What we are honest about

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.

Open items

H5 has run and has not passed: the EOS model does not beat a matched linear probe on the same features.

The first test of emotional memory (Arm M) did not support it: carried EOS state did no better than reading the current message alone.

A plain fine-tuned 1.5B, with no EOS modules, does at least as well as the EOS arms. An early result that it matched large models came from a test-set leak; with the leak removed and the scoring corrected, it sits below them on our current test.

AGP is a proposed algorithm, pre-registered, not yet validated.

If INT4 quantization degrades probe accuracy, serving cost doubles. Tracked in the financial model, not hidden.

The EQ-Bench correlation objection is real; our answer is a falsifiable sub-check, not a dismissal.

Consciousness-adjacent properties of the architecture are governed explicitly, not hand-waved.