Read shard theory (theoretical lineage for User Modeling / Topic 7)

Reading task: work through shard theory as part of the theoretical-lineage reading for the User Modeling / Persona Selection Model thread (Topic 7, research_ideas.md).

What it is

Shard theory (Pope, Turner/TurnTrout et al., Alignment Forum ~2022) is an account of how values and behavioral tendencies form in RL/learning agents: reinforcement creates context-activated decision-influences ("shards") rather than a single coherent utility function. Values are contextual, multiple, and emerge from the training process.

Why it matters here

Connects to the persona/value-formation framing in this project:

• Persona Selection Model treats generation as character simulation; shard theory offers a mechanistic story for how the distribution of "characters"/value-shards gets laid down during training. Sits alongside subtask 7.2 (Janus's Simulators essay + Conditioning Predictive Models + Hubinger 2023) as theoretical lineage rather than a competing frame.
• Relevant to emergent misalignment: if fine-tuning activates a misaligned shard rather than overwriting values, that reframes the "predict EM from training-data signals" angle.
• Possible bridge between the RL-values literature and the LLM persona-space work the project already cites.

Open questions / done condition

• Does shard theory's "contextual decision-influences" map onto persona axes / persona vectors, or is it a different ontology?
• Is there anything actionable for the User Modeling gist's "Theoretical lineage" section (same place subtask 7.2 output lands)?
• Done condition: short synthesis note (3-5 bullets) on whether shard theory joins the project's reference list, appended to the User Modeling gist's theoretical-lineage section.

Source

Captured from Todoist via my-goat auto-processor. Queue file: 2026-05-29T11-00-04_todoist-6gjrvpjrvR7M348M_read-shard-theory.md Related: research_ideas.md Topic 7 (User Modeling), subtask 7.2 (Simulators literature reading).