Skills are shipped flat—a directory of self-contained packages—while they are used relationally: to apply one correctly a model must place it among the others. That relational structure is latent in the library, and a flat list discards it, forcing the model to reconstruct it on every task. The paper develops the structural argument (agent operation as a flow over the structure of its context), documents the cost of the flat default, and demonstrates the filesystem-level correction. The claim is substrate-correction, not agent-performance; the corresponding agent-task benchmark is specified as future work, not claimed. Working paper, June–July 2026.
The construction paper. It specifies and implements a coordinate-addressed tree over an ordinary skill directory, on a substrate (the filesystem, as exposed to a coding agent) that provides no first-class edges. The tree emulates the two things a graph would supply natively: identity, via a hierarchical coordinate (0.2.1) assigned to every node, and edges, via breadcrumbs—instructions, embedded in each node, to load a specific child through the one operation that actually injects it into the model's context. Around this: retrieval that survives the flattening (internal nodes carry a summary of their subtree) and a self-coherence loop (discover/cohere/emit/unemit) that reconstructs, audits, repairs, and losslessly reverts the on-disk structure. Implemented as one MIT-licensed tool (skilltree) with a passing test suite.
The generalization: why a tree is not enough. Skill libraries require graph structure because operational primitives participate in multiple workflows; a shared operational primitive cannot be faithfully represented by a single-parent skill tree without duplication or artificial ownership. A typed hyperedge represents the primitive once and makes it reachable from every workflow that requires it: trees are deployable projections, hyperedges preserve shared operational reality. The runnable existence proof is graph-skills-experimental (MIT, pure stdlib): skills as hypergraph nodes, typed edges as context-injecting state transitions, chart() deriving tree-options, and a lift of the Paper 2 coordinate tree as the degenerate case.
A minimal, deterministic method for the missing hierarchy: assign every node in a skill folder a hierarchical coordinate, render a navigable map (tree, addressable index, branch summaries), and let the agent open a branch and descend to the coordinate it needs. Approximately 90 lines, one command, zero dependencies, non-destructive. The note's own stated scope: the construction and addressing are exact; the retrieval improvement over load-all is the expected, literature-supported direction but is asserted, not benchmarked—a measurement to run, not a result to claim.
Spatial structure answers "where is the skill?"; it does not answer "where am I in the work?" This note adds the missing temporal structure with a dependency-free state machine over the coordinate tree: a persisted current position plus a descent call stack, so the agent resumes instead of restarting—across turns and across sessions. Verified end-to-end at the level of the mechanism (state survives the process); the task-completion improvement over stateless re-reading is asserted as the expected direction, not benchmarked.
A set of formal notes on the order-theoretic semantics underlying the substrate program.
A treatment of graded oversight for autonomous agent systems—under what conditions, and by what mechanism, the external acceptance authority over an agent's work can be progressively substituted while remaining accountable. The mechanism paper of the program's alignment position: trust enforced at the joints between parts, never presumed inside any component. Research-area statement →
Pre-trained corpora embed the relational structures of contemplative traditions (vow mechanics, witnessed transmission, role dynamics). Rather than asking a model to reason about these concepts, the harness has it hold one as a boundary while working, with the environment enforcing what the framing describes. The paper formalizes a three-level loop hierarchy (task commitment, emanation requirement, goal orthogonality) and specifies a five-condition experiment separating roleplay, enforcement, and their coupling. Practitioner-authored, from formal lineage training; mechanics primary throughout. Research-area statement →