Deploy FrontierLab v0.40 source 6f0f7d243154
Browse files- app.py +6 -3
- deployment.json +12 -10
- docs/DEMO_RUNBOOK.md +6 -5
- docs/DEMO_VIDEO_SCRIPT.md +141 -0
- docs/HF_INFERENCE_PROVIDERS_MUSE.md +7 -5
- docs/RESEARCH_OPERATING_INTERFACE.zh-TW.md +10 -9
- docs/SUBMISSION_COPY.md +43 -8
- docs/runs/hf_muse_live_extraction_20261101.json +171 -0
- frontierlab/hf_providers.py +80 -10
- frontierlab/version.py +1 -1
app.py
CHANGED
|
@@ -462,7 +462,8 @@ with st.sidebar:
|
|
| 462 |
f"Primary: {hf_providers.PRIMARY_MODEL} · automatic fallback: {hf_providers.FALLBACK_MODEL}. "
|
| 463 |
f"Session calls left: {hf_session_budget.remaining} / {hf_session_budget.limit} · "
|
| 464 |
f"server calls left today: {hf_providers.daily_budget().remaining} / {hf_providers.daily_budget().limit}. "
|
| 465 |
-
"Every attempt, including a fallback, counts.
|
|
|
|
| 466 |
f"{hf_providers.MAX_LIVE_ABSTRACTS} abstracts, each truncated to "
|
| 467 |
f"{hf_providers.MAX_ABSTRACT_CHARS} characters. Quotes still pass the verbatim evidence gate "
|
| 468 |
"and every step stays human-gated."
|
|
@@ -509,8 +510,10 @@ if run_clicked:
|
|
| 509 |
friendly = europepmc_error_message(exc)
|
| 510 |
st.error(friendly if friendly else f"Research run failed: {type(exc).__name__}: {exc}")
|
| 511 |
if use_hf and client is not None and "frontier_result" not in st.session_state:
|
| 512 |
-
|
| 513 |
-
st.
|
|
|
|
|
|
|
| 514 |
|
| 515 |
with st.expander("Published-literature pilot · 12 cases awaiting researcher annotation"):
|
| 516 |
pilot = load_dataset()
|
|
|
|
| 462 |
f"Primary: {hf_providers.PRIMARY_MODEL} · automatic fallback: {hf_providers.FALLBACK_MODEL}. "
|
| 463 |
f"Session calls left: {hf_session_budget.remaining} / {hf_session_budget.limit} · "
|
| 464 |
f"server calls left today: {hf_providers.daily_budget().remaining} / {hf_providers.daily_budget().limit}. "
|
| 465 |
+
"Every attempt, including a fallback, counts. Both models reason before answering, so "
|
| 466 |
+
"one call can take up to 90 s (fallback 75 s). At most "
|
| 467 |
f"{hf_providers.MAX_LIVE_ABSTRACTS} abstracts, each truncated to "
|
| 468 |
f"{hf_providers.MAX_ABSTRACT_CHARS} characters. Quotes still pass the verbatim evidence gate "
|
| 469 |
"and every step stays human-gated."
|
|
|
|
| 510 |
friendly = europepmc_error_message(exc)
|
| 511 |
st.error(friendly if friendly else f"Research run failed: {type(exc).__name__}: {exc}")
|
| 512 |
if use_hf and client is not None and "frontier_result" not in st.session_state:
|
| 513 |
+
for line in hf_providers.failure_summary(getattr(client, "provenance", None)):
|
| 514 |
+
st.caption("Failed attempt · " + line)
|
| 515 |
+
with st.expander("Live Meta Muse call receipts (failed run)", expanded=True):
|
| 516 |
+
st.json(getattr(client, "provenance", {}))
|
| 517 |
|
| 518 |
with st.expander("Published-literature pilot · 12 cases awaiting researcher annotation"):
|
| 519 |
pilot = load_dataset()
|
deployment.json
CHANGED
|
@@ -1,29 +1,30 @@
|
|
| 1 |
{
|
| 2 |
-
"version": "v0.
|
| 3 |
-
"source_sha": "
|
| 4 |
"space_id": "Jaycee766/FrontierLab-Meta-Hackathon-2026",
|
| 5 |
-
"built_at_utc": "2026-11-
|
| 6 |
"file_sha256": {
|
| 7 |
".streamlit/config.toml": "e1dd42acf57c5311de54b5fc88c4669cfb5142f6d4caf65af2de62e4dc812df9",
|
| 8 |
"Dockerfile": "fb2c4f26e26e6fdc92716f8b10e755d6fdb421fa91ac295fe76c723f3534a19f",
|
| 9 |
"LICENSE": "ce6f35ee0aeea8b99f5bbc542fdcad58aca68859b6e9bc0e82d76e4bf8bcbddb",
|
| 10 |
"NOTICE": "65f86d9ef38f67db4196acc346bfe9f1e87fc6ec5df5cd12d9be0f9cb117ef85",
|
| 11 |
"README.md": "106c1f481f22c8e2d0bd3c73b43b7fbe5d6ee23db4ec4a5043057c6959c70548",
|
| 12 |
-
"app.py": "
|
| 13 |
"datasets/real_ph_excerpt_cases_v1.json": "725d4502bf987f69e3fb4bf2babe806d2e21b42bb2b77bffa68df2086fb54417",
|
| 14 |
"docs/AGENTIC_AI_LANDSCAPE.md": "0c5a6aaf185b01ad9f94d138c4610489c9d917a16c91fa2deb4106d28c7ad5a2",
|
| 15 |
"docs/AGENT_AVAILABILITY_STATEMENT.md": "36f282dd49e683ca3c6c251e9d3e43aeb3a4f2fdf6b518d5044f5425894de5b6",
|
| 16 |
"docs/BIONEMO_SCIENCE_TOOLS.md": "ad7cde9b2237b98d8b65adb336f1c03875a99f25b2da594f53c8c4b74d8f7ae9",
|
| 17 |
"docs/CLAIM_VERIFICATION_AND_NIM.md": "230c1e9d2cc9e340fe64d03f188950016466b921b78494e8215c7e719fa3c813",
|
| 18 |
"docs/D1_MULTIMODAL_EVIDENCE_GATE.md": "1f4846c5bdfd8dba820821b4726b08fe5f6b7f6e5ac7c3cf534b8e1228d5fc8c",
|
| 19 |
-
"docs/DEMO_RUNBOOK.md": "
|
|
|
|
| 20 |
"docs/DETECTORSCOPE_AUDIT.md": "a4730c1323976b797f7df9ade8ff9d84aa8cbd816f2d88b3fae010db6022e537",
|
| 21 |
"docs/ERROR_TAXONOMY.md": "f1a8602229c25ea4234f8ad8d5d6889fc1ba36beb5ae29084505ef16744a20a9",
|
| 22 |
"docs/EVIDENCE_GRAPH.md": "3e3bc44d4a3535f4f8e265c30a3503b2f8ba3a33ce8ffb6697fe9b818f1433fd",
|
| 23 |
"docs/GATE_ABLATION.md": "e9a4907ef2c8555ca6041628e98a99ea822f816ae453c3f11a1a3342c784dec6",
|
| 24 |
"docs/HACKATHON_FINAL_SPRINT.md": "a07efd000bfd5193d389c073be33c4277387e0d0b55329c666009037afdc9842",
|
| 25 |
"docs/HF_DEPLOYMENT.md": "b9536aef203b914eb2c787dc8c33bd1a045f87663b4d807d67b17a3bc3df7580",
|
| 26 |
-
"docs/HF_INFERENCE_PROVIDERS_MUSE.md": "
|
| 27 |
"docs/HF_PRO_SETUP.md": "34033f388bb93a50524d76741e9a4524432c95951a40114469e1df9cb93068af",
|
| 28 |
"docs/HUMAN_REVIEW_GATE.md": "279b161d76496019dd6d64cf2c9cc8273dc37e4c3a45545905fc93946b207e4c",
|
| 29 |
"docs/INFORMATION_MANAGEMENT_EXTENSION.md": "05bd531fe11acca8c41f7efd5cc85a265a34924f10fdc362110aa642b393a75a",
|
|
@@ -47,11 +48,11 @@
|
|
| 47 |
"docs/RELEASE_CHECKLIST.md": "e27f39291682fd3150a1bf1ec874dae9cfbf8b6473e6f51aeb3a7261e515504e",
|
| 48 |
"docs/RESEARCH_FAILURE_LAB.md": "085554767df311a488f59cd2ee865c99d43d180341900acbb7d370731c6f2aaa",
|
| 49 |
"docs/RESEARCH_OPERATING_INTERFACE.md": "eb6dfd4b388509ba819114bad2ccaab06a1838b6e29945945311a8588520c066",
|
| 50 |
-
"docs/RESEARCH_OPERATING_INTERFACE.zh-TW.md": "
|
| 51 |
"docs/RESOURCE_STRATEGY.md": "cbed1ec8d58e3dabd043cd8331f7be73d5bbc230c7cedcc921d1a4e9e30bc3bc",
|
| 52 |
"docs/SPRINGER_NATURE_META.md": "687176f2e1d4ec5477e4fc075b3bfcac979e1dd5d63e7a0f821ab5f69af55dad",
|
| 53 |
"docs/SPRINGER_NATURE_OA.md": "ec956938f89a9023d9ca475e28f5fec028ee0d0dd7db77377263c6b61c189f1c",
|
| 54 |
-
"docs/SUBMISSION_COPY.md": "
|
| 55 |
"docs/TRACE_COVERAGE.md": "8d8cf03077c632b68bbb650c553d836fb51998e4cde6f5eed969c26f2bbea011",
|
| 56 |
"docs/TRAJECTORY_GRAPH.md": "aafba7abeecfb139342d56be146a0c2cbcde35f704053aafc9fc9700807b769f",
|
| 57 |
"docs/VERBATIM_EVIDENCE_GATE.md": "e0e8f2325738267a4ca3dcc3c23616daeb0395752b92b56240b62aff2f4f3f8e",
|
|
@@ -60,6 +61,7 @@
|
|
| 60 |
"docs/assets/flatland-sphere-icon.png": "b1684a4ee4696fa2def476de1970e0187503c2bf83bc0d4baa031e47a0e4c7ce",
|
| 61 |
"docs/assets/flatland-sphere.svg": "51ec3ba0d7428bf516ba48826c57065b567e6f5a499b9bf15cb4d929ee68aa51",
|
| 62 |
"docs/runs/hf_deployment_20261028.json": "073d5134b9c47e943c0aecd5259b9417aba5004a0be6f59670960227f242d2ea",
|
|
|
|
| 63 |
"docs/runs/maintenance_baseline_20261101.json": "3c82e7d4ec231b75ee4ca563d50c7ecda04618cb4eb4346fb54164870988ac3a",
|
| 64 |
"docs/runs/meta_nim_json_20261028.json": "d388d7650eac6c9792d9b13617ef6c51273d21b37018b5c1ebc5244713845d80",
|
| 65 |
"docs/runs/meta_nim_pilot_20261028.json": "e5f7c1337ec01ebb99861b0d3d0681eb985ac5f6d33ee07fb1a30b043415b8c3",
|
|
@@ -88,7 +90,7 @@
|
|
| 88 |
"frontierlab/evidence_graph.py": "c04e22fa270b84e013c71af8056fc91a1d894559d57b8ae3b8bf8c9fe8395d13",
|
| 89 |
"frontierlab/gate_ablation.py": "9ed8c372a938ff41ba876e5d3dbc3a17971360e2a16d922add7fca69b7282425",
|
| 90 |
"frontierlab/hardware.py": "fe983822b632eead0ceb59db77856010954047524b23c0d17d65b1704f245060",
|
| 91 |
-
"frontierlab/hf_providers.py": "
|
| 92 |
"frontierlab/integrity_trials.py": "d63a6a86908df045de6f9262acb4d18589b34e0148e9d3b1c05adedbb403ca10",
|
| 93 |
"frontierlab/interventions.py": "233e13a87bebebb2fb16be612c132098b7629a755936490a4d65c1f8ff3f954c",
|
| 94 |
"frontierlab/isolation.py": "405671ccb7944fc36a108af7ece2054290db2e056aea4cecd161a87762d24845",
|
|
@@ -119,7 +121,7 @@
|
|
| 119 |
"frontierlab/springer_oa.py": "1f2d23a9b528379856aa7e824a7d03826a406c5a93094d4d26e957bd68275747",
|
| 120 |
"frontierlab/traceability.py": "cf15d85d4d7c40cbe9316eac93ed3e4be7d3b9b836be401b447425d81dbc04d6",
|
| 121 |
"frontierlab/trajectory_graph.py": "d13c63c7c2b88920438deb1355a5bae480c777259f0f2b18aa089e0c5461cafa",
|
| 122 |
-
"frontierlab/version.py": "
|
| 123 |
"requirements.txt": "df0f6c06420a638466e7a4da630551bf37c15d1b543aff701bbc608e42197bfb"
|
| 124 |
},
|
| 125 |
"nvidia_secret_transferred": false,
|
|
|
|
| 1 |
{
|
| 2 |
+
"version": "v0.40",
|
| 3 |
+
"source_sha": "6f0f7d2431548621c3f2481939823058d2b9a72d",
|
| 4 |
"space_id": "Jaycee766/FrontierLab-Meta-Hackathon-2026",
|
| 5 |
+
"built_at_utc": "2026-11-02T07:44:13.260397+00:00",
|
| 6 |
"file_sha256": {
|
| 7 |
".streamlit/config.toml": "e1dd42acf57c5311de54b5fc88c4669cfb5142f6d4caf65af2de62e4dc812df9",
|
| 8 |
"Dockerfile": "fb2c4f26e26e6fdc92716f8b10e755d6fdb421fa91ac295fe76c723f3534a19f",
|
| 9 |
"LICENSE": "ce6f35ee0aeea8b99f5bbc542fdcad58aca68859b6e9bc0e82d76e4bf8bcbddb",
|
| 10 |
"NOTICE": "65f86d9ef38f67db4196acc346bfe9f1e87fc6ec5df5cd12d9be0f9cb117ef85",
|
| 11 |
"README.md": "106c1f481f22c8e2d0bd3c73b43b7fbe5d6ee23db4ec4a5043057c6959c70548",
|
| 12 |
+
"app.py": "56548384e533181e842a9174c19acde77a77aaa90afc642e538159e54977bbc2",
|
| 13 |
"datasets/real_ph_excerpt_cases_v1.json": "725d4502bf987f69e3fb4bf2babe806d2e21b42bb2b77bffa68df2086fb54417",
|
| 14 |
"docs/AGENTIC_AI_LANDSCAPE.md": "0c5a6aaf185b01ad9f94d138c4610489c9d917a16c91fa2deb4106d28c7ad5a2",
|
| 15 |
"docs/AGENT_AVAILABILITY_STATEMENT.md": "36f282dd49e683ca3c6c251e9d3e43aeb3a4f2fdf6b518d5044f5425894de5b6",
|
| 16 |
"docs/BIONEMO_SCIENCE_TOOLS.md": "ad7cde9b2237b98d8b65adb336f1c03875a99f25b2da594f53c8c4b74d8f7ae9",
|
| 17 |
"docs/CLAIM_VERIFICATION_AND_NIM.md": "230c1e9d2cc9e340fe64d03f188950016466b921b78494e8215c7e719fa3c813",
|
| 18 |
"docs/D1_MULTIMODAL_EVIDENCE_GATE.md": "1f4846c5bdfd8dba820821b4726b08fe5f6b7f6e5ac7c3cf534b8e1228d5fc8c",
|
| 19 |
+
"docs/DEMO_RUNBOOK.md": "2ed9cb2cccbe636ad9b8abd813155073806885e825d99555fb44589f928faf37",
|
| 20 |
+
"docs/DEMO_VIDEO_SCRIPT.md": "da55d78364a03883f05bce2d1632d8c48b4e83bd858768f6ee087d78118870ff",
|
| 21 |
"docs/DETECTORSCOPE_AUDIT.md": "a4730c1323976b797f7df9ade8ff9d84aa8cbd816f2d88b3fae010db6022e537",
|
| 22 |
"docs/ERROR_TAXONOMY.md": "f1a8602229c25ea4234f8ad8d5d6889fc1ba36beb5ae29084505ef16744a20a9",
|
| 23 |
"docs/EVIDENCE_GRAPH.md": "3e3bc44d4a3535f4f8e265c30a3503b2f8ba3a33ce8ffb6697fe9b818f1433fd",
|
| 24 |
"docs/GATE_ABLATION.md": "e9a4907ef2c8555ca6041628e98a99ea822f816ae453c3f11a1a3342c784dec6",
|
| 25 |
"docs/HACKATHON_FINAL_SPRINT.md": "a07efd000bfd5193d389c073be33c4277387e0d0b55329c666009037afdc9842",
|
| 26 |
"docs/HF_DEPLOYMENT.md": "b9536aef203b914eb2c787dc8c33bd1a045f87663b4d807d67b17a3bc3df7580",
|
| 27 |
+
"docs/HF_INFERENCE_PROVIDERS_MUSE.md": "e275598ef54b15fea3bfebd483f7a0607d7ab8e92fefdf8576d980fe5f993b32",
|
| 28 |
"docs/HF_PRO_SETUP.md": "34033f388bb93a50524d76741e9a4524432c95951a40114469e1df9cb93068af",
|
| 29 |
"docs/HUMAN_REVIEW_GATE.md": "279b161d76496019dd6d64cf2c9cc8273dc37e4c3a45545905fc93946b207e4c",
|
| 30 |
"docs/INFORMATION_MANAGEMENT_EXTENSION.md": "05bd531fe11acca8c41f7efd5cc85a265a34924f10fdc362110aa642b393a75a",
|
|
|
|
| 48 |
"docs/RELEASE_CHECKLIST.md": "e27f39291682fd3150a1bf1ec874dae9cfbf8b6473e6f51aeb3a7261e515504e",
|
| 49 |
"docs/RESEARCH_FAILURE_LAB.md": "085554767df311a488f59cd2ee865c99d43d180341900acbb7d370731c6f2aaa",
|
| 50 |
"docs/RESEARCH_OPERATING_INTERFACE.md": "eb6dfd4b388509ba819114bad2ccaab06a1838b6e29945945311a8588520c066",
|
| 51 |
+
"docs/RESEARCH_OPERATING_INTERFACE.zh-TW.md": "752aa81b8eb43b85cedc6f62e18a8a6ff9097678a40ab7f04a1f6267be465f23",
|
| 52 |
"docs/RESOURCE_STRATEGY.md": "cbed1ec8d58e3dabd043cd8331f7be73d5bbc230c7cedcc921d1a4e9e30bc3bc",
|
| 53 |
"docs/SPRINGER_NATURE_META.md": "687176f2e1d4ec5477e4fc075b3bfcac979e1dd5d63e7a0f821ab5f69af55dad",
|
| 54 |
"docs/SPRINGER_NATURE_OA.md": "ec956938f89a9023d9ca475e28f5fec028ee0d0dd7db77377263c6b61c189f1c",
|
| 55 |
+
"docs/SUBMISSION_COPY.md": "1381783abac79ca10a7b28c1c95f14802f7957ed7c335ceb2a368fa00171be6d",
|
| 56 |
"docs/TRACE_COVERAGE.md": "8d8cf03077c632b68bbb650c553d836fb51998e4cde6f5eed969c26f2bbea011",
|
| 57 |
"docs/TRAJECTORY_GRAPH.md": "aafba7abeecfb139342d56be146a0c2cbcde35f704053aafc9fc9700807b769f",
|
| 58 |
"docs/VERBATIM_EVIDENCE_GATE.md": "e0e8f2325738267a4ca3dcc3c23616daeb0395752b92b56240b62aff2f4f3f8e",
|
|
|
|
| 61 |
"docs/assets/flatland-sphere-icon.png": "b1684a4ee4696fa2def476de1970e0187503c2bf83bc0d4baa031e47a0e4c7ce",
|
| 62 |
"docs/assets/flatland-sphere.svg": "51ec3ba0d7428bf516ba48826c57065b567e6f5a499b9bf15cb4d929ee68aa51",
|
| 63 |
"docs/runs/hf_deployment_20261028.json": "073d5134b9c47e943c0aecd5259b9417aba5004a0be6f59670960227f242d2ea",
|
| 64 |
+
"docs/runs/hf_muse_live_extraction_20261101.json": "04939dd3f5f5b91d791b4bfef19af62653406ef25219f4fda720730b969d4a4d",
|
| 65 |
"docs/runs/maintenance_baseline_20261101.json": "3c82e7d4ec231b75ee4ca563d50c7ecda04618cb4eb4346fb54164870988ac3a",
|
| 66 |
"docs/runs/meta_nim_json_20261028.json": "d388d7650eac6c9792d9b13617ef6c51273d21b37018b5c1ebc5244713845d80",
|
| 67 |
"docs/runs/meta_nim_pilot_20261028.json": "e5f7c1337ec01ebb99861b0d3d0681eb985ac5f6d33ee07fb1a30b043415b8c3",
|
|
|
|
| 90 |
"frontierlab/evidence_graph.py": "c04e22fa270b84e013c71af8056fc91a1d894559d57b8ae3b8bf8c9fe8395d13",
|
| 91 |
"frontierlab/gate_ablation.py": "9ed8c372a938ff41ba876e5d3dbc3a17971360e2a16d922add7fca69b7282425",
|
| 92 |
"frontierlab/hardware.py": "fe983822b632eead0ceb59db77856010954047524b23c0d17d65b1704f245060",
|
| 93 |
+
"frontierlab/hf_providers.py": "b594c5e16dffb66bbe319089e324b716464fa5266ee2777bfbfb2f13c1c95010",
|
| 94 |
"frontierlab/integrity_trials.py": "d63a6a86908df045de6f9262acb4d18589b34e0148e9d3b1c05adedbb403ca10",
|
| 95 |
"frontierlab/interventions.py": "233e13a87bebebb2fb16be612c132098b7629a755936490a4d65c1f8ff3f954c",
|
| 96 |
"frontierlab/isolation.py": "405671ccb7944fc36a108af7ece2054290db2e056aea4cecd161a87762d24845",
|
|
|
|
| 121 |
"frontierlab/springer_oa.py": "1f2d23a9b528379856aa7e824a7d03826a406c5a93094d4d26e957bd68275747",
|
| 122 |
"frontierlab/traceability.py": "cf15d85d4d7c40cbe9316eac93ed3e4be7d3b9b836be401b447425d81dbc04d6",
|
| 123 |
"frontierlab/trajectory_graph.py": "d13c63c7c2b88920438deb1355a5bae480c777259f0f2b18aa089e0c5461cafa",
|
| 124 |
+
"frontierlab/version.py": "bf6ebe0d35d1d665b8be3508a99cf145d12ba7b6a8ddfe747286bf0c19f1d80f",
|
| 125 |
"requirements.txt": "df0f6c06420a638466e7a4da630551bf37c15d1b543aff701bbc608e42197bfb"
|
| 126 |
},
|
| 127 |
"nvidia_secret_transferred": false,
|
docs/DEMO_RUNBOOK.md
CHANGED
|
@@ -13,12 +13,13 @@ Use a revision that has passed CI and note the exact source SHA.
|
|
| 13 |
|
| 14 |
Current software baseline at preparation time:
|
| 15 |
|
| 16 |
-
- FrontierLab v0.
|
| 17 |
-
-
|
| 18 |
-
- **
|
| 19 |
-
- Maintenance Canary: **HEALTHY / HEALTHY_STABLE**, no accepted-baseline drift, advisory result **NO_REPAIR_NEEDED** (main run 41298101823; artifact 13079632919)
|
| 20 |
- Stage 5 status: **readiness complete; planning not approved or generated**
|
| 21 |
-
- currently deployed HF baseline is **v0.
|
|
|
|
|
|
|
| 22 |
|
| 23 |
Do not call a recorded result "live". Do not call researcher review "independent gold". Do not call software tests model accuracy.
|
| 24 |
|
|
|
|
| 13 |
|
| 14 |
Current software baseline at preparation time:
|
| 15 |
|
| 16 |
+
- FrontierLab v0.39 on main (merge `5958770f581d9afa4fc8054172f888941c4d46ee`; later docs-only commits may change the final main SHA without changing runtime code)
|
| 17 |
+
- **398 software tests passed after merge** (run 41311197281)
|
| 18 |
+
- Maintenance Canary: **HEALTHY** (398 tests), advisory result **DRIFT_REVIEW_REQUIRED / HEALTHY_DRIFT** because the `requirements.txt` hash differs from the accepted Nov 1 baseline (main run 41311197269); accepting a new baseline is the owner's decision
|
|
|
|
| 19 |
- Stage 5 status: **readiness complete; planning not approved or generated**
|
| 20 |
+
- currently deployed HF baseline is **v0.39**: source `5958770…` → Space commit `1e162166…` (run 41311275823, 398 tests, receipt BUILD_PENDING; public Space API later showed runtime RUNNING with a v0.39 source marker). Re-verify the rendered source marker and the sidebar's live-Muse status before recording; any later merge needs one more deploy for exact parity
|
| 21 |
+
|
| 22 |
+
The shot-by-shot video script that follows the in-app Judge path (recorded replay → live Meta Muse via Hugging Face → provenance/governance) is in [DEMO_VIDEO_SCRIPT.md](DEMO_VIDEO_SCRIPT.md). This runbook remains the claims checklist.
|
| 23 |
|
| 24 |
Do not call a recorded result "live". Do not call researcher review "independent gold". Do not call software tests model accuracy.
|
| 25 |
|
docs/DEMO_VIDEO_SCRIPT.md
ADDED
|
@@ -0,0 +1,141 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# FrontierLab — 3-minute demo video script
|
| 2 |
+
|
| 3 |
+
**Target length:** ~3:00 (internal production target; recheck the participant portal for any official limit).
|
| 4 |
+
**Follows:** the in-app **Judge path · start here** box, in this order: (1) *Start here · Replay recorded real M2 case (no model call)* → (2) live Meta Muse via Hugging Face, only if the Space shows it available → (3) provenance and governance views.
|
| 5 |
+
**Claims checklist:** [DEMO_RUNBOOK.md](DEMO_RUNBOOK.md). Every sentence below must stay true of the build you record.
|
| 6 |
+
|
| 7 |
+
Judging weights to keep in mind while cutting: **Impact 30% · Use of Meta AI tooling 25% · Technical depth 20% · Demo 25%**. The judges watch the video itself, so picture, sound and pacing count as much as the content.
|
| 8 |
+
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
## Before you press record
|
| 12 |
+
|
| 13 |
+
| Check | How | Pass condition |
|
| 14 |
+
| --- | --- | --- |
|
| 15 |
+
| Deployed version | Open `https://huggingface.co/spaces/Jaycee766/FrontierLab-Meta-Hackathon-2026/raw/main/deployment.json` | `version` and `source_sha` match the revision you describe (v0.39: `5958770…`) |
|
| 16 |
+
| Live Muse status | App sidebar, mode list | Reads **Live Meta Muse (via Hugging Face)** with *no* "· unavailable" suffix. If it still says "unavailable (no Hugging Face token configured)", use **Plan B** in Shot 5 |
|
| 17 |
+
| Call budget | Sidebar caption under the HF mode | "Session calls left: 6 / 6". Use a fresh browser session for the take |
|
| 18 |
+
| Spending limit | Hugging Face account billing | A spending limit is set (the app caps are not HF billing limits) |
|
| 19 |
+
| Screen | Browser zoom 110–125%, 1920×1080, light theme, notifications off, bookmarks bar hidden | Text readable on a phone |
|
| 20 |
+
| Audio | External or headset mic, quiet room | No clipping, no room echo |
|
| 21 |
+
|
| 22 |
+
Do one full live rehearsal **off camera** so you know what the model returns, then record. Never paste or re-type a model answer and call it live.
|
| 23 |
+
|
| 24 |
+
## Visual style
|
| 25 |
+
|
| 26 |
+
Black and white, the Flatland homage used across the app and README: white background, black type, the hero and architecture SVGs in `docs/assets/`. Title cards in plain black sans-serif on white. No stock footage, no music with lyrics; if you use music, keep it quiet and instrumental. Burn in English captions (judges often watch muted).
|
| 27 |
+
|
| 28 |
+
---
|
| 29 |
+
|
| 30 |
+
## Shot list and narration
|
| 31 |
+
|
| 32 |
+
Times are cumulative. Narration is about 140 words per minute; the bracketed **[zh-TW]** notes are optional guidance for a Traditional Chinese voice-over or subtitles, not a word-for-word translation.
|
| 33 |
+
|
| 34 |
+
### Shot 1 · 0:00–0:15 · Hook
|
| 35 |
+
|
| 36 |
+
**Screen:** `docs/assets/flatland-hero.svg` full screen (line → triangle → square → hexagon → sphere), then cut to the app landing page.
|
| 37 |
+
**On-screen text:** "A real citation is not the same as real support."
|
| 38 |
+
|
| 39 |
+
> **Narration:** "A language model can cite a real paper and still claim something the paper never says. In biomedical research, that can steer which experiment gets run. FrontierLab is my answer: an evidence-grounded, human-gated research pipeline built on Meta Muse models."
|
| 40 |
+
|
| 41 |
+
[zh-TW] 開場點出問題:模型可以引用真的論文,卻說出論文沒寫的事;FrontierLab 用 Meta Muse 模型,但每一步都要有證據、要經過研究者把關。
|
| 42 |
+
|
| 43 |
+
### Shot 2 · 0:15–0:30 · The Judge path
|
| 44 |
+
|
| 45 |
+
**Screen:** The **Judge path · start here (about 3 minutes)** box at the top of the app. Slowly highlight steps 1–3 (cursor or a light zoom).
|
| 46 |
+
|
| 47 |
+
> **Narration:** "Five stages — Research, Evidence, Hypothesis, Evaluation, Planning. Models propose; deterministic checks and the researcher decide what moves forward. Here's the judge path: a recorded real case, a live Meta Muse run, then the provenance behind both."
|
| 48 |
+
|
| 49 |
+
[zh-TW] 簡介五個階段,說明「模型提出、規則與研究者決定」。
|
| 50 |
+
|
| 51 |
+
### Shot 3 · 0:30–1:05 · Replay the recorded real M2 case
|
| 52 |
+
|
| 53 |
+
**Screen:** The expanded **Start here · Replay recorded real M2 case (no model call)** panel. Hold on the blue **RECORDED REAL RUN · REPLAY ONLY** banner for two seconds, then on the four metrics.
|
| 54 |
+
**Call-outs (zoom or box):** *Source trace coverage 100%* · *Strict EXTRACTED coverage 50%* · *Planning obligations 9* · *Model calls in this replay 0*.
|
| 55 |
+
|
| 56 |
+
> **Narration:** "This is a real run, recorded on October 28th: Meta Muse Glimmer 30B read two published antibody-engineering abstracts, every claim was reviewed by me, and I selected method M2. Opening it makes zero model calls — every number is recomputed from the recorded files. All four artifacts trace back to a paper: 100 percent. But only two of four have a fully extracted path: 50 percent. That gap is a real disagreement between the model and me, and FrontierLab keeps it visible instead of smoothing it over."
|
| 57 |
+
|
| 58 |
+
**Must say on screen or aloud:** "recorded", "replay", "provenance, not accuracy".
|
| 59 |
+
|
| 60 |
+
[zh-TW] 強調這是「已記錄的真實執行」,重播時不呼叫模型;100% 是來源可追溯,50% 是嚴格 EXTRACTED,差距保留了模型與研究者的真實分歧。這是溯源指標,不是準確度。
|
| 61 |
+
|
| 62 |
+
### Shot 4 · 1:05–1:25 · Readiness is not approval
|
| 63 |
+
|
| 64 |
+
**Screen:** Replay tabs → **Planning readiness**: open data / code / environment checks, the 9 obligations, `Planning approved: NO`. Then a 3-second glance at the **Skeptic → Hypothesis** tab.
|
| 65 |
+
|
| 66 |
+
> **Narration:** "The researcher chose M2, but readiness is not approval. Open data, code and environment checks travel into Stage 5, along with nine obligations — like monomeric-Fab SPR so an avidity effect isn't mistaken for intrinsic affinity. Planning stays locked until a human authorizes it."
|
| 67 |
+
|
| 68 |
+
**Do not** tick the planning-approval checkbox for the video.
|
| 69 |
+
|
| 70 |
+
[zh-TW] 研究者選了 M2,但「準備好」不等於「核准」;九項義務(例如 monomeric-Fab SPR)會帶入第五階段,規劃仍未核准。
|
| 71 |
+
|
| 72 |
+
### Shot 5 · 1:25–2:10 · Live Meta Muse via Hugging Face
|
| 73 |
+
|
| 74 |
+
**Plan A — the sidebar shows the mode as available.**
|
| 75 |
+
|
| 76 |
+
**Screen:**
|
| 77 |
+
1. Sidebar → select **Live Meta Muse (via Hugging Face)**. Show the caption "Primary: meta-models/Muse-Spark-Open-816B-A42B:baseten · automatic fallback: meta-models/Muse-Glimmer-30B" and the call counters.
|
| 78 |
+
2. Tick the consent checkbox (let viewers read it).
|
| 79 |
+
3. Keep *Maximum live abstracts* at 2 → **Run research audit**. Cut the waiting time; show a small "time skipped" label if you cut.
|
| 80 |
+
4. Show **Evidence claims and citation checks**: accepted claims with quotes, any rejected claim and its reason.
|
| 81 |
+
5. Open **Model request receipts**: model ID, provider, latency, tokens, status.
|
| 82 |
+
6. Show the run stopping at **Researcher claim review · required before synthesis**.
|
| 83 |
+
|
| 84 |
+
> **Narration:** "Now live. FrontierLab calls Meta Muse Spark Open through Hugging Face Inference Providers, with automatic fallback to Muse Glimmer 30B. Each visitor has to consent, and every session is capped at six calls. Every claim has to quote its source word for word, or it's rejected before synthesis. Each call leaves a receipt — model, provider, tokens, latency — and the token never appears. And the run stops right here: nothing moves forward until a researcher reviews these claims."
|
| 85 |
+
|
| 86 |
+
Say exactly which model the receipts show. If the receipt shows the Glimmer fallback, say "the primary failed and it fell back to Muse Glimmer — that's the fallback doing its job."
|
| 87 |
+
|
| 88 |
+
Optional, only if time and budget allow: review the claims, click **Resume model synthesis after human review**, and show the live Skeptic → Hypothesis output.
|
| 89 |
+
|
| 90 |
+
**Plan B — the sidebar still says "unavailable (no Hugging Face token configured)", or the live call fails.**
|
| 91 |
+
|
| 92 |
+
**Screen:** Show the sidebar label as it is. Then the receipt panel or error message if a call failed. Then cut back to the recorded replay.
|
| 93 |
+
|
| 94 |
+
> **Narration:** "On this deployment the live Meta Muse path is turned off: no token is configured, so the app labels it unavailable instead of faking an answer. The recorded run you just saw came from the same pipeline and the same gates."
|
| 95 |
+
|
| 96 |
+
Never record Plan A with pre-written output. A visible, honest fallback is stronger than a fake live demo.
|
| 97 |
+
|
| 98 |
+
After a successful live take, download the provenance JSON and commit it under `docs/runs/` before describing the live result anywhere else (owner checklist in [HF_INFERENCE_PROVIDERS_MUSE.md](HF_INFERENCE_PROVIDERS_MUSE.md)).
|
| 99 |
+
|
| 100 |
+
[zh-TW] 現場示範:Muse Spark Open 為主、Muse Glimmer 30B 自動備援;需使用者同意、每個 session 限 6 次;每個主張都要逐字引用,否則被擋下;每次呼叫都有收據,且不顯示 token;結果停在人工審查。若顯示 unavailable,就如實說明並回到已記錄的重播。
|
| 101 |
+
|
| 102 |
+
### Shot 6 · 2:10–2:40 · Provenance and governance
|
| 103 |
+
|
| 104 |
+
**Screen (about 7 seconds each):**
|
| 105 |
+
1. **Research Operating Interface** tab: next required stage, the two trace-coverage metrics, "Auto repair: OFF", and the note that provenance disagreements are preserved rather than promoted to strict evidence.
|
| 106 |
+
2. **Evidence graph**: Paper ← Passage ← Claim ← Hypothesis, with EXTRACTED / INFERRED / AMBIGUOUS tags.
|
| 107 |
+
3. **Trajectory** tab: model calls, human interventions and method evaluation as an audit-only graph.
|
| 108 |
+
4. GitHub Actions page for Maintenance canary run 41311197269: `HEALTHY`, advisory `DRIFT_REVIEW_REQUIRED`. Or skip this and click **Download recorded trajectory.json** instead.
|
| 109 |
+
|
| 110 |
+
> **Narration:** "Two kinds of provenance, kept separate. The evidence graph asks: can this claim trace back through a checked passage to a paper? The trajectory graph asks: what happened — which model ran, who approved what? Process events never count as evidence. And maintainability: a canary re-runs all 398 tests on every change. Here it passed, but it flagged a dependency drift for human review. It can't repair, commit or deploy on its own."
|
| 111 |
+
|
| 112 |
+
[zh-TW] 證據圖(科學溯源)與軌跡圖(流程溯源)分開;Maintenance Canary 偵測到相依套件漂移時,只會提出建議、交由人審查,不會自動修復或部署。
|
| 113 |
+
|
| 114 |
+
### Shot 7 · 2:40–3:00 · Close
|
| 115 |
+
|
| 116 |
+
**Screen:** `docs/assets/flatland-architecture.svg`, then an end card: **FrontierLab** · github.com/Jaycee871/FrontierLab-Meta-Hackathon-2026 · huggingface.co/spaces/Jaycee766/FrontierLab-Meta-Hackathon-2026 · Pack Kwan Low (Jaycee).
|
| 117 |
+
|
| 118 |
+
> **Narration:** "Paper2Agent showed that papers can become agents. My Information Management extension asks how to keep those agents maintainable, traceable and governed, and the Research Operating Interface reports exactly that. FrontierLab is deliberately a Level 1 assisted pipeline, not an autonomous scientist. Meta models reason, critique and propose. Evidence and researchers decide what happens next."
|
| 119 |
+
|
| 120 |
+
**Credit rule:** Paper2Agent (Miao et al., *Nature* 2026) is cited prior work. The three-pillar extension and the Research Operating Interface are the author's own contribution. If Agent Availability comes up, credit the concept to the Paper2Agent authors; FrontierLab adds only the field list.
|
| 121 |
+
|
| 122 |
+
[zh-TW] 結尾:Paper2Agent 是引用的前人研究;可維護性、溯源、研究治理三支柱與研究操作介面是作者本人的延伸貢獻。FrontierLab 刻意定位為 Level 1 輔助流程。
|
| 123 |
+
|
| 124 |
+
---
|
| 125 |
+
|
| 126 |
+
## Phrases to avoid
|
| 127 |
+
|
| 128 |
+
- "live" for anything replayed from `docs/runs/`;
|
| 129 |
+
- "accuracy", "validated", "proven" for trace coverage, test counts or researcher review;
|
| 130 |
+
- "autonomous scientist", "self-repairing", "auto-fixes";
|
| 131 |
+
- any live Muse Spark Open result that is not in a saved receipt;
|
| 132 |
+
- "independent gold" for the researcher's own review.
|
| 133 |
+
|
| 134 |
+
## Post-production checklist
|
| 135 |
+
|
| 136 |
+
- [ ] Total length ≈ 3:00; no dead air longer than 2 s; cut long model waits and label the cuts.
|
| 137 |
+
- [ ] Captions burned in; every number on screen matches the app at recording time.
|
| 138 |
+
- [ ] Banners "RECORDED REAL RUN · REPLAY ONLY" and "Planning is NOT approved" visible when on screen.
|
| 139 |
+
- [ ] No token, account e-mail or billing page visible in any frame.
|
| 140 |
+
- [ ] Export 1080p MP4 (H.264, AAC); upload as public or unlisted; test the link in a private window.
|
| 141 |
+
- [ ] Add the URL to `docs/SUBMISSION_COPY.md` and README (replace the `[TODO Jaycee: …]` placeholders) and to the submission form.
|
docs/HF_INFERENCE_PROVIDERS_MUSE.md
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
-
# Live Meta Muse via Hugging Face Inference Providers (v0.39)
|
| 2 |
|
| 3 |
-
Status: **implemented
|
| 4 |
|
| 5 |
## What it does
|
| 6 |
|
|
@@ -8,12 +8,14 @@ Status: **implemented, no live result recorded yet.** All tests mock the OpenAI
|
|
| 8 |
|
| 9 |
| Position | Model ID sent to the router | Provider | Timeout |
|
| 10 |
| --- | --- | --- | --- |
|
| 11 |
-
| Primary | `meta-models/Muse-Spark-Open-816B-A42B:baseten` | Baseten (pinned) |
|
| 12 |
-
| Fallback | `meta-models/Muse-Glimmer-30B` | chosen by the router (Together, DeepInfra or Featherless AI) |
|
| 13 |
|
| 14 |
- Router: `https://router.huggingface.co/v1` (`chat/completions`), per <https://huggingface.co/docs/inference-providers>. `model:provider` pins a provider; no suffix means the router's default `:fastest` policy.
|
| 15 |
- Fallback happens on any provider error, timeout, HTTP error, empty or truncated output (`finish_reason != "stop"`), unexpected reported model identity, or unparseable / non-object / duplicate-key JSON. If both fail, the step stops with a plain message that points to the recorded replay.
|
| 16 |
-
- Request: system + user message, `max_tokens=
|
|
|
|
|
|
|
| 17 |
- Roles: Evidence Extraction; after researcher claim review, Adversarial Skeptic and Hypothesis; then Method Evaluation suggestions and the approved draft plan through the same client.
|
| 18 |
|
| 19 |
## What it does not change
|
|
|
|
| 1 |
+
# Live Meta Muse via Hugging Face Inference Providers (v0.39, reasoning-budget fix v0.40)
|
| 2 |
|
| 3 |
+
Status: **implemented; one live extraction recorded on 2026-11-01** (`docs/runs/hf_muse_live_extraction_20261101.json`). All tests mock the OpenAI client.
|
| 4 |
|
| 5 |
## What it does
|
| 6 |
|
|
|
|
| 8 |
|
| 9 |
| Position | Model ID sent to the router | Provider | Timeout |
|
| 10 |
| --- | --- | --- | --- |
|
| 11 |
+
| Primary | `meta-models/Muse-Spark-Open-816B-A42B:baseten` | Baseten (pinned) | 90 s |
|
| 12 |
+
| Fallback | `meta-models/Muse-Glimmer-30B` | chosen by the router (Together, DeepInfra or Featherless AI) | 75 s |
|
| 13 |
|
| 14 |
- Router: `https://router.huggingface.co/v1` (`chat/completions`), per <https://huggingface.co/docs/inference-providers>. `model:provider` pins a provider; no suffix means the router's default `:fastest` policy.
|
| 15 |
- Fallback happens on any provider error, timeout, HTTP error, empty or truncated output (`finish_reason != "stop"`), unexpected reported model identity, or unparseable / non-object / duplicate-key JSON. If both fail, the step stops with a plain message that points to the recorded replay.
|
| 16 |
+
- Request: system + user message, `max_tokens=8192`, `temperature=0.2`, `stream=false`, OpenAI SDK `max_retries=0`.
|
| 17 |
+
- Both models are **reasoning models**: hidden reasoning tokens count against `max_tokens` before any answer text. In v0.39 `max_tokens=1536` was used up entirely by reasoning on both models (`finish_reason="length"`, empty answer), so every live run failed. v0.40 raises the budget to 8,192 (a real extraction used 2,975 tokens on Spark and 3,269 on Glimmer) and timeouts to 90 s / 75 s. Reasoning text is never used as an answer, stored or shown; receipts record only `reasoning_tokens`, `reasoning_text_present` and `answer_text_present`.
|
| 18 |
+
- Failed attempts are labelled `ReasoningBudgetExhausted`, `IncompleteOutput`, `InvalidJSONOutput`, `ModelIdentityMismatch` or the SDK error class plus HTTP status. A failed run shows one plain line per attempt and an expanded receipts panel.
|
| 19 |
- Roles: Evidence Extraction; after researcher claim review, Adversarial Skeptic and Hypothesis; then Method Evaluation suggestions and the approved draft plan through the same client.
|
| 20 |
|
| 21 |
## What it does not change
|
docs/RESEARCH_OPERATING_INTERFACE.zh-TW.md
CHANGED
|
@@ -252,17 +252,18 @@ FrontierLab 已經在 `frontierlab/traceability.py` 實作了兩個追蹤指標
|
|
| 252 |
|
| 253 |
### 相關方法
|
| 254 |
|
| 255 |
-
簡報中列出三篇相關方法論文:
|
| 256 |
|
| 257 |
-
- **AlphaGenome**
|
| 258 |
-
- **Scanpy**
|
| 259 |
-
- **TISSUE**
|
| 260 |
|
| 261 |
### 程式碼與部署
|
| 262 |
|
| 263 |
-
- **Paper2Agent** 的程式碼公開在 GitHub
|
| 264 |
-
- AlphaGenome、Scanpy 和 TISSUE 的代理
|
|
|
|
|
|
|
|
|
|
| 265 |
|
| 266 |
-
以上
|
| 267 |
-
|
| 268 |
-
> **待補:** 上面三篇論文的完整書目資料(作者、卷期、頁碼、DOI)和各代理的網址還沒有填入。本文刻意不自行補上,避免寫出錯的引用。Paper2Agent 本身的書目見 [Information Management extension](INFORMATION_MANAGEMENT_EXTENSION.md#references)。
|
|
|
|
| 252 |
|
| 253 |
### 相關方法
|
| 254 |
|
| 255 |
+
簡報中列出三篇相關方法論文。以下書目資料已於 2026-11-01(Asia/Taipei)以 Crossref 的 DOI 紀錄核對:
|
| 256 |
|
| 257 |
+
- **AlphaGenome**:Avsec, Ž., Latysheva, N., Cheng, J. *et al.* Advancing regulatory variant effect prediction with AlphaGenome. *Nature* **649**(8099), 1206–1218 (2026). https://doi.org/10.1038/s41586-025-10014-0
|
| 258 |
+
- **Scanpy**:Wolf, F. A., Angerer, P. & Theis, F. J. SCANPY: large-scale single-cell gene expression data analysis. *Genome Biology* **19**, 15 (2018). https://doi.org/10.1186/s13059-017-1382-0
|
| 259 |
+
- **TISSUE**:Sun, E. D., Ma, R., Navarro Negredo, P., Brunet, A. & Zou, J. TISSUE: uncertainty-calibrated prediction of single-cell spatial transcriptomics improves downstream analyses. *Nature Methods* **21**(3), 444–454 (2024). https://doi.org/10.1038/s41592-024-02184-y
|
| 260 |
|
| 261 |
### 程式碼與部署
|
| 262 |
|
| 263 |
+
- **Paper2Agent** 的程式碼公開在 GitHub:https://github.com/jmiao24/Paper2Agent
|
| 264 |
+
- AlphaGenome、Scanpy 和 TISSUE 的 Paper2Agent 代理(MCP server)部署在 Hugging Face:
|
| 265 |
+
- AlphaGenome:https://huggingface.co/spaces/Paper2Agent/alphagenome_mcp
|
| 266 |
+
- Scanpy:https://huggingface.co/spaces/Paper2Agent/scanpy_mcp
|
| 267 |
+
- TISSUE:https://huggingface.co/spaces/Paper2Agent/tissue_mcp
|
| 268 |
|
| 269 |
+
代理網址與 Paper2Agent 程式碼網址取自 [Information Management extension](INFORMATION_MANAGEMENT_EXTENSION.md#references) 的參考資料;Paper2Agent 本身的完整書目也見該處。以上文獻是 Paper2Agent 作者的案例研究(【文獻】),不是 FrontierLab 的成果。
|
|
|
|
|
|
docs/SUBMISSION_COPY.md
CHANGED
|
@@ -2,13 +2,27 @@
|
|
| 2 |
|
| 3 |
This file contains reusable public-facing copy. Update only claims that have actually changed.
|
| 4 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 5 |
## One-line description
|
| 6 |
|
| 7 |
**FrontierLab is an evidence-grounded, human-gated scientific research pipeline that uses Meta models to turn literature into reviewable hypotheses and plans without hiding the evidence, uncertainty or researcher decisions in between.**
|
| 8 |
|
| 9 |
## Short description
|
| 10 |
|
| 11 |
-
FrontierLab uses Meta Muse Glimmer 30B inside a five-stage scientific workflow: Research → Evidence → Hypothesis → Evaluation → Planning. Model claims must cite exact source passages, critical transitions require researcher review, and evidence/process provenance are exported separately. The system is designed to make scientific AI more auditable rather than more theatrically autonomous.
|
| 12 |
|
| 13 |
## Project description
|
| 14 |
|
|
@@ -18,6 +32,24 @@ A Meta Muse Glimmer 30B backend proposes source-linked claims, receives adversar
|
|
| 18 |
|
| 19 |
The current biomedical demonstration uses published pH-dependent antibody-engineering literature. The researcher selected a structure-based dual-pH design route, while FrontierLab carries unresolved reproducibility checks and orthogonal-validation requirements into a pre-planning contract. Planning remains deliberately human-authorized.
|
| 20 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
## Why this fits Push the Frontier
|
| 22 |
|
| 23 |
FrontierLab applies Meta models to the scientific-research process itself: evidence interpretation, skeptical reasoning, hypothesis generation and experiment-planning preparation.
|
|
@@ -38,7 +70,9 @@ Core implementation:
|
|
| 38 |
|
| 39 |
- Python + Streamlit;
|
| 40 |
- Europe PMC literature retrieval;
|
| 41 |
-
- Meta Muse Glimmer 30B through
|
|
|
|
|
|
|
| 42 |
- strict structured JSON contracts;
|
| 43 |
- deterministic source-ID and verbatim-passage gates;
|
| 44 |
- source-bound human review;
|
|
@@ -51,7 +85,8 @@ Core implementation:
|
|
| 51 |
- separate Trajectory Audit Graph;
|
| 52 |
- recorded gate-ablation tooling;
|
| 53 |
- source-bound 12-case review workspace;
|
| 54 |
-
- deterministic M2 planning-readiness contract
|
|
|
|
| 55 |
|
| 56 |
Optional adapters for Muse Spark Open, Llama 3.1 8B, Muse2-230M, Liquid d1-2B and DetectorScope are intentionally separated from the verified Glimmer path.
|
| 57 |
|
|
@@ -83,11 +118,11 @@ Software verification and scientific evaluation are reported separately.
|
|
| 83 |
|
| 84 |
Current software baseline:
|
| 85 |
|
| 86 |
-
- v0.
|
| 87 |
-
- **
|
| 88 |
-
- matched main Maintenance Canary run
|
| 89 |
- automatic repair, dependency/code edits, commit, merge and deploy remain disabled;
|
| 90 |
-
- v0.
|
| 91 |
- no normal test invokes a model.
|
| 92 |
|
| 93 |
Recorded 12-case Glimmer gate replay:
|
|
@@ -121,7 +156,7 @@ The repository also includes a manual same-model/same-cases two-call prompt-abla
|
|
| 121 |
|
| 122 |
## Deployment state
|
| 123 |
|
| 124 |
-
Latest deployed baseline: **v0.
|
| 125 |
|
| 126 |
## What remains deliberately incomplete
|
| 127 |
|
|
|
|
| 2 |
|
| 3 |
This file contains reusable public-facing copy. Update only claims that have actually changed.
|
| 4 |
|
| 5 |
+
## Submission at a glance
|
| 6 |
+
|
| 7 |
+
| Field | Value |
|
| 8 |
+
| --- | --- |
|
| 9 |
+
| Project | FrontierLab — evidence-grounded, human-gated multi-agent AI for scientific research |
|
| 10 |
+
| Event / track / theme | Meta Global AI Developer Hackathon 2026 · Indie Developer Track · Push the Frontier |
|
| 11 |
+
| Representative / author | Pack Kwan Low (Jaycee) · GitHub [Jaycee871](https://github.com/Jaycee871) · Hugging Face [Jaycee766](https://huggingface.co/Jaycee766) |
|
| 12 |
+
| Team members | [TODO Jaycee: confirm "solo entry (Pack Kwan Low)" or list every team member exactly as registered in the participant portal] |
|
| 13 |
+
| Code repository | https://github.com/Jaycee871/FrontierLab-Meta-Hackathon-2026 (MIT License) |
|
| 14 |
+
| Live demo (HF Space) | https://huggingface.co/spaces/Jaycee766/FrontierLab-Meta-Hackathon-2026 · direct app: https://jaycee766-frontierlab-meta-hackathon-2026.hf.space |
|
| 15 |
+
| Demo video | [TODO Jaycee: public demo-video URL after recording, following [DEMO_VIDEO_SCRIPT.md](DEMO_VIDEO_SCRIPT.md)] |
|
| 16 |
+
| Version | v0.39 on main (`5958770`), deployed to the Space (Space commit `1e162166…`) |
|
| 17 |
+
| Submission deadline | 2026-11-04 21:00 ET = 2026-11-05 10:00 Asia/Taipei (recheck the participant portal) |
|
| 18 |
+
|
| 19 |
## One-line description
|
| 20 |
|
| 21 |
**FrontierLab is an evidence-grounded, human-gated scientific research pipeline that uses Meta models to turn literature into reviewable hypotheses and plans without hiding the evidence, uncertainty or researcher decisions in between.**
|
| 22 |
|
| 23 |
## Short description
|
| 24 |
|
| 25 |
+
FrontierLab uses Meta Muse models — Muse Spark Open (primary) and Muse Glimmer 30B (fallback) through Hugging Face Inference Providers, plus recorded Muse Glimmer 30B runs through NVIDIA NIM — inside a five-stage scientific workflow: Research → Evidence → Hypothesis → Evaluation → Planning. Model claims must cite exact source passages, critical transitions require researcher review, and evidence/process provenance are exported separately. The system is designed to make scientific AI more auditable rather than more theatrically autonomous.
|
| 26 |
|
| 27 |
## Project description
|
| 28 |
|
|
|
|
| 32 |
|
| 33 |
The current biomedical demonstration uses published pH-dependent antibody-engineering literature. The researcher selected a structure-based dual-pH design route, while FrontierLab carries unresolved reproducibility checks and orthogonal-validation requirements into a pre-planning contract. Planning remains deliberately human-authorized.
|
| 34 |
|
| 35 |
+
## Impact
|
| 36 |
+
|
| 37 |
+
**Who it is for.** Biomedical researchers who use language models to read literature, the PIs and reviewers who must sign off on what those models produce, and the research-integrity, journal and funder staff who increasingly have to ask "where did this AI-assisted claim come from, and who approved it?"
|
| 38 |
+
|
| 39 |
+
**The problem it addresses.** A model can cite a real paper and still assert something the paper does not support. In biomedical work that error is not cosmetic: an unsupported claim about binding, safety or mechanism can steer which experiment gets funded or run. At the same time, work such as Paper2Agent (Miao et al., *Nature* 2026, cited prior work) shows that papers and their code/data can become operable agents. Once research artefacts become agents, the open question is no longer only "can the model answer?" but "can anyone reproduce, trace and govern what the agent did?"
|
| 40 |
+
|
| 41 |
+
**What FrontierLab changes, concretely and checkably.**
|
| 42 |
+
|
| 43 |
+
- **Claims are bound to verbatim evidence.** A model claim that does not cite a retrieved source ID with a contiguous 24–500-character quotation from that source is rejected before synthesis. Quote presence is then separately reviewed by a human, because a matching quote is not proof of support.
|
| 44 |
+
- **Provenance becomes a number a reviewer can recompute.** On the recorded, researcher-reviewed real M2 case (two published abstracts; real Meta Muse Glimmer 30B outputs recorded 2026-10-28, Asia/Taipei), the app recomputes from the files in `docs/runs/`: **100% source trace coverage (4/4 artifacts reach a known paper), 50% strict EXTRACTED coverage (2/4 have an all-EXTRACTED path), 9 planning obligations carried into Stage 5, and 0 model calls to replay**. The 50% is not hidden or "fixed": it preserves a real model/researcher provenance disagreement on PMC4314675. These are provenance metrics on a very small sample, not accuracy.
|
| 45 |
+
- **Reproducibility does not require trust, keys or money.** Anyone can open the public Space and replay the real case with no API key and no model call; tampered recorded files or a selection that would authorize planning fail closed.
|
| 46 |
+
- **Governance is enforced, not described.** Claim review, method selection and Stage 5 planning authorization are explicit researcher actions bound to content hashes. In the recorded case the researcher selected M2, yet planning remains **not approved**: readiness carries open data/code/environment checks and obligations such as monomeric-Fab SPR, avidity/stoichiometry handling, search-bias disclosure, replication and falsification.
|
| 47 |
+
- **Maintainability is measured, not assumed.** A scheduled Maintenance Canary re-runs the full regression suite and compares the environment with a human-accepted baseline. On the v0.39 main merge (run 41311197269) all 398 tests passed and the canary reported `HEALTHY`, while the advisory proposal correctly flagged `DRIFT_REVIEW_REQUIRED` because `requirements.txt` changed — detection plus human review, with automatic repair, commit, merge and deploy disabled.
|
| 48 |
+
|
| 49 |
+
**Why it matters beyond this demo.** The five-stage contract, the evidence gate and the provenance/trajectory graphs are domain-agnostic; the pH-dependent antibody-engineering case is the first worked domain. The broader contribution is Pack Kwan Low's Information Management extension — maintainability, provenance and research governance as the three pillars for managing paper-derived agents as research infrastructure — and the Research Operating Interface that reports those pillars as operational status. FrontierLab also proposes a concrete field list for an Agent Availability statement; the Agent Availability concept itself was anticipated by the Paper2Agent authors.
|
| 50 |
+
|
| 51 |
+
**What we do not claim.** No scientific-accuracy score, no independently adjudicated gold standard, no wet-lab or clinical validation, and no recorded result yet from the live Hugging Face Muse path. The impact claimed is narrower and verifiable: model-assisted scientific steps become inspectable, challengeable and reproducible.
|
| 52 |
+
|
| 53 |
## Why this fits Push the Frontier
|
| 54 |
|
| 55 |
FrontierLab applies Meta models to the scientific-research process itself: evidence interpretation, skeptical reasoning, hypothesis generation and experiment-planning preparation.
|
|
|
|
| 70 |
|
| 71 |
- Python + Streamlit;
|
| 72 |
- Europe PMC literature retrieval;
|
| 73 |
+
- Meta Muse Spark Open (primary) and Muse Glimmer 30B (automatic fallback) through the Hugging Face Inference Providers router for the live path (v0.39);
|
| 74 |
+
- Meta Muse Glimmer 30B through NVIDIA NIM for the recorded model-backed research runs;
|
| 75 |
+
- zero-call recorded replay of the real reviewed M2 case;
|
| 76 |
- strict structured JSON contracts;
|
| 77 |
- deterministic source-ID and verbatim-passage gates;
|
| 78 |
- source-bound human review;
|
|
|
|
| 85 |
- separate Trajectory Audit Graph;
|
| 86 |
- recorded gate-ablation tooling;
|
| 87 |
- source-bound 12-case review workspace;
|
| 88 |
+
- deterministic M2 planning-readiness contract;
|
| 89 |
+
- scheduled Maintenance Canary with advisory, human-gated repair proposals.
|
| 90 |
|
| 91 |
Optional adapters for Muse Spark Open, Llama 3.1 8B, Muse2-230M, Liquid d1-2B and DetectorScope are intentionally separated from the verified Glimmer path.
|
| 92 |
|
|
|
|
| 118 |
|
| 119 |
Current software baseline:
|
| 120 |
|
| 121 |
+
- v0.39 on main (merge `5958770`, PR #35);
|
| 122 |
+
- **398 software tests passed** after merge (GitHub Actions run 41311197281);
|
| 123 |
+
- matched main Maintenance Canary run 41311197269 is **HEALTHY** (398 tests passed); its advisory proposal is **DRIFT_REVIEW_REQUIRED / HEALTHY_DRIFT** because the `requirements.txt` hash differs from the accepted Nov 1 baseline (v0.38–v0.39 dependency changes). Accepting a new baseline is a human decision;
|
| 124 |
- automatic repair, dependency/code edits, commit, merge and deploy remain disabled;
|
| 125 |
+
- v0.39 is the deployed HF Space baseline: source `5958770…` → Space commit `1e162166…` (run 41311275823, 398 tests, receipt `HF_SOURCE_COMMITTED_BUILD_PENDING`; a read-only check of the public Space API on 2026-11-01 showed runtime `RUNNING` and a served `deployment.json` with `version: v0.39` and that source SHA). Any later merge needs one more manual deploy for exact Space/main parity;
|
| 126 |
- no normal test invokes a model.
|
| 127 |
|
| 128 |
Recorded 12-case Glimmer gate replay:
|
|
|
|
| 156 |
|
| 157 |
## Deployment state
|
| 158 |
|
| 159 |
+
Latest deployed baseline: **v0.39**. v0.39 main source `5958770f581d9afa4fc8054172f888941c4d46ee` was uploaded by run [41311275823](https://github.com/Jaycee871/FrontierLab-Meta-Hackathon-2026/actions/runs/41311275823) (2026-11-01 21:27 Asia/Taipei) to HF Space commit `1e162166450c254b3136ecd0fc7378a1db72d3ff` after **398 software tests passed**. The receipt reported `HF_SOURCE_COMMITTED_BUILD_PENDING`; a later read-only check of the public Space API (2026-11-01, Asia/Taipei) showed runtime stage `RUNNING` and a served `deployment.json` with `version: v0.39` and that source SHA. At that check the live sidebar still read "Live Meta Muse (via Hugging Face) · unavailable (no Hugging Face token configured)", so the live Muse path must be re-verified on the Space before it is demonstrated. Any later merge (including this docs change) changes the GitHub revision and should be followed by one final deploy if exact Space/main source parity is required.
|
| 160 |
|
| 161 |
## What remains deliberately incomplete
|
| 162 |
|
docs/runs/hf_muse_live_extraction_20261101.json
ADDED
|
@@ -0,0 +1,171 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"record_type": "live_hf_inference_providers_muse_extraction",
|
| 3 |
+
"recorded_date_taipei": "2026-11-01",
|
| 4 |
+
"recorded_time_taipei": "23:18",
|
| 5 |
+
"codebase_version": "0.40",
|
| 6 |
+
"code_state": "branch fix/live-muse-reasoning (base main 5958770) working tree with the v0.40 reasoning-budget fix, run locally (not on the Space)",
|
| 7 |
+
"entry_point": "frontierlab.pipeline.run_research(mode='live', max_papers=1) with frontierlab.hf_providers.HFMuseResearchClient(authorized=True, session_budget=SessionBudget(2)); default app question and Europe PMC query",
|
| 8 |
+
"source": "Provenance export of the run (frontierlab.reporting.export_json); no credentials, no reasoning text and no full abstract text included",
|
| 9 |
+
"receipt": {
|
| 10 |
+
"schema_version": 1,
|
| 11 |
+
"question": "How might pH-dependent protein binder affinity be evaluated reproducibly?",
|
| 12 |
+
"query": "(pH dependent antibody binding) AND (histidine OR engineering)",
|
| 13 |
+
"paper": {
|
| 14 |
+
"paper_id": "MED:42959777",
|
| 15 |
+
"title": "Targeting BIGH3 in selected solid tumors through a pH-Dependent antibody engineering strategy.",
|
| 16 |
+
"year": "2026",
|
| 17 |
+
"url": "https://doi.org/10.1016/j.bbrep.2026.102769"
|
| 18 |
+
},
|
| 19 |
+
"backend": "hf-inference-providers",
|
| 20 |
+
"model": "meta-models/Muse-Spark-Open-816B-A42B",
|
| 21 |
+
"inference": {
|
| 22 |
+
"provider": "Hugging Face Inference Providers",
|
| 23 |
+
"backend": "hf-inference-providers",
|
| 24 |
+
"router_url": "https://router.huggingface.co/v1",
|
| 25 |
+
"primary_model": "meta-models/Muse-Spark-Open-816B-A42B",
|
| 26 |
+
"primary_provider_requested": "baseten",
|
| 27 |
+
"fallback_model": "meta-models/Muse-Glimmer-30B",
|
| 28 |
+
"fallback_provider_requested": "auto",
|
| 29 |
+
"session_call_cap": 2,
|
| 30 |
+
"daily_call_cap": 60,
|
| 31 |
+
"max_tokens": 8192,
|
| 32 |
+
"temperature": 0.2,
|
| 33 |
+
"max_abstract_chars": 3000,
|
| 34 |
+
"attempted_calls": 1,
|
| 35 |
+
"receipts": [
|
| 36 |
+
{
|
| 37 |
+
"role": "Evidence Extraction Agent",
|
| 38 |
+
"chain_position": "primary",
|
| 39 |
+
"requested_model": "meta-models/Muse-Spark-Open-816B-A42B",
|
| 40 |
+
"router_model_string": "meta-models/Muse-Spark-Open-816B-A42B:baseten",
|
| 41 |
+
"provider_requested": "baseten",
|
| 42 |
+
"started_at_utc": "2026-11-01T15:18:50.170685+00:00",
|
| 43 |
+
"input_sha256": "11286e8324237d5fe1c4726238819f596b039b42ef8eec2137969ce521ca6c65",
|
| 44 |
+
"input_abstract_truncated": false,
|
| 45 |
+
"max_tokens": 8192,
|
| 46 |
+
"temperature": 0.2,
|
| 47 |
+
"timeout_seconds": 90.0,
|
| 48 |
+
"answer_text_present": true,
|
| 49 |
+
"reasoning_text_present": true,
|
| 50 |
+
"provider_reported": "baseten",
|
| 51 |
+
"reported_model": "meta-models/muse-spark-open",
|
| 52 |
+
"finish_reason": "stop",
|
| 53 |
+
"usage": {
|
| 54 |
+
"prompt_tokens": 810,
|
| 55 |
+
"completion_tokens": 3477,
|
| 56 |
+
"total_tokens": 4287,
|
| 57 |
+
"reasoning_tokens": 3182
|
| 58 |
+
},
|
| 59 |
+
"output_sha256": "9035cc063607546b656895727457365cd1218631d928d84482cec45aa50be308",
|
| 60 |
+
"model_identity": "MATCHED",
|
| 61 |
+
"finished_at_utc": "2026-11-01T15:18:58.159638+00:00",
|
| 62 |
+
"elapsed_seconds": 7.989,
|
| 63 |
+
"status": "JSON_OBJECT_RETURNED"
|
| 64 |
+
}
|
| 65 |
+
],
|
| 66 |
+
"billed_to": "Hugging Face account of the project owner",
|
| 67 |
+
"direct_meta_model_api_used": false,
|
| 68 |
+
"scientific_outputs_independently_verified": false
|
| 69 |
+
},
|
| 70 |
+
"status": "JSON_OBJECT_RETURNED",
|
| 71 |
+
"decision": "INSUFFICIENT_EVIDENCE",
|
| 72 |
+
"audit": {
|
| 73 |
+
"verified_claims": [
|
| 74 |
+
{
|
| 75 |
+
"paper_id": "MED:42959777",
|
| 76 |
+
"claim_key": "kd_ec50_acidic_vs_neutral_ph_affinity_evaluation",
|
| 77 |
+
"stance": "supports",
|
| 78 |
+
"statement": "Abstract reports BIGH3 antibody binding and affinity was evaluated using KD or EC50 measurements under TME-mimicking acidic conditions versus physiological pH, suggesting comparative KD/EC50 assessment as a method to evaluate pH-dependent binder affinity for this target.",
|
| 79 |
+
"limitation": "Proof-of-concept study limited to BIGH3; abstract notes direct in vivo PK, biodistribution and tumor accumulation studies remain to be performed and provides no quantitative reproducibility metrics.",
|
| 80 |
+
"evidence_quote": "without detectable loss of antibody binding and affinity (Equilibrium dissociation constant (KD) or half-maximal effective concentration (EC50))",
|
| 81 |
+
"passage_matched": true
|
| 82 |
+
},
|
| 83 |
+
{
|
| 84 |
+
"paper_id": "MED:42959777",
|
| 85 |
+
"claim_key": "biophysical_in_vitro_ph_dependent_engagement",
|
| 86 |
+
"stance": "supports",
|
| 87 |
+
"statement": "Abstract indicates pH-dependent target engagement was assessed via biophysical analyses for stability/manufacturability combined with in vitro studies demonstrating differential binding under acidic TME-mimicking versus neutral pH conditions.",
|
| 88 |
+
"limitation": "Findings described as proof-of-concept with in vitro data and single colorectal cancer mouse model; abstract explicitly states in vivo PK/biodistribution validation is still needed.",
|
| 89 |
+
"evidence_quote": "Biophysical analyses confirmed preserved stability and manufacturability profiles, while in vitro studies demonstrated pH-dependent target engagement",
|
| 90 |
+
"passage_matched": true
|
| 91 |
+
}
|
| 92 |
+
],
|
| 93 |
+
"rejected_claims": [],
|
| 94 |
+
"warnings": [
|
| 95 |
+
"Passage grounding checked against cited abstracts; semantic entailment is NOT verified. Researcher review of methods and full text is required.",
|
| 96 |
+
"Fewer than two distinct sources with validated source IDs; hypothesis generation stopped."
|
| 97 |
+
]
|
| 98 |
+
},
|
| 99 |
+
"log": [
|
| 100 |
+
{
|
| 101 |
+
"time_utc": "2026-11-01T15:18:50.170410+00:00",
|
| 102 |
+
"role": "Retriever",
|
| 103 |
+
"status": "COMPLETE",
|
| 104 |
+
"note": "Retrieved 1 indexed publications with abstracts"
|
| 105 |
+
},
|
| 106 |
+
{
|
| 107 |
+
"time_utc": "2026-11-01T15:18:58.159990+00:00",
|
| 108 |
+
"role": "Extraction Agent",
|
| 109 |
+
"status": "COMPLETE",
|
| 110 |
+
"note": "2 proposed; 2 abstract-quoted; 0 rejected"
|
| 111 |
+
},
|
| 112 |
+
{
|
| 113 |
+
"time_utc": "2026-11-01T15:18:58.160000+00:00",
|
| 114 |
+
"role": "Audit Gate",
|
| 115 |
+
"status": "BLOCKED",
|
| 116 |
+
"note": "Insufficient distinct sourced claims for synthesis"
|
| 117 |
+
}
|
| 118 |
+
]
|
| 119 |
+
},
|
| 120 |
+
"diagnosis": {
|
| 121 |
+
"symptom": "v0.39 live mode on the HF Space: Run consumed 2 calls (primary + fallback) and showed no result.",
|
| 122 |
+
"root_cause": "Both Muse models are reasoning models. Reasoning tokens count against max_tokens; with max_tokens=1536 both spent the whole budget on reasoning (finish_reason=length, answer content null/empty), which the client correctly rejected as incomplete output, so both attempts failed.",
|
| 123 |
+
"evidence_calls": [
|
| 124 |
+
{
|
| 125 |
+
"model": "meta-models/Muse-Spark-Open-816B-A42B:baseten",
|
| 126 |
+
"max_tokens": 1536,
|
| 127 |
+
"http": 200,
|
| 128 |
+
"provider_reported": "baseten",
|
| 129 |
+
"finish_reason": "length",
|
| 130 |
+
"answer_content": "null",
|
| 131 |
+
"reasoning_tokens": 1533,
|
| 132 |
+
"completion_tokens": 1536,
|
| 133 |
+
"latency_seconds": 3.84
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"model": "meta-models/Muse-Glimmer-30B",
|
| 137 |
+
"max_tokens": 1536,
|
| 138 |
+
"http": 200,
|
| 139 |
+
"provider_reported": "deepinfra",
|
| 140 |
+
"finish_reason": "length",
|
| 141 |
+
"answer_content": "empty string",
|
| 142 |
+
"completion_tokens": 1536,
|
| 143 |
+
"latency_seconds": 7.5
|
| 144 |
+
},
|
| 145 |
+
{
|
| 146 |
+
"model": "meta-models/Muse-Spark-Open-816B-A42B:baseten",
|
| 147 |
+
"max_tokens": 8192,
|
| 148 |
+
"http": 200,
|
| 149 |
+
"provider_reported": "baseten",
|
| 150 |
+
"finish_reason": "stop",
|
| 151 |
+
"answer": "valid JSON, 2 claims",
|
| 152 |
+
"reasoning_tokens": 2600,
|
| 153 |
+
"completion_tokens": 2975,
|
| 154 |
+
"latency_seconds": 7.39
|
| 155 |
+
},
|
| 156 |
+
{
|
| 157 |
+
"model": "meta-models/Muse-Glimmer-30B",
|
| 158 |
+
"max_tokens": 8192,
|
| 159 |
+
"http": 200,
|
| 160 |
+
"provider_reported": "deepinfra",
|
| 161 |
+
"finish_reason": "stop",
|
| 162 |
+
"answer": "valid JSON, 1 claim",
|
| 163 |
+
"completion_tokens": 3269,
|
| 164 |
+
"latency_seconds": 17.1
|
| 165 |
+
}
|
| 166 |
+
],
|
| 167 |
+
"input_sha256_all_evidence_calls": "11286e8324237d5fe1c4726238819f596b039b42ef8eec2137969ce521ca6c65",
|
| 168 |
+
"note": "All four diagnostic calls used the same Evidence Extraction prompt the app builds (same input_sha256 as this receipt) with a 180 s observation timeout."
|
| 169 |
+
},
|
| 170 |
+
"scope": "One real Europe PMC abstract, one live Evidence Extraction call. Both quotes passed the verbatim evidence gate; semantic entailment is NOT verified and researcher review is pending. One source is below the two-source threshold, so no skeptic, hypothesis or plan was generated. Not a benchmark or scientific validation."
|
| 171 |
+
}
|
frontierlab/hf_providers.py
CHANGED
|
@@ -14,7 +14,12 @@ Safety and cost guards:
|
|
| 14 |
- a hard per-session call cap (default 6) and a per-process daily cap
|
| 15 |
(``FRONTIERLAB_HF_DAILY_CALL_CAP``, default 60) count every HTTP attempt,
|
| 16 |
including fallbacks; a capped request is never sent;
|
| 17 |
-
- strict timeouts, ``max_retries=0``,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 18 |
- source abstracts are truncated to a bounded prefix before sending (a quote
|
| 19 |
from the prefix is still a verbatim excerpt of the full abstract, and the
|
| 20 |
unchanged fail-closed quote gate in ``audit.validate_claims`` checks it
|
|
@@ -48,8 +53,8 @@ FALLBACK_PROVIDER = None # router auto-selection among live Glimmer providers
|
|
| 48 |
|
| 49 |
# (role in chain, HF model id, pinned provider or None, timeout seconds, family key)
|
| 50 |
ATTEMPTS = (
|
| 51 |
-
("primary", PRIMARY_MODEL, PRIMARY_PROVIDER,
|
| 52 |
-
("fallback", FALLBACK_MODEL, FALLBACK_PROVIDER,
|
| 53 |
)
|
| 54 |
|
| 55 |
TOKEN_ENV_VARS = ("HF_TOKEN", "HF_INFERENCE_TOKEN", "HUGGINGFACEHUB_API_TOKEN")
|
|
@@ -61,7 +66,9 @@ MAX_LIVE_ABSTRACTS = 2
|
|
| 61 |
MAX_ABSTRACT_CHARS = 3000
|
| 62 |
MAX_QUESTION_CHARS = 1000
|
| 63 |
MAX_PROMPT_CHARS = 16000
|
| 64 |
-
|
|
|
|
|
|
|
| 65 |
TEMPERATURE = 0.2
|
| 66 |
|
| 67 |
ROLES = {
|
|
@@ -223,14 +230,52 @@ def parse_json_object(text: str) -> dict:
|
|
| 223 |
return answer
|
| 224 |
|
| 225 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 226 |
def _safe_usage(usage: Any) -> dict:
|
|
|
|
| 227 |
if usage is None:
|
| 228 |
return {}
|
| 229 |
if not isinstance(usage, dict):
|
|
|
|
| 230 |
usage = {k: getattr(usage, k, None)
|
| 231 |
for k in ("prompt_tokens", "completion_tokens", "total_tokens")}
|
| 232 |
-
|
| 233 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 234 |
|
| 235 |
|
| 236 |
def _bounded_payload(payload: dict) -> tuple[dict, bool]:
|
|
@@ -368,7 +413,10 @@ class HFMuseResearchClient:
|
|
| 368 |
content = choice.message.content
|
| 369 |
finish_reason = choice.finish_reason
|
| 370 |
reported = getattr(completion, "model", None)
|
|
|
|
| 371 |
receipt.update(
|
|
|
|
|
|
|
| 372 |
provider_reported=served_by if isinstance(served_by, str) and served_by else "NOT_REPORTED",
|
| 373 |
reported_model=reported if isinstance(reported, str) and reported else "NOT_REPORTED",
|
| 374 |
finish_reason=finish_reason if finish_reason in ("stop", "length") else "other",
|
|
@@ -379,10 +427,15 @@ class HFMuseResearchClient:
|
|
| 379 |
identity = _identity(reported, family)
|
| 380 |
receipt["model_identity"] = identity
|
| 381 |
if identity == "MISMATCH":
|
| 382 |
-
raise
|
| 383 |
-
if finish_reason
|
| 384 |
-
raise
|
| 385 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 386 |
except Exception as exc: # fail closed; never surface message text
|
| 387 |
receipt["status"] = "FAILED_FALLBACK_NEXT" if has_next else "FAILED"
|
| 388 |
receipt["error"] = _error_label(exc)
|
|
@@ -393,3 +446,20 @@ class HFMuseResearchClient:
|
|
| 393 |
receipt["status"] = "JSON_OBJECT_RETURNED"
|
| 394 |
self.model = model
|
| 395 |
return answer
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
- a hard per-session call cap (default 6) and a per-process daily cap
|
| 15 |
(``FRONTIERLAB_HF_DAILY_CALL_CAP``, default 60) count every HTTP attempt,
|
| 16 |
including fallbacks; a capped request is never sent;
|
| 17 |
+
- strict timeouts, ``max_retries=0``, a bounded ``max_tokens``, low temperature;
|
| 18 |
+
- both Muse models are reasoning models: hidden reasoning tokens are counted
|
| 19 |
+
against ``max_tokens`` before any answer text is produced, so the budget must
|
| 20 |
+
leave room for reasoning *and* the JSON answer (v0.40; 1536 left the answer
|
| 21 |
+
empty with ``finish_reason="length"`` on both models). Reasoning text is never
|
| 22 |
+
used as an answer, stored or shown; only its token count is receipted;
|
| 23 |
- source abstracts are truncated to a bounded prefix before sending (a quote
|
| 24 |
from the prefix is still a verbatim excerpt of the full abstract, and the
|
| 25 |
unchanged fail-closed quote gate in ``audit.validate_claims`` checks it
|
|
|
|
| 53 |
|
| 54 |
# (role in chain, HF model id, pinned provider or None, timeout seconds, family key)
|
| 55 |
ATTEMPTS = (
|
| 56 |
+
("primary", PRIMARY_MODEL, PRIMARY_PROVIDER, 90.0, "muse-spark"),
|
| 57 |
+
("fallback", FALLBACK_MODEL, FALLBACK_PROVIDER, 75.0, "muse-glimmer"),
|
| 58 |
)
|
| 59 |
|
| 60 |
TOKEN_ENV_VARS = ("HF_TOKEN", "HF_INFERENCE_TOKEN", "HUGGINGFACEHUB_API_TOKEN")
|
|
|
|
| 66 |
MAX_ABSTRACT_CHARS = 3000
|
| 67 |
MAX_QUESTION_CHARS = 1000
|
| 68 |
MAX_PROMPT_CHARS = 16000
|
| 69 |
+
# Reasoning + answer budget. Measured 2026-11-01 on one real extraction prompt:
|
| 70 |
+
# Muse Spark Open used 2,975 completion tokens (2,600 reasoning), Muse Glimmer 3,269.
|
| 71 |
+
MAX_TOKENS = 8192
|
| 72 |
TEMPERATURE = 0.2
|
| 73 |
|
| 74 |
ROLES = {
|
|
|
|
| 230 |
return answer
|
| 231 |
|
| 232 |
|
| 233 |
+
def _count(value: Any) -> bool:
|
| 234 |
+
return isinstance(value, int) and not isinstance(value, bool) and value >= 0
|
| 235 |
+
|
| 236 |
+
|
| 237 |
def _safe_usage(usage: Any) -> dict:
|
| 238 |
+
"""Token counts only (prompt/completion/total, plus reasoning when reported)."""
|
| 239 |
if usage is None:
|
| 240 |
return {}
|
| 241 |
if not isinstance(usage, dict):
|
| 242 |
+
details = getattr(usage, "completion_tokens_details", None)
|
| 243 |
usage = {k: getattr(usage, k, None)
|
| 244 |
for k in ("prompt_tokens", "completion_tokens", "total_tokens")}
|
| 245 |
+
usage["completion_tokens_details"] = (
|
| 246 |
+
{"reasoning_tokens": getattr(details, "reasoning_tokens", None)}
|
| 247 |
+
if details is not None and not isinstance(details, dict) else details)
|
| 248 |
+
out = {k: usage[k] for k in ("prompt_tokens", "completion_tokens", "total_tokens")
|
| 249 |
+
if _count(usage.get(k))}
|
| 250 |
+
details = usage.get("completion_tokens_details")
|
| 251 |
+
if isinstance(details, dict) and _count(details.get("reasoning_tokens")):
|
| 252 |
+
out["reasoning_tokens"] = details["reasoning_tokens"]
|
| 253 |
+
return out
|
| 254 |
+
|
| 255 |
+
|
| 256 |
+
def _has_reasoning(message: Any) -> bool:
|
| 257 |
+
"""True when the provider returned separate reasoning text (never read further)."""
|
| 258 |
+
extra = getattr(message, "model_extra", None)
|
| 259 |
+
fields = [getattr(message, "reasoning_content", None), getattr(message, "reasoning", None)]
|
| 260 |
+
if isinstance(extra, dict):
|
| 261 |
+
fields += [extra.get("reasoning_content"), extra.get("reasoning")]
|
| 262 |
+
return any(isinstance(f, str) and f.strip() for f in fields)
|
| 263 |
+
|
| 264 |
+
|
| 265 |
+
class ReasoningBudgetExhausted(ValueError):
|
| 266 |
+
"""The reasoning model spent max_tokens before producing any answer text."""
|
| 267 |
+
|
| 268 |
+
|
| 269 |
+
class IncompleteOutput(ValueError):
|
| 270 |
+
"""Truncated, empty or non-final answer text."""
|
| 271 |
+
|
| 272 |
+
|
| 273 |
+
class InvalidJSONOutput(ValueError):
|
| 274 |
+
"""Answer text was not one strict JSON object."""
|
| 275 |
+
|
| 276 |
+
|
| 277 |
+
class ModelIdentityMismatch(ValueError):
|
| 278 |
+
"""The provider reported a model outside the requested family."""
|
| 279 |
|
| 280 |
|
| 281 |
def _bounded_payload(payload: dict) -> tuple[dict, bool]:
|
|
|
|
| 413 |
content = choice.message.content
|
| 414 |
finish_reason = choice.finish_reason
|
| 415 |
reported = getattr(completion, "model", None)
|
| 416 |
+
has_content = isinstance(content, str) and bool(content.strip())
|
| 417 |
receipt.update(
|
| 418 |
+
answer_text_present=has_content,
|
| 419 |
+
reasoning_text_present=_has_reasoning(choice.message),
|
| 420 |
provider_reported=served_by if isinstance(served_by, str) and served_by else "NOT_REPORTED",
|
| 421 |
reported_model=reported if isinstance(reported, str) and reported else "NOT_REPORTED",
|
| 422 |
finish_reason=finish_reason if finish_reason in ("stop", "length") else "other",
|
|
|
|
| 427 |
identity = _identity(reported, family)
|
| 428 |
receipt["model_identity"] = identity
|
| 429 |
if identity == "MISMATCH":
|
| 430 |
+
raise ModelIdentityMismatch("Provider reported an unexpected model identity")
|
| 431 |
+
if finish_reason == "length" and not has_content:
|
| 432 |
+
raise ReasoningBudgetExhausted("Reasoning used the whole token budget; no answer")
|
| 433 |
+
if finish_reason != "stop" or not has_content:
|
| 434 |
+
raise IncompleteOutput("Model output incomplete")
|
| 435 |
+
try:
|
| 436 |
+
answer = parse_json_object(content)
|
| 437 |
+
except ValueError:
|
| 438 |
+
raise InvalidJSONOutput("Model response was not one JSON object") from None
|
| 439 |
except Exception as exc: # fail closed; never surface message text
|
| 440 |
receipt["status"] = "FAILED_FALLBACK_NEXT" if has_next else "FAILED"
|
| 441 |
receipt["error"] = _error_label(exc)
|
|
|
|
| 446 |
receipt["status"] = "JSON_OBJECT_RETURNED"
|
| 447 |
self.model = model
|
| 448 |
return answer
|
| 449 |
+
|
| 450 |
+
|
| 451 |
+
def failure_summary(provenance: dict | None) -> list[str]:
|
| 452 |
+
"""One plain line per failed attempt (model, provider, error class, latency). No message text."""
|
| 453 |
+
lines = []
|
| 454 |
+
for r in (provenance or {}).get("receipts", []):
|
| 455 |
+
if not isinstance(r, dict) or not str(r.get("status", "")).startswith("FAILED"):
|
| 456 |
+
continue
|
| 457 |
+
provider = r.get("provider_reported") or r.get("provider_requested") or "auto"
|
| 458 |
+
line = (f"{r.get('chain_position', '?')}: {r.get('requested_model', '?')} via {provider} · "
|
| 459 |
+
f"{r.get('error', 'unknown error')}")
|
| 460 |
+
if r.get("finish_reason"):
|
| 461 |
+
line += f" · finish_reason={r['finish_reason']}"
|
| 462 |
+
if isinstance(r.get("elapsed_seconds"), (int, float)):
|
| 463 |
+
line += f" · {r['elapsed_seconds']:.1f} s"
|
| 464 |
+
lines.append(line)
|
| 465 |
+
return lines
|
frontierlab/version.py
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
"""Single codebase version used by the UI and tests."""
|
| 2 |
|
| 3 |
-
__version__ = "0.
|
|
|
|
| 1 |
"""Single codebase version used by the UI and tests."""
|
| 2 |
|
| 3 |
+
__version__ = "0.40"
|