Publish 9 November leaderboard snapshot with GPT-6.1 Sol high effort

Add leaderboard-results-20261109.json; show reasoning effort in the dataset-card table; update GPT-6.1 Sol task scores to the high-effort 107-task run and keep Hypothesis Discovery unchanged.

LEADERBOARD.md

@@ -1,6 +1,6 @@
 # Native Hub leaderboard definitions
 
-This specification describes the website results synchronized on **8 November 2026**. It records published
+This specification describes the website results synchronized on **9 November 2026**. It records published
 model-and-harness results, not a reaggregation of later attempts. The complete table is in the dataset card and on
 the [website leaderboard](https://sciences-last-exam.org/leaderboard).
 
@@ -37,13 +37,14 @@ progress rubric are diagnostic outputs, not replacements for the official leader
 
 ## Frozen values and provenance
 
-`leaderboard-results-20261108.json` contains the current published rows, all scored assessment metrics, the source
+`leaderboard-results-20261109.json` contains the current published rows, all scored assessment metrics, the source
 website commit and the Hub model mappings. Earlier dated snapshots are retained unchanged for provenance; they
 record the scope and aggregation in use at the time. Missing infrastructure outcomes are not converted to zero.
 
 A published row can be partial: Gemini 4 Argon reports the four task-based assessments, and its `composite-score`
 stays null until Hypothesis Discovery is available. Cost and time are still reported because all four contributing
-assessments are present.
+assessments are present. The 9 November snapshot updates GPT-6.1 Sol's four task-based assessments to the
+high-effort run; its Hypothesis Discovery row is unchanged, and the composite uses that value.
 
 Evaluations use model API endpoints with task-specific agent runtimes. Native result notes identify the harness and
 reasoning setting. These are API endpoint measurements, not measurements of locally served or quantized Hub

README.md

@@ -81,20 +81,22 @@ Watch recorded agent runs on **42 out of 113 tasks** in [Replays](https://scienc
 ## Published results
 
 The table below matches the [website leaderboard](https://sciences-last-exam.org/leaderboard), including its
-End-to-end research component. The snapshot was synchronized on **8 November 2026** from the website's
-8 November results. Each entry evaluates a model with its agent harness.
-
-| Model | Harness | Composite (0–100) | Cost / task (USD) | Agent time / task (s) |
-| --- | --- | ---: | ---: | ---: |
-| Claude Opus 5 | Claude Code | 37.50 | $19.70 | 1,051 |
-| GPT-6.1 Sol | Codex | 36.15 | $2.85 | 692 |
-| Kimi K3.1 Pro | Claude Code | 28.92 | $2.89 | 1,273 |
-| DeepSeek V4.1 Flash | Claude Code | 26.83 | $0.42 | 1,266 |
-| MiMo-V2.6-Pro | Claude Code | 26.40 | $0.56 | 2,341 |
-| Qwen3.8 Max | Claude Code | 23.30 | $6.43 | 2,136 |
-| GLM-5.3 Flash | Claude Code | 21.33 | $0.38 | 1,726 |
-| MiniMax-M3 | Claude Code | 16.82 | $1.15 | 1,727 |
-| Gemini 4 Argon | Codex | — | $4.26 | 1,833 |
+End-to-end research component. The snapshot was synchronized on **9 November 2026** from the website's
+9 November results. Each entry evaluates a model with its agent harness and records its reasoning effort.
+
+| Model | Harness | Effort | Composite (0–100) | Cost / task (USD) | Agent time / task (s) |
+| --- | --- | --- | ---: | ---: | ---: |
+| GPT-6.1 Sol | Codex | High | 44.66 | $3.07 | 811 |
+| Claude Opus 5 | Claude Code | High | 37.50 | $19.70 | 1,051 |
+| Kimi K3.1 Pro | Claude Code | — | 28.92 | $2.89 | 1,273 |
+| DeepSeek V4.1 Flash | Claude Code | High | 26.83 | $0.42 | 1,266 |
+| MiMo-V2.6-Pro | Claude Code | — | 26.40 | $0.56 | 2,341 |
+| Qwen3.8 Max | Claude Code | — | 23.30 | $6.43 | 2,136 |
+| GLM-5.3 Flash | Claude Code | High | 21.33 | $0.38 | 1,726 |
+| MiniMax-M3 | Claude Code | High | 16.82 | $1.15 | 1,727 |
+| Gemini 4 Argon | Codex | Default | — | $4.26 | 1,833 |
+
+Effort is the reasoning setting recorded with the published run; **—** means the setting was not specified.
 
 **Claude Opus 5.5 and GPT-6 Astra have not yet been evaluated. Their biological safeguards can restrict biology research workflows. We are applying for research access and will add results once access is approved and evaluation is complete.** Official explanations: [Anthropic](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-opus-5-5#safeguard-refusals) · [OpenAI](https://deploymentsafety.openai.com/gpt-6-astra/biology---trusted-access-for-biology-research).
 
@@ -128,7 +130,7 @@ Infrastructure failures remain unscored. The runner's mean task score is a diagn
 
 See [scoring rules](https://sciences-last-exam.org/docs-scoring) and the [task specification](https://github.com/Gen-Verse/Science-Last-Exam/blob/main/docs/TASKS.md).
 The [native leaderboard specification](LEADERBOARD.md) defines Hub task IDs and metric units. The
-[current result snapshot](leaderboard-results-20261108.json) contains the published model rows and all scored
+[current result snapshot](leaderboard-results-20261109.json) contains the published model rows and all scored
 assessment metrics. Earlier dated result snapshots remain available for provenance; they use the scope and
 aggregation recorded in those files. Native Benchmark activation is managed by Hugging Face.
 
@@ -196,6 +198,7 @@ Use matching source-code, input-data and reference-data revisions. Record the da
 
 - **2026-11-05 — website alignment:** descriptions use six scientific capabilities and one end-to-end research assessment. The composite includes End-to-end research at weight 1/6, and the published table matches the website. Task instructions and inputs are unchanged.
 - **2026-11-08 — Gemini 4 Argon:** added Protocol Planning, Scientific Computing, Scientific Software Operation and Experiment Observation results from the 8 November evaluation. Hypothesis Discovery is pending, so no composite score is reported. The task cards and other models are unchanged.
+- **2026-11-09 — GPT-6.1 Sol high effort:** updated the four task-based assessments to the high-effort 107-task run (composite 44.66, cost $3.07/task, 811 s/task). Hypothesis Discovery remains 4.92 from the medium-effort measurement. The published table now shows reasoning effort for every model.
 
 ## License and intended use
 

leaderboard-results-20261109.json

@@ -0,0 +1,276 @@
+{
+  "snapshot": "2026-11-09",
+  "scope": {
+    "protocol-planning": 40,
+    "experiment-observation": 30,
+    "scientific-computing": 23,
+    "scientific-software-operation": 14,
+    "hypothesis-discovery": {
+      "harness": "DiscoveryRSI",
+      "research_databases": [
+        "MIMIC-IV",
+        "eICU",
+        "HCC clinical cohort",
+        "UK Biobank",
+        "GEO",
+        "ArrayExpress"
+      ],
+      "unit": "hypothesis versions"
+    },
+    "end-to-end-science": 1
+  },
+  "excluded": [
+    "wet-lab-execution"
+  ],
+  "source": "https://sciences-last-exam.org/leaderboard",
+  "source_website_commit": "6927a475f8f15e1c9eaf4b5dd432bac19d598c63",
+  "precision": "Published chart-label precision; not a rescore of later attempts.",
+  "models": [
+    {
+      "id": "gpt",
+      "name": "GPT-6.1 Sol",
+      "harness": "Codex",
+      "reasoning_effort": "high",
+      "hub_model": null,
+      "metrics": {
+        "composite-score": 44.66,
+        "hypothesis-discovery-inspire-rate": 4.92,
+        "protocol-planning-mean-score": 57.3,
+        "scientific-computing-full-credit-rate": 73.33,
+        "scientific-software-operation-full-credit-rate": 85.71,
+        "experiment-observation-pass-rate": 46.67,
+        "end-to-end-research-score": 0.0
+      },
+      "hypothesis_discovery": {
+        "harness": "DiscoveryRSI",
+        "positives": 6,
+        "hypothesis_versions": 122
+      },
+      "cost_per_task_usd": 3.071625,
+      "agent_time_per_task_s": 811.05
+    },
+    {
+      "id": "claude",
+      "name": "Claude Opus 5",
+      "harness": "Claude Code",
+      "reasoning_effort": "high",
+      "hub_model": null,
+      "metrics": {
+        "composite-score": 37.5,
+        "hypothesis-discovery-inspire-rate": 7.33,
+        "protocol-planning-mean-score": 55.0,
+        "scientific-computing-full-credit-rate": 57.0,
+        "scientific-software-operation-full-credit-rate": 79.0,
+        "experiment-observation-pass-rate": 26.67,
+        "end-to-end-research-score": 0.0
+      },
+      "hypothesis_discovery": {
+        "harness": "DiscoveryRSI",
+        "positives": 25,
+        "hypothesis_versions": 341
+      },
+      "cost_per_task_usd": 19.7025,
+      "agent_time_per_task_s": 1051.25
+    },
+    {
+      "id": "gemini",
+      "name": "Gemini 4 Argon",
+      "harness": "Codex",
+      "reasoning_effort": "default",
+      "hub_model": null,
+      "metrics": {
+        "composite-score": null,
+        "hypothesis-discovery-inspire-rate": null,
+        "protocol-planning-mean-score": 46.1,
+        "scientific-computing-full-credit-rate": 39.13,
+        "scientific-software-operation-full-credit-rate": 57.14,
+        "experiment-observation-pass-rate": 50.0,
+        "end-to-end-research-score": 0.0
+      },
+      "hypothesis_discovery": null,
+      "cost_per_task_usd": 4.262225,
+      "agent_time_per_task_s": 1832.875
+    },
+    {
+      "id": "deepseek",
+      "name": "DeepSeek V4.1 Flash",
+      "harness": "Claude Code",
+      "reasoning_effort": "high",
+      "hub_model": "deepseek-ai/DeepSeek-V4.1-Flash-1017",
+      "metrics": {
+        "composite-score": 26.83,
+        "hypothesis-discovery-inspire-rate": 21.28,
+        "protocol-planning-mean-score": 55.0,
+        "scientific-computing-full-credit-rate": 39.0,
+        "scientific-software-operation-full-credit-rate": 29.0,
+        "experiment-observation-pass-rate": 16.67,
+        "end-to-end-research-score": 0.0
+      },
+      "hypothesis_discovery": {
+        "harness": "DiscoveryRSI",
+        "positives": 116,
+        "hypothesis_versions": 545
+      },
+      "cost_per_task_usd": 0.422,
+      "agent_time_per_task_s": 1265.75
+    },
+    {
+      "id": "qwen",
+      "name": "Qwen3.8 Max",
+      "harness": "Claude Code",
+      "reasoning_effort": null,
+      "hub_model": "Qwen/Qwen3.8-2.4T-A95B",
+      "metrics": {
+        "composite-score": 23.3,
+        "hypothesis-discovery-inspire-rate": 15.75,
+        "protocol-planning-mean-score": 52.0,
+        "scientific-computing-full-credit-rate": 8.7,
+        "scientific-software-operation-full-credit-rate": 50.0,
+        "experiment-observation-pass-rate": 13.33,
+        "end-to-end-research-score": 0.0
+      },
+      "mapping_sources": [
+        "https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B",
+        "https://www.qwencloud.com/news/qwen-3-8-max"
+      ],
+      "mapping_note": "Official model card identifies Qwen3.8-Max as the API version based on this model; API features differ from locally served weights.",
+      "hypothesis_discovery": {
+        "harness": "DiscoveryRSI",
+        "positives": 23,
+        "hypothesis_versions": 146
+      },
+      "cost_per_task_usd": 6.4325,
+      "agent_time_per_task_s": 2136.25
+    },
+    {
+      "id": "glm",
+      "name": "GLM-5.3 Flash",
+      "harness": "Claude Code",
+      "reasoning_effort": "high",
+      "hub_model": "zai-org/GLM-5.3-Flash",
+      "metrics": {
+        "composite-score": 21.33,
+        "hypothesis-discovery-inspire-rate": 11.63,
+        "protocol-planning-mean-score": 58.0,
+        "scientific-computing-full-credit-rate": 8.7,
+        "scientific-software-operation-full-credit-rate": 43.0,
+        "experiment-observation-pass-rate": 6.67,
+        "end-to-end-research-score": 0.0
+      },
+      "hypothesis_discovery": {
+        "harness": "DiscoveryRSI",
+        "positives": 5,
+        "hypothesis_versions": 43
+      },
+      "cost_per_task_usd": 0.38375,
+      "agent_time_per_task_s": 1725.75
+    },
+    {
+      "id": "kimi",
+      "name": "Kimi K3.1 Pro",
+      "harness": "Claude Code",
+      "reasoning_effort": null,
+      "hub_model": null,
+      "metrics": {
+        "composite-score": 28.92,
+        "hypothesis-discovery-inspire-rate": 9.21,
+        "protocol-planning-mean-score": 58.0,
+        "scientific-computing-full-credit-rate": 26.0,
+        "scientific-software-operation-full-credit-rate": 57.0,
+        "experiment-observation-pass-rate": 23.33,
+        "end-to-end-research-score": 0.0
+      },
+      "hypothesis_discovery": {
+        "harness": "DiscoveryRSI",
+        "positives": 7,
+        "hypothesis_versions": 76
+      },
+      "cost_per_task_usd": 2.8935,
+      "agent_time_per_task_s": 1272.75
+    },
+    {
+      "id": "minimax",
+      "name": "MiniMax-M3",
+      "harness": "Claude Code",
+      "reasoning_effort": "high",
+      "hub_model": "MiniMaxAI/MiniMax-M3",
+      "metrics": {
+        "composite-score": 16.82,
+        "hypothesis-discovery-inspire-rate": 3.61,
+        "protocol-planning-mean-score": 43.0,
+        "scientific-computing-full-credit-rate": 22.0,
+        "scientific-software-operation-full-credit-rate": 29.0,
+        "experiment-observation-pass-rate": 3.33,
+        "end-to-end-research-score": 0.0
+      },
+      "hypothesis_discovery": {
+        "harness": "DiscoveryRSI",
+        "positives": 3,
+        "hypothesis_versions": 83
+      },
+      "cost_per_task_usd": 1.14525,
+      "agent_time_per_task_s": 1727.25
+    },
+    {
+      "id": "mimo",
+      "name": "MiMo-V2.6-Pro",
+      "harness": "Claude Code",
+      "reasoning_effort": null,
+      "hub_model": "XiaomiMiMo/MiMo-V2.6-Pro-MOPD",
+      "metrics": {
+        "composite-score": 26.4,
+        "hypothesis-discovery-inspire-rate": 12.41,
+        "protocol-planning-mean-score": 49.0,
+        "scientific-computing-full-credit-rate": 30.0,
+        "scientific-software-operation-full-credit-rate": 57.0,
+        "experiment-observation-pass-rate": 10.0,
+        "end-to-end-research-score": 0.0
+      },
+      "mapping_sources": [
+        "https://mimo.xiaomi.com/blog/mimo-v2-6-tool-call-repetition",
+        "https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-MOPD"
+      ],
+      "mapping_note": "Mapped from the vendor API transition notice effective 2026-09-25 and recorded evaluation dates 2026-10-29/30; an API endpoint measurement, not a local-weight run.",
+      "hypothesis_discovery": {
+        "harness": "DiscoveryRSI",
+        "positives": 17,
+        "hypothesis_versions": 137
+      },
+      "cost_per_task_usd": 0.557,
+      "agent_time_per_task_s": 2341
+    }
+  ],
+  "composite": "Equal-weight mean of six assessments on the 0–100 scale: Hypothesis Discovery, Protocol Planning, Scientific Computing, Scientific Software Operation, Experiment Observation and End-to-end research. End-to-end research scores 0 for all listed models; Wet-lab Execution is not included.",
+  "website_results_updated": "2026-11-09",
+  "framework": "Six scientific capabilities and one end-to-end research assessment",
+  "status": {
+    "wet-lab-execution": "In construction",
+    "end-to-end-science": "To be expanded"
+  },
+  "aggregation": {
+    "members": [
+      "pp",
+      "eo",
+      "sc",
+      "so",
+      "hd",
+      "e2e"
+    ],
+    "weights": [
+      0.16666666666666666,
+      0.16666666666666666,
+      0.16666666666666666,
+      0.16666666666666666,
+      0.16666666666666666,
+      0.16666666666666666
+    ],
+    "score": "Equal-weight mean of six assessments on the 0–100 scale: Hypothesis Discovery, Protocol Planning, Scientific Computing, Scientific Software Operation, Experiment Observation and End-to-end research. End-to-end research scores 0 for all listed models; Wet-lab Execution is not included.",
+    "cost_time": "Equal-weight mean of the per-task means for Protocol Planning, Scientific Computing, Scientific Software Operation and Experiment Observation. Hypothesis Discovery and End-to-end research report neither. Not a task-count-weighted pooled mean.",
+    "precision": "Protocol Planning, Scientific Computing, Scientific Software Operation and Experiment Observation values are calculated from the rounded source chart values; Hypothesis Discovery values are reported to two decimals.",
+    "missing": "The score remains null unless all six assessment scores are available; cost and time remain null unless all four reported values are available."
+  },
+  "notes": [
+    "Gemini 4 Argon is evaluated on Protocol Planning, Scientific Computing, Scientific Software Operation and Experiment Observation. Hypothesis Discovery is pending, so its composite score is null.",
+    "GPT-6.1 Sol task-based scores are from the 9 November high-effort run; Hypothesis Discovery is the previously published medium-effort measurement, and the composite uses that unchanged value."
+  ]
+}