Skip to content
Affective
← Research

The bet, stated falsifiably

2026-09-14

Historical note. This article records the September model thesis and may describe mechanisms or schedules that are no longer the active POC plan. Read the current status first.
Lever study for the H5 bet: a 1.5B model with EOS on the long arm balancing a 70B conventional model on the short arm. Pre-registered, not a result.1.5B + EOS70B · CONVENTIONALABH5 · PRE-REGISTERED
Lever study. H5 as pre-registered. It has run and has not passed.

H5

A 1.5B-parameter EOS-augmented model should beat a 70B conventional model on emotional intelligence benchmarks. If that holds, architecture, not scale, is the emotional-intelligence axis, and a foundation-model raise is justified.

If it is false, we say so and stop. Testing it costs under $1,000. Every pass/fail threshold is pre-registered before the experiments run, and the result posts publicly either way.

Why pre-register

Because a benchmark you control after the fact is not a benchmark. The preregistration document fixes the baselines, the thresholds, and the kill criteria before a single experiment runs. This is the part most AI claims skip, and the part we consider non-negotiable.