若水研究院

Docs

Beginner's guide

One diagram to see what MEI rates, how we rate it, and how to read a conclusion. World models only. No cross-domain leaderboard.

World-model rating system at a glanceFlow from rated object to dual-axis conclusion, seven dimensions, three method layers, and run conditions.01What we rateModel version × use-domain02Dual-axis callLevel × evidence strength03Seven dimsShown separatelyThree method layersL1Replication audit

Independently re-run public benchmarks; publish values and conditions.

L2Independent probes

Fill gaps on uncertainty calibration and failure boundaries.

L3Rating judgment

Level × strength by use-domain, with a reasoning chain.

04Run conditions are evidenceN · precision · GPU · subset / offload — on the same card
Read left to right, then down: version × domain → level × strength → seven dimensions → L1–L3 sources → run conditions as the trust boundary.

The rated object is always world-model version × use-domain. Domains lack a reliable common order, so we never publish a total ranking.

Conclusions use two axes: level (1 / 2 / 3 / NR) for maturity, evidence strength (high / medium / low) for how strong the evidence is. They must appear as a pair.

Seven dimensions stay separate—never one total score. Run conditions (N, precision, GPU, subset, …) ship on the same card. Beyond “does it look good,” we ask “dare you trust it.”

How to read a model cardRead the methodology