# Native Hub leaderboard definitions

This specification describes the website results synchronized on **5 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).

**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).

## Task scope and metrics

SLE covers **six scientific capabilities and one end-to-end research assessment**, with **113 tasks in v1**.
Wet-lab Execution is **In construction** and is not scored. End-to-end research is **To be expanded**.
The `default` configuration / `v1` split indexes 113 task cards: Hypothesis Discovery 5, Protocol Planning 40,
Scientific Computing 23, Scientific Software Operation 14, Experiment Observation 30 and End-to-end research 1.
Wet-lab Execution has no released task. All scored metrics use a 0–100 scale and higher is better.

| Hub task ID | Definition |
| --- | --- |
| `composite-score` | Equal-weight mean of the scored assessments below, at 1/6 each. |
| `hypothesis-discovery-inspire-rate` | Percentage of hypothesis versions that anticipate research published after the model's knowledge cutoff; reported with positives / hypothesis versions. |
| `protocol-planning-mean-score` | Mean task score for 40 Protocol Planning tasks, on a 0–100 scale. |
| `scientific-computing-full-credit-rate` | Percentage of 23 tasks passing every scientific check; not mean repair gain. |
| `scientific-software-operation-full-credit-rate` | Percentage of 14 tasks receiving full credit; not mean partial-credit reward. |
| `experiment-observation-pass-rate` | Percentage of 30 tasks with every requested field correct. |
| `end-to-end-research-score` | End-to-end research score; 0 for all currently listed models, included in the composite. |

The composite equally weights Hypothesis Discovery, Protocol Planning, Scientific Computing, Scientific
Software Operation, Experiment Observation and End-to-end research on a 0–100 scale, at **1/6 each**.
End-to-end research scores **0 for all currently listed models**; this is a published result, not a fixed scoring rule. Wet-lab Execution is In construction and is not scored.
Cost and time equally average the per-task means for Protocol Planning, Scientific Computing, Scientific Software Operation
and Experiment Observation; Hypothesis Discovery and End-to-end research report neither.

The composite is not an overall task pass rate. The runner's raw capability-balanced mean and the end-to-end
progress rubric are diagnostic outputs, not replacements for the official leaderboard metrics.

## Frozen values and provenance

`leaderboard-results-20261105.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.

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
checkpoints. No cryptographic verification badge is claimed.

## Model repository mapping

Native submissions target these official repositories:

| Website model | Official Hub repository | Harness (task-based / Hypothesis Discovery) | Reported effort |
| --- | --- | --- | --- |
| DeepSeek V4.1 Flash | `deepseek-ai/DeepSeek-V4.1-Flash-1017` | Claude Code / DiscoveryRSI | High |
| GLM-5.3 Flash | `zai-org/GLM-5.3-Flash` | Claude Code / DiscoveryRSI | High |
| MiniMax-M3 | `MiniMaxAI/MiniMax-M3` | Claude Code / DiscoveryRSI | High |
| Qwen3.8 Max API | `Qwen/Qwen3.8-2.4T-A95B` | Claude Code / DiscoveryRSI | Not reported |
| MiMo-V2.6-Pro API (MOPD update) | `XiaomiMiMo/MiMo-V2.6-Pro-MOPD` | Claude Code / DiscoveryRSI | Not reported |

The [Qwen model card](https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B) identifies Qwen3.8-Max as the official API version based on Qwen3.8-2.4T-A95B, with additional API features. The submission is explicitly labeled as an API-version measurement; it does not claim identical behavior from locally served weights.

[Xiaomi's update notice](https://mimo.xiaomi.com/blog/mimo-v2-6-tool-call-repetition) states that the MOPD update reached its API on 25 September 2026 without changing `mimo-v2.6-pro`. The recorded SLE MiMo attempts are dated 29–30 October 2026. This is the evidence used to map the measured API version to `MiMo-V2.6-Pro-MOPD`; the [MOPD model card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-MOPD) also lists API and OpenRouter availability.

As checked on 3 November 2026, GPT-6.1 Sol, Claude Opus 5 and Kimi K3.1 Pro do not have an established matching official Hub model repository. They remain in the complete eight-model dataset-card table. Kimi K3 is a different model and is not substituted for K3.1 Pro.

## Hub integration

The root `eval.yaml` declares the composite and its scored assessment metrics with the Harbor evaluation framework. Result records belong in `.eval_results/*.yaml` in each matching **model** repository, not in this dataset repository. Each record links to this dataset, its task ID, a pinned dataset revision and the published result source.

Hugging Face controls Benchmark allow-list activation. Configuration and submitted result files alone are not a claim of Official Benchmark status. The three SLE data repositories remain in private preview unless their maintainers explicitly change access.