Earned Growth: A Criteria-First Framework for Developmental Models of a Brain
Year: 2026
Type: preprint
Abstract: Technical note (v1.1, 2026-11-09). This note sets out three commitments for models that claim to develop rather than to be trained, and reports the criteria, the evidence available for each, and the negative results that constrain the framework. 1. Earned growth - a structural addition is admitted only after a pre-registered criterion for a specific capability has failed; a failed criterion means no growth rather than slower growth.2. Counterfactual agency - "I did it" is defined as the residual surprise under the counterfactual of not acting, indexed by ABI = agency - agency_bound.3. Developmental ladder - levels L0-L8, each with its own criterion, none of which can stand in for another. All mechanisms are disabled by default; with mechanisms disabled the implementation reproduces the earlier behaviour exactly (max|delta| = 0.0 across five ablations). Contents: 5-page PDF. Code and reproduction scripts will be released with the accompanying software paper. Use of AI tools: the framework, criteria and interpretation are the author's; drafting, language editing and implementation were assisted by an AI coding/writing assistant; every reported number was checked against the original run files.
Author: Jiang, Yuntong
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225656723
Cites:
Cited by:
Related to: