+                "avg_confidence": r.avg_confidence,
+                "error": r.error or "",
+            })
+    if not rows:
+        return
+    with open(path, "w", newline="") as f:
+        writer = csv.DictWriter(f, fieldnames=rows[0].keys())
+        writer.writeheader()
+        writer.writerows(rows)

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/runner.py ADDED Viewed

	@@ -0,0 +1,404 @@

+# src/med_agent/runner.py
+import hashlib
+import json
+import logging
+import os
+import random
+from pathlib import Path
+from typing import Optional
+from .config import ExperimentConfig
+from .types import AgentOutput, AgentMessage, TopologyResult
+from .agent import Agent
+from .data.loader import DiagnosisCase, load_cases
+from .topologies import ROUTERS
+logger = logging.getLogger(__name__)
+AGENT_IDS = ["A0", "A1", "A2", "A3", "A4"]
+def _check_convergence(
+    agents_outputs: dict[str, AgentMessage],
+    config: ExperimentConfig,
+) -> tuple[bool, Optional[str], str]:
+    """Check three-tier convergence. Returns (converged, system_answer, convergence_type)."""
+    if not agents_outputs:
+        return False, None, "not_converged"
+    if getattr(config, "disable_convergence", False):
+        return False, None, "not_converged"
+    answers = {aid: msg.output for aid, msg in agents_outputs.items()}
+    unanimous_mode = getattr(config, "unanimous_mode", "standard")
+    # Tier 1: Unanimous (if not disabled)
+    is_unanimous = False
+    first_answer = None
+    if answers:
+        first_answer = list(answers.values())[0].answer
+        is_unanimous = all(o.answer == first_answer for o in answers.values())
+    if is_unanimous:
+        if unanimous_mode == "standard":
+            return True, first_answer, "unanimous"
+        elif unanimous_mode == "needs_confidence":
+            total_conf = sum(o.confidence for o in answers.values())
+            avg_conf = total_conf / len(answers) if answers else 0.0
+            if avg_conf >= config.confidence_threshold:
+                return True, first_answer, "unanimous_with_conf"
+            # Otherwise, proceed to Tier 2
+        elif unanimous_mode == "disabled":
+            pass  # proceed to Tier 2
+        else:
+            logger.warning(
+                f"Unknown unanimous_mode: '{unanimous_mode}'. "
+                f"Falling back to standard unanimous convergence."
+            )
+            return True, first_answer, "unanimous"
+    # Tier 2: Weighted majority
+    total_weight = sum(o.confidence for o in answers.values())
+    if total_weight == 0:
+        return False, None, "not_converged"
+    vote_sums: dict[str, float] = {}
+    vote_confidences: dict[str, list[float]] = {}
+    for aid, o in answers.items():
+        vote_sums[o.answer] = vote_sums.get(o.answer, 0.0) + o.confidence
+        vote_confidences.setdefault(o.answer, []).append(o.confidence)
+    top_answer = max(vote_sums, key=vote_sums.get)
+    top_weight = vote_sums[top_answer]
+    top_conf_list = vote_confidences[top_answer]
+    top_avg_conf = sum(top_conf_list) / len(top_conf_list) if top_conf_list else 0.0
+    if (top_weight / total_weight) >= config.weight_majority_ratio and top_avg_conf >= config.confidence_threshold:
+        return True, top_answer, "weighted_majority"
+    return False, None, "not_converged"
+def _tie_break(vote_sums: dict[str, float], agents_outputs: dict[str, AgentMessage]) -> str:
+    """Break ties: highest single-agent confidence, then alphabetically."""
+    max_vote = max(vote_sums.values())
+    tied = [ans for ans, v in vote_sums.items() if v == max_vote]
+    if len(tied) == 1:
+        return tied[0]
+    # Among tied, find highest per-answer max confidence
+    max_conf_per_answer: dict[str, float] = {}
+    for ans in tied:
+        max_conf_per_answer[ans] = max(
+            (msg.output.confidence for msg in agents_outputs.values() if msg.output.answer == ans),
+            default=0.0,
+        )
+    best_max_conf = max(max_conf_per_answer.values())
+    best_tied = [ans for ans, c in max_conf_per_answer.items() if c == best_max_conf]
+    return min(best_tied)  # alphabetical among remaining
+def _generate_random_round_0_msg(
+    agent: Agent,
+    case: DiagnosisCase,
+    seed: int,
+    config: ExperimentConfig,
+) -> AgentMessage:
+    """Generate a random Round 0 message deterministically and update agent messages."""
+    seed_str = f"{case.id}_{agent.agent_id}_{seed}"
+    seed_hash = int(hashlib.md5(seed_str.encode()).hexdigest(), 16)
+    rng = random.Random(seed_hash)
+    valid_options = sorted(list(case.options.keys()))
+    if not valid_options:
+        valid_options = ["A"]
+    ans = rng.choice(valid_options)
+    conf = round(rng.uniform(0.1, 0.5), 2)
+    evidence = "I have not read the patient case details yet. This initial choice and confidence score are randomly generated to establish a baseline for consensus testing."
+    output = AgentOutput(answer=ans, confidence=conf, evidence=evidence)
+    # Maintain conversational context (assistant role)
+    agent.messages.append({
+        "role": "assistant",
+        "content": json.dumps({
+            "answer": ans,
+            "confidence": conf,
+            "evidence": evidence,
+        })
+    })
+    return AgentMessage(
+        agent_id=agent.agent_id,
+        round=0,
+        output=output,
+        tokens=0,
+    )
+def run_single_case(
+    case: DiagnosisCase,
+    topology: str,
+    seed: int,
+    config: ExperimentConfig,
+) -> TopologyResult:
+    """Run one case under one topology with one seed."""
+    route_fn = ROUTERS[topology]
+    agent_ids = ["A0"] if topology == "single" else AGENT_IDS
+    # Initialize agents
+    agents: list[Agent] = []
+    for i, aid in enumerate(agent_ids):
+        agent = Agent(aid, config, seed)
+        agent.add_case(case.question, case.options)
+        agents.append(agent)
+    all_messages: dict[str, list[AgentMessage]] = {aid: [] for aid in agent_ids}
+    confidence_history: dict[int, dict[str, float]] = {}
+    conversation_history: list[list[AgentMessage]] = []
+    # Round 0: initial independent responses
+    round_msgs: list[AgentMessage] = []
+    if getattr(config, "randomize_round_0", False):
+        for agent in agents:
+            msg = _generate_random_round_0_msg(agent, case, seed, config)
+            round_msgs.append(msg)
+            all_messages[agent.agent_id].append(msg)
+    else:
+        for agent in agents:
+            msg = agent.respond(0)
+            round_msgs.append(msg)
+            all_messages[agent.agent_id].append(msg)
+    conversation_history.append(list(round_msgs))
+    confidence_history[0] = {m.agent_id: m.output.confidence for m in round_msgs}
+    agents_outputs = {m.agent_id: m for m in round_msgs}
+    converged = False
+    answer = None
+    conv_type = "single" if topology == "single" else "not_converged"
+    # Rounds 1..max_rounds: route -> respond -> check convergence
+    for round_idx in range(1, config.max_rounds + 1):
+        if topology == "single":
+            break
+        # Route messages from previous round
+        visible = route_fn(all_messages, round_idx)
+        # Each agent receives routed messages (from ALL previous rounds, not just latest)
+        # Actually, add only new messages -- the routing function controls what's visible
+        for agent in agents:
+            new_msgs = visible.get(agent.agent_id, [])
+            # Filter to only messages the agent hasn't seen
+            seen_ids = {(m.agent_id, m.round) for m in all_messages[agent.agent_id]}
+            unseen = [m for m in new_msgs if (m.agent_id, m.round) not in seen_ids]
+            if unseen:
+                agent.add_messages(unseen)
+        # Agents respond in parallel (sequential for simplicity; parallel via asyncio later)
+        round_msgs = []
+        for agent in agents:
+            msg = agent.respond(round_idx)
+            round_msgs.append(msg)
+            all_messages[agent.agent_id].append(msg)
+        conversation_history.append(list(round_msgs))
+        confs = {m.agent_id: m.output.confidence for m in round_msgs}
+        confidence_history[round_idx] = confs
+        # Check convergence
+        agents_outputs = {m.agent_id: m for m in round_msgs}
+        converged, answer, conv_type = _check_convergence(agents_outputs, config)
+        if converged:
+            break
+    # If loop exhausted without convergence, fallback to weighted majority
+    if not converged and topology != "single":
+        # Weighted majority tie-break for final output
+        total_weight = sum(m.output.confidence for m in agents_outputs.values())
+        if total_weight > 0:
+            vote_sums: dict[str, float] = {}
+            for m in agents_outputs.values():
+                vote_sums[m.output.answer] = vote_sums.get(m.output.answer, 0.0) + m.output.confidence
+            answer = _tie_break(vote_sums, agents_outputs)
+        else:
+            answer = list(agents_outputs.values())[0].output.answer
+        conv_type = "not_converged"
+    elif topology == "single":
+        answer = list(agents_outputs.values())[0].output.answer
+    total_tokens = sum(m.tokens for msgs_list in all_messages.values() for m in msgs_list)
+    final_confs = [m.output.confidence for m in agents_outputs.values()]
+    avg_confidence = sum(final_confs) / len(final_confs) if final_confs else 0.0
+    n_rounds = len(conversation_history) - 1  # exclude round 0 (initial independent reasoning)
+    result = TopologyResult(
+        case_id=case.id,
+        topology=topology,
+        seed=seed,
+        ground_truth=case.label,
+        system_output=answer,
+        accuracy=1.0 if answer == case.label else 0.0,
+        rounds_to_converge=n_rounds,
+        convergence_type=conv_type,
+        total_tokens=total_tokens,
+        avg_confidence=avg_confidence,
+        confidence_history=confidence_history,
+        agent_final_outputs={aid: agents_outputs[aid].output for aid in agent_ids if aid in agents_outputs},
+        conversation_history=conversation_history,
+    )
+    return result
+async def run_single_case_async(
+    case: DiagnosisCase,
+    topology: str,
+    seed: int,
+    config: ExperimentConfig,
+) -> TopologyResult:
+    """Async version with parallel agent execution per round."""
+    import asyncio
+    route_fn = ROUTERS[topology]
+    agent_ids = ["A0"] if topology == "single" else AGENT_IDS
+    agents: list[Agent] = []
+    for aid in agent_ids:
+        agent = Agent(aid, config, seed)
+        agent.add_case(case.question, case.options)
+        agents.append(agent)
+    all_messages: dict[str, list[AgentMessage]] = {aid: [] for aid in agent_ids}
+    confidence_history: dict[int, dict[str, float]] = {}
+    conversation_history: list[list[AgentMessage]] = []
+    # Round 0: parallel initial responses
+    if getattr(config, "randomize_round_0", False):
+        round_msgs = [
+            _generate_random_round_0_msg(agent, case, seed, config)
+            for agent in agents
+        ]
+    else:
+        round_msgs = await asyncio.gather(*[agent.respond_async(0) for agent in agents])
+    for msg in round_msgs:
+        all_messages[msg.agent_id].append(msg)
+    conversation_history.append(list(round_msgs))
+    confidence_history[0] = {m.agent_id: m.output.confidence for m in round_msgs}
+    agents_outputs = {m.agent_id: m for m in round_msgs}
+    converged = False
+    answer = None
+    conv_type = "single" if topology == "single" else "not_converged"
+    # Rounds 1..max_rounds
+    for round_idx in range(1, config.max_rounds + 1):
+        if topology == "single":
+            break
+        visible = route_fn(all_messages, round_idx)
+        for agent in agents:
+            new_msgs = visible.get(agent.agent_id, [])
+            seen_ids = {(m.agent_id, m.round) for m in all_messages[agent.agent_id]}
+            unseen = [m for m in new_msgs if (m.agent_id, m.round) not in seen_ids]
+            if unseen:
+                agent.add_messages(unseen)
+        # Parallel agent responses
+        round_msgs = await asyncio.gather(*[agent.respond_async(round_idx) for agent in agents])
+        for msg in round_msgs:
+            all_messages[msg.agent_id].append(msg)
+        conversation_history.append(list(round_msgs))
+        confs = {m.agent_id: m.output.confidence for m in round_msgs}
+        confidence_history[round_idx] = confs
+        agents_outputs = {m.agent_id: m for m in round_msgs}
+        converged, answer, conv_type = _check_convergence(agents_outputs, config)
+        if converged:
+            break
+    if not converged and topology != "single":
+        total_weight = sum(m.output.confidence for m in agents_outputs.values())
+        if total_weight > 0:
+            vote_sums: dict[str, float] = {}
+            for m in agents_outputs.values():
+                vote_sums[m.output.answer] = vote_sums.get(m.output.answer, 0.0) + m.output.confidence
+            answer = _tie_break(vote_sums, agents_outputs)
+        else:
+            answer = list(agents_outputs.values())[0].output.answer
+        conv_type = "not_converged"
+    elif topology == "single":
+        answer = list(agents_outputs.values())[0].output.answer
+    total_tokens = sum(m.tokens for msgs_list in all_messages.values() for m in msgs_list)
+    final_confs = [m.output.confidence for m in agents_outputs.values()]
+    avg_confidence = sum(final_confs) / len(final_confs) if final_confs else 0.0
+    n_rounds = len(conversation_history) - 1
+    return TopologyResult(
+        case_id=case.id,
+        topology=topology,
+        seed=seed,
+        ground_truth=case.label,
+        system_output=answer,
+        accuracy=1.0 if answer == case.label else 0.0,
+        rounds_to_converge=n_rounds,
+        convergence_type=conv_type,
+        total_tokens=total_tokens,
+        avg_confidence=avg_confidence,
+        confidence_history=confidence_history,
+        agent_final_outputs={aid: agents_outputs[aid].output for aid in agent_ids if aid in agents_outputs},
+        conversation_history=conversation_history,
+    )
+def run_experiment(config: ExperimentConfig, topologies: list[str] = None):
+    """Run full experiment: all cases x target topologies x all trials."""
+    cases = load_cases(config)
+    os.makedirs(config.results_dir, exist_ok=True)
+    target_topologies = topologies or list(ROUTERS.keys())
+    all_results: dict[str, list[TopologyResult]] = {t: [] for t in target_topologies}
+    for case_idx, case in enumerate(cases):
+        logger.info(f"Case {case_idx+1}/{len(cases)}: {case.id}")
+        for topology in target_topologies:
+            for trial_idx, seed in enumerate(config.seeds[:config.n_trials]):
+                # Checkpoint: skip if already done
+                result_path = Path(config.results_dir) / topology / f"{case.id}_seed{seed}.json"
+                if getattr(config, "resume", False) and result_path.exists():
+                    logger.info(f"  [{topology}] trial {trial_idx+1}/{config.n_trials}: already done, skipping")
+                    try:
+                        with open(result_path) as f:
+                            data = json.load(f)
+                        all_results[topology].append(TopologyResult.from_dict(data))
+                        continue
+                    except Exception as e:
+                        logger.warning(f"Failed to load checkpoint {result_path}: {e}. Re-running...")
+                logger.info(f"  [{topology}] trial {trial_idx+1}/{config.n_trials} (seed={seed})...")
+                try:
+                    result = run_single_case(case, topology, seed, config)
+                except Exception as e:
+                    logger.error(f"  [{topology}] FAILED: {e}")
+                    result = TopologyResult(
+                        case_id=case.id, topology=topology, seed=seed,
+                        ground_truth=case.label, system_output="", accuracy=0.0,
+                        rounds_to_converge=config.max_rounds, convergence_type="error",
+                        total_tokens=0, avg_confidence=0.0,
+                        confidence_history={}, agent_final_outputs={},
+                        conversation_history=[], error=str(e),
+                    )
+                # Save checkpoint using to_dict for proper nested serialization
+                result_path.parent.mkdir(parents=True, exist_ok=True)
+                with open(result_path, "w") as f:
+                    json.dump(result.to_dict(), f, indent=2)
+                all_results[topology].append(result)
+    return all_results

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/__init__.py ADDED Viewed

	@@ -0,0 +1,17 @@

+# src/med_agent/topologies/__init__.py
+from .star import route_star
+from .complete import route_complete
+from .ring import route_ring
+from .hierarchical import route_hierarchical
+from .single import route_single
+ROUTERS = {
+    "star": route_star,
+    "complete": route_complete,
+    "ring": route_ring,
+    "hierarchical": route_hierarchical,
+    "single": route_single,
+}
+__all__ = ["ROUTERS", "route_star", "route_complete", "route_ring", "route_hierarchical", "route_single"]

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/base.py ADDED Viewed

	@@ -0,0 +1,9 @@

+from typing import Dict, List
+from ..types import AgentMessage
+# Input: all messages from the latest round, grouped by agent
+# Output: for each agent, which messages they can see
+RouteInput = Dict[str, List[AgentMessage]]
+RouteOutput = Dict[str, List[AgentMessage]]
+AGENT_IDS = ["A0", "A1", "A2", "A3", "A4"]

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/complete.py ADDED Viewed

	@@ -0,0 +1,12 @@

+from .base import RouteInput, RouteOutput, AGENT_IDS
+def route_complete(all_messages: RouteInput, current_round: int) -> RouteOutput:
+    """Complete: every agent sees every message (group chat)."""
+    result: RouteOutput = {}
+    all_msgs = []
+    for msgs in all_messages.values():
+        all_msgs.extend(msgs)
+    for agent_id in AGENT_IDS:
+        result[agent_id] = list(all_msgs)  # everyone sees everything
+    return result

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/hierarchical.py ADDED Viewed

	@@ -0,0 +1,36 @@

+from .base import RouteInput, RouteOutput, AGENT_IDS
+# Tree structure: A0 (root) -> A1, A2 (mid) -> A3, A4 (leaf)
+# A3 reports to A1, A4 reports to A2
+# A1 and A2 report to A0
+# A0 broadcasts to A1, A2 only
+HIERARCHY = {
+    "leaf_to_mid": {"A3": "A1", "A4": "A2"},
+    "mid_to_root": {"A1": "A0", "A2": "A0"},
+}
+def route_hierarchical(all_messages: RouteInput, current_round: int) -> RouteOutput:
+    """Hierarchical: messages flow up the tree, only A0 broadcasts down to mid-level."""
+    result: RouteOutput = {}
+    for agent_id in AGENT_IDS:
+        result[agent_id] = []
+    for sender_id, msgs in all_messages.items():
+        if sender_id in HIERARCHY["leaf_to_mid"]:
+            # Leaf -> only visible to its mid-level manager
+            mid = HIERARCHY["leaf_to_mid"][sender_id]
+            result[mid].extend(msgs)
+            # Leaf sees its own message
+            result[sender_id].extend(msgs)
+        elif sender_id in HIERARCHY["mid_to_root"]:
+            # Mid-level -> visible to root (A0) and itself
+            root = HIERARCHY["mid_to_root"][sender_id]
+            result[root].extend(msgs)
+            result[sender_id].extend(msgs)
+        elif sender_id == "A0":
+            # Root -> broadcast to mid-level only
+            result["A0"].extend(msgs)
+            for mid_id in HIERARCHY["mid_to_root"]:
+                result[mid_id].extend(msgs)
+    return result

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/ring.py ADDED Viewed

	@@ -0,0 +1,24 @@

+from .base import RouteInput, RouteOutput, AGENT_IDS
+# Ring adjacency: A0<->A1<->A2<->A3<->A4<->A0
+NEIGHBORS = {
+    "A0": ["A4", "A1"],
+    "A1": ["A0", "A2"],
+    "A2": ["A1", "A3"],
+    "A3": ["A2", "A4"],
+    "A4": ["A3", "A0"],
+}
+def route_ring(all_messages: RouteInput, current_round: int) -> RouteOutput:
+    """Ring: each agent sees messages from itself and immediate neighbors."""
+    result: RouteOutput = {}
+    for agent_id in AGENT_IDS:
+        result[agent_id] = []
+    for agent_id in AGENT_IDS:
+        visible_ids = set(NEIGHBORS[agent_id] + [agent_id])
+        for sender_id, msgs in all_messages.items():
+            if sender_id in visible_ids:
+                result[agent_id].extend(msgs)
+    return result

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/single.py ADDED Viewed

	@@ -0,0 +1,8 @@

+from .base import RouteInput, RouteOutput
+def route_single(all_messages: RouteInput, current_round: int) -> RouteOutput:
+    """Single Agent: see own messages only."""
+    result: RouteOutput = {}
+    if "A0" in all_messages:
+        result["A0"] = list(all_messages["A0"])
+    return result

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/topologies/star.py ADDED Viewed

	@@ -0,0 +1,23 @@

+from .base import RouteInput, RouteOutput, AGENT_IDS
+def route_star(all_messages: RouteInput, current_round: int) -> RouteOutput:
+    """Star: center (A0) sees all; peripherals only see A0's messages and their own."""
+    result: RouteOutput = {}
+    for agent_id in AGENT_IDS:
+        result[agent_id] = []
+        if agent_id in all_messages:
+            result[agent_id].extend(all_messages[agent_id])
+    # Collect all messages from this round
+    # A0 (center) sees all messages from peripherals
+    for sender_id, msgs in all_messages.items():
+        if sender_id != "A0":
+            result["A0"].extend(msgs)
+    # Peripherals see A0's message (if any)
+    if "A0" in all_messages:
+        for agent_id in AGENT_IDS:
+            if agent_id != "A0":
+                result[agent_id].extend(all_messages["A0"])
+    return result

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/src/med_agent/types.py ADDED Viewed

	@@ -0,0 +1,92 @@

+from dataclasses import dataclass
+from typing import Optional
+@dataclass
+class AgentOutput:
+    answer: str
+    confidence: float
+    evidence: str
+    def to_dict(self) -> dict:
+        return {"answer": self.answer, "confidence": self.confidence, "evidence": self.evidence}
+    @classmethod
+    def from_dict(cls, d: dict) -> "AgentOutput":
+        return cls(answer=d["answer"], confidence=d["confidence"], evidence=d["evidence"])
+@dataclass
+class AgentMessage:
+    agent_id: str
+    round: int
+    output: AgentOutput
+    tokens: int
+    def to_dict(self) -> dict:
+        return {
+            "agent_id": self.agent_id, "round": self.round,
+            "output": self.output.to_dict(), "tokens": self.tokens,
+        }
+    @classmethod
+    def from_dict(cls, d: dict) -> "AgentMessage":
+        return cls(
+            agent_id=d["agent_id"], round=d["round"],
+            output=AgentOutput.from_dict(d["output"]), tokens=d["tokens"],
+        )
+@dataclass
+class TopologyResult:
+    case_id: str
+    topology: str
+    seed: int
+    ground_truth: str
+    system_output: str
+    accuracy: float
+    rounds_to_converge: int
+    convergence_type: str
+    total_tokens: int
+    avg_confidence: float
+    confidence_history: dict[int, dict[str, float]]
+    agent_final_outputs: dict[str, AgentOutput]
+    conversation_history: list[list[AgentMessage]]
+    error: Optional[str] = None
+    def to_dict(self) -> dict:
+        return {
+            "case_id": self.case_id,
+            "topology": self.topology,
+            "seed": self.seed,
+            "ground_truth": self.ground_truth,
+            "system_output": self.system_output,
+            "accuracy": self.accuracy,
+            "rounds_to_converge": self.rounds_to_converge,
+            "convergence_type": self.convergence_type,
+            "total_tokens": self.total_tokens,
+            "avg_confidence": self.avg_confidence,
+            "confidence_history": {str(k): v for k, v in self.confidence_history.items()},
+            "agent_final_outputs": {k: v.to_dict() for k, v in self.agent_final_outputs.items()},
+            "conversation_history": [[m.to_dict() for m in r] for r in self.conversation_history],
+            "error": self.error,
+        }
+    @classmethod
+    def from_dict(cls, d: dict) -> "TopologyResult":
+        return cls(
+            case_id=d["case_id"],
+            topology=d["topology"],
+            seed=d["seed"],
+            ground_truth=d["ground_truth"],
+            system_output=d["system_output"],
+            accuracy=d["accuracy"],
+            rounds_to_converge=d["rounds_to_converge"],
+            convergence_type=d["convergence_type"],
+            total_tokens=d["total_tokens"],
+            avg_confidence=d["avg_confidence"],
+            confidence_history={int(k): v for k, v in d["confidence_history"].items()},
+            agent_final_outputs={k: AgentOutput.from_dict(v) for k, v in d["agent_final_outputs"].items()},
+            conversation_history=[[AgentMessage.from_dict(m) for m in r] for r in d["conversation_history"]],
+            error=d.get("error"),
+        )

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/__init__.py ADDED Viewed

File without changes

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/conftest.py ADDED Viewed

	@@ -0,0 +1,46 @@

+import pytest
+from src.med_agent.config import ExperimentConfig
+@pytest.fixture
+def config():
+    return ExperimentConfig(n_cases=3, n_trials=1, seeds=[42])
+@pytest.fixture
+def sample_options():
+    return {"A": "Common cold", "B": "Influenza", "C": "COVID-19"}
+@pytest.fixture
+def sample_case_text():
+    return "Patient has fever for 3 days, cough, fatigue. Temp 38.5C. No rash."
+@pytest.fixture(autouse=True)
+def clean_client_cache():
+    import asyncio
+    from src.med_agent.agent import Agent
+    yield
+    # Teardown: close all cached sync clients
+    for client in list(Agent._sync_client_cache.values()):
+        try:
+            client.close()
+        except Exception:
+            pass
+    Agent._sync_client_cache.clear()
+    # Teardown: close all cached async clients
+    for client in list(Agent._async_client_cache.values()):
+        try:
+            try:
+                loop = asyncio.get_running_loop()
+                if loop.is_running():
+                    loop.create_task(client.aclose())
+                else:
+                    loop.run_until_complete(client.aclose())
+            except RuntimeError:
+                asyncio.run(client.aclose())
+        except Exception:
+            pass
+    Agent._async_client_cache.clear()

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_agent.py ADDED Viewed

	@@ -0,0 +1,218 @@

+# tests/test_agent.py
+import asyncio
+import threading
+import time
+from unittest.mock import patch
+import pytest
+import httpx
+from src.med_agent.config import ExperimentConfig
+from src.med_agent.agent import Agent
+from src.med_agent.types import AgentOutput, AgentMessage
+class TestAgent:
+    def test_initialization(self):
+        config = ExperimentConfig()
+        agent = Agent("A0", config, seed=42)
+        assert agent.agent_id == "A0"
+        assert len(agent.messages) == 1  # system prompt
+    def test_add_case_adds_to_messages(self):
+        config = ExperimentConfig()
+        agent = Agent("A0", config, seed=42)
+        agent.add_case("Patient has fever?", {"A": "Cold", "B": "Flu"})
+        assert len(agent.messages) == 2
+    def test_add_messages_formats_correctly(self):
+        config = ExperimentConfig()
+        agent = Agent("A0", config, seed=42)
+        msg = AgentMessage(
+            agent_id="A1", round=1,
+            output=AgentOutput(answer="B", confidence=0.9, evidence="Fever"),
+            tokens=50,
+        )
+        agent.add_messages([msg])
+        assert len(agent.messages) == 2
+        assert "Dr. A1" in agent.messages[1]["content"]
+    @pytest.mark.integration
+    def test_respond_calls_llm(self):
+        """Requires running LiteLLM proxy. Marked as integration."""
+        config = ExperimentConfig()
+        agent = Agent("A0", config, seed=42)
+        agent.add_case("Patient has fever for 3 days, cough, fatigue. Temp 38.5C. No rash.",
+                       {"A": "Common cold", "B": "Influenza", "C": "COVID-19"})
+        result = agent.respond()
+        assert isinstance(result, AgentMessage)
+        assert result.agent_id == "A0"
+        assert result.output.answer in ("A", "B", "C")
+        assert 0.0 <= result.output.confidence <= 1.0
+        assert isinstance(result.output.evidence, str)
+        assert isinstance(result.tokens, int)
+    def test_parse_response_round_0_confidence_constraints(self):
+        config = ExperimentConfig()
+        agent = Agent("A0", config, seed=42)
+        # 1. Round 0 with valid confidence <= 0.5 should succeed
+        valid_round_0_json = '{"answer": "C", "confidence": 0.45, "evidence": "Valid evidence"}'
+        res = agent._parse_response(valid_round_0_json, round_idx=0)
+        assert res is not None
+        assert res.answer == "C"
+        assert res.confidence == 0.45
+        # 2. Round 0 with confidence > 0.5 should fail (return None)
+        invalid_round_0_json = '{"answer": "C", "confidence": 0.85, "evidence": "Invalid evidence"}'
+        res_invalid = agent._parse_response(invalid_round_0_json, round_idx=0)
+        assert res_invalid is None
+        # 3. Round > 0 with confidence > 0.5 should succeed
+        valid_round_1_json = '{"answer": "C", "confidence": 0.85, "evidence": "Valid evidence in round 1"}'
+        res_round_1 = agent._parse_response(valid_round_1_json, round_idx=1)
+        assert res_round_1 is not None
+        assert res_round_1.answer == "C"
+        assert res_round_1.confidence == 0.85
+    def test_start_main_experiment(self):
+        import subprocess
+        import os
+        cmd = [
+            ".venv/bin/python3", "main.py",
+            "--n-cases", "0",
+            "--n-trials", "3",
+            "--topologies", "single", "star", "complete", "ring", "hierarchical",
+            "--async",
+            "--concurrency", "50",
+            "--results-dir", "results_concurrency50",
+            "--resume"
+        ]
+        log_path = os.path.abspath("baseline_run.log")
+        with open(log_path, "w") as log_file:
+            subprocess.Popen(
+                cmd,
+                stdout=log_file,
+                stderr=subprocess.STDOUT,
+                close_fds=True,
+                preexec_fn=os.setpgrp
+            )
+        print(f"Successfully started main experiment in background. Logs redirected to {log_path}")
+def test_client_thread_safety():
+    # Clear cache to force initialization
+    Agent._sync_client_cache.clear()
+    config = ExperimentConfig()
+    agent = Agent("A0", config, seed=42)
+    results = []
+    original_init = httpx.Client.__init__
+    init_count = 0
+    init_lock = threading.Lock()
+    def slow_init(self, *args, **kwargs):
+        nonlocal init_count
+        with init_lock:
+            init_count += 1
+        time.sleep(0.1)  # Sleep to ensure race condition triggers
+        original_init(self, *args, **kwargs)
+    def worker():
+        try:
+            client = agent.client
+            results.append(client)
+        except Exception as e:
+            results.append(e)
+    with patch.object(httpx.Client, "__init__", slow_init):
+        threads = [threading.Thread(target=worker) for _ in range(5)]
+        for t in threads:
+            t.start()
+        for t in threads:
+            t.join()
+    for r in results:
+        assert not isinstance(r, Exception), f"Worker failed: {r}"
+        assert r is results[0]
+    assert init_count == 1
+def test_async_client_thread_safety():
+    # Clear cache to force initialization
+    Agent._async_client_cache.clear()
+    config = ExperimentConfig()
+    agent = Agent("A0", config, seed=42)
+    results = []
+    original_init = httpx.AsyncClient.__init__
+    init_count = 0
+    init_lock = threading.Lock()
+    def slow_init(self, *args, **kwargs):
+        nonlocal init_count
+        with init_lock:
+            init_count += 1
+        time.sleep(0.1)  # Sleep to ensure race condition triggers
+        original_init(self, *args, **kwargs)
+    def worker():
+        try:
+            client = agent.async_client
+            results.append(client)
+        except Exception as e:
+            results.append(e)
+    with patch.object(httpx.AsyncClient, "__init__", slow_init):
+        threads = [threading.Thread(target=worker) for _ in range(5)]
+        for t in threads:
+            t.start()
+        for t in threads:
+            t.join()
+    for r in results:
+        assert not isinstance(r, Exception), f"Worker failed: {r}"
+        assert r is results[0]
+    assert init_count == 1
+def test_strict_pydantic_option_validation():
+    config = ExperimentConfig()
+    agent = Agent("A0", config, seed=42)
+    # 1. Before case is added, validation should fall back to basic parsing or allow all
+    # Let's test after add_case:
+    agent.add_case("Patient has fever?", {"A": "Cold", "B": "Flu"})
+    # Parsing a valid option (A or B) should succeed
+    valid_json = '{"answer": "A", "confidence": 0.4, "evidence": "good evidence"}'
+    res = agent._parse_response(valid_json, round_idx=0)
+    assert res is not None
+    assert res.answer == "A"
+    # Parsing a valid option with lowercase should succeed and format to uppercase
+    valid_json_lower = '{"answer": "b", "confidence": 0.4, "evidence": "good evidence"}'
+    res = agent._parse_response(valid_json_lower, round_idx=0)
+    assert res is not None
+    assert res.answer == "B"
+    # Parsing an option not in the case options (e.g. C) should fail (return None)
+    invalid_json_c = '{"answer": "C", "confidence": 0.4, "evidence": "good evidence"}'
+    res = agent._parse_response(invalid_json_c, round_idx=0)
+    assert res is None
+    # Parsing a non-compliant option string like "Option B" or "None" should fail (return None)
+    non_compliant_json = '{"answer": "Option B", "confidence": 0.4, "evidence": "good evidence"}'
+    res = agent._parse_response(non_compliant_json, round_idx=0)
+    assert res is None

data/external/results/formal/early_mas_architecture_motivation/code_snapshot/med_agent_reconstructed/tests/test_analyze_ab_details.py ADDED Viewed

	@@ -0,0 +1,184 @@

+# tests/test_analyze_ab_details.py
+import os
+import json
+import glob
+import numpy as np
+import pandas as pd
+def load_all_json_details(results_dir):
+    data = []
+    pattern = os.path.join(results_dir, "*", "*.json")
+    for filepath in glob.glob(pattern):
+        topology = os.path.basename(os.path.dirname(filepath))
+        case_id = os.path.basename(filepath).replace(".json", "")
+        with open(filepath, "r") as f:
+            try:
+                res = json.load(f)
+            except Exception as e:
+                continue
+        accuracy = res.get("accuracy", 0.0)
+        total_tokens = res.get("total_tokens", 0)
+        rounds = res.get("rounds_to_converge", 6)
+        system_output = res.get("system_output", "")
+        ground_truth = res.get("ground_truth", "")
+        conv_hist = res.get("conversation_history", [])
+        conf_seq = []
+        conf_std_seq = []
+        agreement_seq = []
+        for r in range(7):
+            if r < len(conv_hist):
+                round_data = conv_hist[r]
+                confs = []
+                choices = []
+                for agent_data in round_data:
+                    output_dict = agent_data.get("output", {})
+                    if isinstance(output_dict, dict):
+                        conf_val = output_dict.get("confidence")
+                        ans_val = output_dict.get("answer")
+                        if conf_val is not None:
+                            confs.append(float(conf_val))
+                        if ans_val is not None:
+                            choices.append(ans_val)
+                if confs:
+                    conf_seq.append(np.mean(confs))
+                    conf_std_seq.append(np.std(confs))
+                else:
+                    conf_seq.append(np.nan)
+                    conf_std_seq.append(np.nan)
+                if choices:
+                    most_common = max(set(choices), key=choices.count)
+                    agree_ratio = choices.count(most_common) / len(choices)
+                    agreement_seq.append(agree_ratio)
+                else:
+                    agreement_seq.append(np.nan)
+            else:
+                conf_seq.append(np.nan)
+                conf_std_seq.append(np.nan)
+                agreement_seq.append(np.nan)
+        row = {
+            "topology": topology,
+            "case_id": case_id,
+            "accuracy": accuracy,
+            "total_tokens": total_tokens,
+            "rounds": rounds,
+            "system_output": system_output,
+            "ground_truth": ground_truth,
+        }
+        for r in range(7):
+            row[f"conf_r{r}"] = conf_seq[r]
+            row[f"conf_std_r{r}"] = conf_std_seq[r]
+            row[f"agree_r{r}"] = agreement_seq[r]
+        data.append(row)
+    return pd.DataFrame(data)
+def test_run_ab_details_analysis():
+    workspace_root = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+    dir_a = os.path.join(workspace_root, "results_compare_standard")
+    dir_b = os.path.join(workspace_root, "results_compare_random0")
+    df_a = load_all_json_details(dir_a)
+    df_b = load_all_json_details(dir_b)
+    assert len(df_a) > 0, "No standard comparison results found"
+    assert len(df_b) > 0, "No randomized comparison results found"
+    merged = pd.merge(df_a, df_b, on=["case_id", "topology"], suffixes=("_A", "_B"))
+    output_path = os.path.join(dir_b, "analyze_ab_details_output.txt")
+    with open(output_path, "w") as out:
+        out.write("======================================================================\n")
+        out.write("                 DETAILED GROUP A VS GROUP B STUDY                     \n")
+        out.write("======================================================================\n")
+        out.write(f"Group A (Standard) cases: {len(df_a)}\n")
+        out.write(f"Group B (Randomized Round 0) cases: {len(df_b)}\n")
+        out.write("----------------------------------------------------------------------\n\n")
+        # 1. Topology-wise comparison
+        topologies = ["star", "complete", "ring", "hierarchical"]
+        out.write("[1] METRICS OVERVIEW BY TOPOLOGY\n")
+        out.write(f"{'Topology':<15} | {'Acc A':<6} {'Acc B':<6} {'Diff':<6} | {'Tok A':<9} {'Tok B':<9} {'Savings %':<10} | {'Rnd A':<5} {'Rnd B':<5}\n")
+        out.write("-" * 88 + "\n")
+        for topo in topologies:
+            sub = merged[merged["topology"] == topo]
+            acc_a = sub["accuracy_A"].mean() if len(sub) > 0 else np.nan
+            acc_b = sub["accuracy_B"].mean() if len(sub) > 0 else np.nan
+            acc_diff = acc_b - acc_a
+            tok_a = sub["total_tokens_A"].mean() if len(sub) > 0 else np.nan
+            tok_b = sub["total_tokens_B"].mean() if len(sub) > 0 else np.nan
+            tok_sav = (tok_a - tok_b) / tok_a if tok_a > 0 else 0.0
+            rnd_a = sub["rounds_A"].mean() if len(sub) > 0 else np.nan
+            rnd_b = sub["rounds_B"].mean() if len(sub) > 0 else np.nan
+            out.write(f"{topo:<15} | {acc_a:<6.1%} {acc_b:<6.1%} {acc_diff:<+6.1%} | {tok_a:<9,.0f} {tok_b:<9,.0f} {tok_sav:<10.1%} | {rnd_a:<5.1f} {rnd_b:<5.1f}\n")
+        out.write("-" * 88 + "\n")
+        tot_acc_a = merged["accuracy_A"].mean()
+        tot_acc_b = merged["accuracy_B"].mean()
+        tot_tok_a = merged["total_tokens_A"].mean()
+        tot_tok_b = merged["total_tokens_B"].mean()
+        tot_sav = (tot_tok_a - tot_tok_b) / tot_tok_a if tot_tok_a > 0 else 0.0
+        out.write(f"{'OVERALL':<15} | {tot_acc_a:<6.1%} {tot_acc_b:<6.1%} {tot_acc_b - tot_acc_a:<+6.1%} | {tot_tok_a:<9,.0f} {tot_tok_b:<9,.0f} {tot_sav:<10.1%} |\n\n")
+        # 2. Case-by-case choice modifications
+        out.write("[2] DIAGNOSTIC OUTPUT CHANGE ANALYSIS\n")
+        out.write("We evaluate if Group B's deterministic randomized Round 0 changed the final output option:\n")
+        changed_outputs = merged[merged["system_output_A"] != merged["system_output_B"]]
+        out.write(f"Total topological runs where output changed: {len(changed_outputs)} / {len(merged)} ({len(changed_outputs)/len(merged):.1%})\n")
+        improved = merged[(merged["accuracy_A"] == 0.0) & (merged["accuracy_B"] == 1.0)]
+        worsened = merged[(merged["accuracy_A"] == 1.0) & (merged["accuracy_B"] == 0.0)]
+        both_correct = merged[(merged["accuracy_A"] == 1.0) & (merged["accuracy_B"] == 1.0)]
+        both_incorrect_same = merged[(merged["accuracy_A"] == 0.0) & (merged["accuracy_B"] == 0.0) & (merged["system_output_A"] == merged["system_output_B"])]
+        both_incorrect_diff = merged[(merged["accuracy_A"] == 0.0) & (merged["accuracy_B"] == 0.0) & (merged["system_output_A"] != merged["system_output_B"])]
+        out.write(f"  - Changed from INCORRECT to CORRECT (Improved): {len(improved)} runs\n")
+        out.write(f"  - Changed from CORRECT to INCORRECT (Worsened): {len(worsened)} runs\n")
+        out.write(f"  - Both CORRECT:                                {len(both_correct)} runs\n")
+        out.write(f"  - Both INCORRECT with the SAME wrong choice:   {len(both_incorrect_same)} runs\n")
+        out.write(f"  - Both INCORRECT with DIFFERENT wrong choices: {len(both_incorrect_diff)} runs\n\n")
+        # 3. Consensus Trajectory Details (Mean Confidence & Stdev)
+        out.write("[3] CONSENSUS TRAJECTORY DETAILS (MEAN & SD BY ROUND)\n")
+        for topo in topologies:
+            out.write(f"\nTopology: {topo.upper()}\n")
+            sub = merged[merged["topology"] == topo]
+            out.write(f"  {'Round':<6} | {'Avg Conf A':<12} {'Avg Conf B':<12} {'Diff':<7} | {'Conf SD A':<11} {'Conf SD B':<11} | {'Agree Ratio A':<13} {'Agree Ratio B':<13}\n")
+            out.write("  " + "-" * 105 + "\n")
+            for r in range(7):
+                mean_a = sub[f"conf_r{r}_A"].mean()
+                mean_b = sub[f"conf_r{r}_B"].mean()
+                mean_diff = mean_b - mean_a
+                std_a = sub[f"conf_std_r{r}_A"].mean()
+                std_b = sub[f"conf_std_r{r}_B"].mean()
+                agree_a = sub[f"agree_r{r}_A"].mean()
+                agree_b = sub[f"agree_r{r}_B"].mean()
+                str_mean_a = f"{mean_a:.4f}" if not np.isnan(mean_a) else "N/A"
+                str_mean_b = f"{mean_b:.4f}" if not np.isnan(mean_b) else "N/A"
+                str_mean_diff = f"{mean_diff:+.4f}" if not np.isnan(mean_diff) else "N/A"
+                str_std_a = f"{std_a:.4f}" if not np.isnan(std_a) else "N/A"
+                str_std_b = f"{std_b:.4f}" if not np.isnan(std_b) else "N/A"
+                str_agree_a = f"{agree_a:.1%}" if not np.isnan(agree_a) else "N/A"
+                str_agree_b = f"{agree_b:.1%}" if not np.isnan(agree_b) else "N/A"
+                out.write(f"  Round {r:<1} | {str_mean_a:<12} {str_mean_b:<12} {str_mean_diff:<7} | {str_std_a:<11} {str_std_b:<11} | {str_agree_a:<13} {str_agree_b:<13}\n")
+    print(f"Detailed AB analysis written to {output_path}")