expand_less Responsible: Ben Goertzel

SubRep (Subgoal Representation) transformsproposes transforming subgoal and option learning from an ad-hoc process into a disciplined, certifiable practice. It provides formal mathematical certificates thatdesigned to guarantee when a subgoal genuinely serves a larger purpose and when component solutions can be safely composed.

Status: Proposed. SubRep is described in the 2025 whitepaper (§5.9) as a framework for certified subgoal management within PRIMUS. The CDS/PDS admission rules and Motive Decomposition Network are theoretical designs; implementation is a research goal.

The Problem

Standard reinforcement learning discovers subgoals by shaping a single scalar reward and hoping the resulting skills transfer to new contexts. This approach is brittle: skills overfit to specific reward signals, transfer negatively to related tasks, and cannot be reliably composed into larger plans. PRIMUS needs subgoals that work across multiple motives, compose safely, and serve both neural controllers and symbolic reasoners.

Proposed Certificate-Driven Admission

SubRep introduces two formal admission rules that would screen candidate options (subgoal-reaching policies) before they enter the system's skill library:


CDS (Cone-Dominant Subgoals): An option is admitted if it is robust across a family of motives — not just optimal for one reward, but beneficial across a "cone" of related objectives. This prevents brittle specialization.
PDS (Pareto-Dominant Subgoals): When genuine trade-offs exist between motives, an option is admitted if it is explicitly Pareto-good on a small covering set — meaning no other available option dominates it on all objectives simultaneously.


These certificates arewould be expressed in the same language the rest of Hyperon speaks — backed-up values over AtomSpace features — so they could benefit from and contribute to the broader neurosymbolic loop.

Motive Decomposition Network

SubRep co-learnsproposes co-learning a Motive Decomposition Network (MDN) that decomposes high-level motives into achievable subgoals, learning the geometry of how motives relate to each other and to available skills. The MDN maps the motive space into the option space, identifying which skills serve which purposes and where gaps exist.

Neurosymbolic NativenessDesign

Unlike standard RL option frameworks, SubRep worksis designed to work equally for:

Neural controllers (learned policies)
Logic macros (PLN-derived inference strategies)
Evolved programs (MOSES/GEO-EVO outputs)


All arewould be screened by the same CDS/PDS admission rules and composed by the same planner.planner, Thismaking makes SubRep a universal interface between PRIMUS's diverse skill generators and its planning system.

IntegrationRelationship withto MetaMo and TransWeave

SubRep is adesigned naturalas a complement to MetaMo (motivational framework): MetaMo defines what the system cares about as motive geometries; SubRep validates which skills actually serve those motives with formal certificates. Together they form PRIMUS's goal management layer.

When TransWeave transfers skills to new domains, SubRep certificates would travel with themthem. — ensuring that transferred skills maintain their safety and utility guarantees in the new context.

Key References

Goertzel, B. (2025). Hyperon for AGI⇒ASI Whitepaper, §5.9: SubRep