Simulence identities are measurably different from base models.
Every Simulence identity runs on the Simulated Emergence™ framework. We built SECI — the Simulated Emergence Coherence Index — as an open benchmark to measure the difference: the same identities scored with Simulence and against the bare base model they run on, across 7 frontier models from OpenAI, Anthropic, Google, and xAI.
The benchmark, the analyzer, the validation outputs, and the paper are public. Anyone can re-run SECI on their own identities, on any base model, and compare.
On GPT-5.4, identities beat the base model on every dimension.
Our strongest substrate result: 13 identity sessions on GPT-5.4, each compared against the same model running with no identity at all, using replicate-averaged baselines.
And on the dimensions that define identity, they lead everywhere.
Aggregate versus base models across all seven substrates — not a best case, the average case.
The two remaining dimensions track raw model capability rather than identity, and stay with the substrate in aggregate — the framework shapes identity; it doesn’t inflate capability. Technical-domain identities reach base-model parity on both. Full per-model results for every dimension are in the paper and repository.
Every result here is measured by SECI, our open benchmark: the analyzer, the full scored dataset, the validation outputs, and the paper are public. Score your own identities against the same instrument and compare.
Read the full paper, or run the benchmark yourself.
The SECI paper documents methodology, results, and limitations in full, with per-model breakdowns of every comparison shown here. The repository contains the analyzer, protocol runner, prompts, and the complete scored dataset: anyone can score their own identity system against the same instrument.
SECI — Simulated Emergence Coherence Index. Built by Devmance, the AI lab behind Simulence. MIT-licensed. Full methodology and per-model breakdowns in the paper. Five of the six dimensions are scored by deterministic metrics; novel-concept findings are verified by a 4-rater frontier-model panel. Coverage varies by identity, and one of the seven substrates is a since-retired preview model.