Attribution graphs for persistent skills

[WGlynn]

Odysseus already tracks user/AI co-evolution through the persistent-skills system. The "your agent evolves over time as it better understands you" property is real architectural state, not just a feature description.

I've been thinking about what it would take to surface that evolution as an attribution graph. Specifically: track contribution events (user prompt, AI generation, user correction, refinement, derived-from-skill references) as nodes in a DAG, and compute Shapley-value shares over the graph when someone asks "who contributed what to this skill."

Shapley has the property that four classical axioms (efficiency, symmetry, linearity, null player) uniquely characterize the decomposition. There's no other allocation rule that satisfies all four. For skills that mutate over time, an extra pairwise-proportionality axiom keeps the decomposition fair across rule-set changes, so if someone adds a new contribution, prior shares don't get reshuffled relative to each other.

A few concrete things this would answer. "Did I do this or did the agent?" turns from a memory question into a computed one. "Who owns the evolved skill if two people contributed?" goes from undefined to well-defined. And the AGPL discussion raises real derivative-ownership questions that a legal license alone can't answer at per-skill granularity. A data-layer attribution graph can.

Implementation-wise it's lighter than it sounds. The contribution events sit in a new SQLite table next to the existing skills schema. Shapley computation is on-demand, not on every event, so the write-path cost is zero. For skills with many contributors the exact computation gets expensive at O(2^N), so for N greater than about 8 you'd use the Owen-value approximation, which groups by actor class first and gives bounded-error results in polynomial time. The UI surface is a read-only "provenance" panel, opt-in.

There's a longer trajectory where this becomes the substrate for federated multi-instance Odysseus, attribution-preserving skill marketplaces, cooperative ownership of evolved skills, and optional cryptographic anchoring of attribution roots for instances that want strong portability. Naming those for transparency. None of them are part of what I'm proposing here. The data-layer piece stands on its own.

Happy to put together a longer RFC with the formal definitions, the lazy-compute algorithm, and a reference Python module if there's interest. Or scope it smaller (record-only, no UI) if that fits better.

Does attribution-thinking fit anywhere in the Stabilization v1 milestone, sit downstream, or out of scope? Curious how the existing coordination thread and the AGPL thread map onto it from your side.