Closure-SDK boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

This run analyses the current Closure EA Rust geometric memory and learning subsystem within Closure-SDK, not the entire monorepo and not just the Python data-integrity API. Its intended use is to explain how this subsystem admits carriers into memory, reads retained state, changes it through feedback and consolidation, and serializes it for reuse. The target class is memory/knowledge/context-engineering system; the boundary kind is subsystem-only.

The repository presents several distinct products. The root README calls closure_ea/ the Geometric Computer, closure_sdk/ the data-integrity SDK, and closure_dna/ the geometric database. The Python package publishes integrity and database commands, whereas the current learning subsystem has its own Rust crate and public ThreeCell surface. These are separate shipped responsibilities, not an external host inferred merely from a library-call boundary (SRC-1, README.md, pyproject.toml, closure_ea/Cargo.toml, closure_ea/src/lib.rs, closure_ea/__main__.py).

The subsystem qualifies on its implemented memory and learning functions. This does not classify it as an LLM agent harness or establish a model-provider integration. Its declared substrate is quaternion geometry; the inspected learning caller obtains a prediction from a field read, commits it, and evaluates it against a supplied target. The source's use of “brain,” “semantic,” or “cognitive” is not evidence of human-like cognition (SRC-1, closure_ea/src/lib.rs, closure_ea/src/teach.rs).

//! - [genome] — DNA + epigenetic layers, Law 3 ingest, nearest neighbor. //! Bootstrap seeds DNA; chunked learning writes epigenetic. //! - [field] — RESONATE and ZREAD over genome ∪ buffer. --- closure_ea/src/lib.rs

brain.commit_prediction(predicted, source); brain.evaluate_prediction(target); --- closure_ea/src/teach.rs

Functional scope

Included paths are the current crate's closure_ea/src/ modules, its manifest, closure_ea/examples/ shipped drivers and evaluators, and the inline tests in closure_ea/src/lib.rs. This includes substrate embedding and addressing, transient buffering, genome reads and writes, hierarchy emission, prediction feedback, consolidation, generation, state save/load, curriculum drivers, and runtime-backed execution consumers. The subsystem's own closure_ea/README.md and closure_ea/docs/ supply declared design; documentation claims remain distinct from implementation evidence. The initial inspected paths below are not an allowlist: further inspection of included paths at the same commit is permitted under SRC-1.

The entry-point trace establishes more than an API inventory:

• closure_ea/src/lib.rs exports ThreeCell, BrainState, memory modules and curriculum functions. ThreeCell::ingest receives a carrier, reads the genome and buffer, records closure-related admission signals, promotes eligible expired entries, and invokes consolidation. The boundary includes the entire implementation of these responsibilities, not only the initially read ranges (SRC-1, closure_ea/src/lib.rs, closure_ea/src/three_cell.rs).
• closure_ea/src/teach.rs drives ingestion and delayed target feedback. closure_ea/examples/exp_neuromodulated_learning.rs is a shipped caller that supplies context, commits predictions, evaluates them and forces consolidation. It is included, not treated as an excluded application because it supplies state (SRC-1, closure_ea/src/teach.rs, closure_ea/examples/exp_neuromodulated_learning.rs).
• closure_ea/examples/exp_associative_memory.rs writes input/target pairs through learn_response and later consumes the retained pairs through brain.evaluate. Its recall and interference checks are included evaluator code, not observed experimental results (SRC-1, closure_ea/examples/exp_associative_memory.rs).
• closure_ea/src/execution.rs embeds ThreeCell in OrbitRuntime, seeds orbit memory, and consumes it through evaluation. Those helpers and their shipped Rust example callers remain inside this boundary (SRC-1, closure_ea/src/execution.rs).
• ThreeCell exposes JSON file persistence and reconstruction. The inline round-trip test constructs nontrivial state, snapshots it, restores it and checks retained fields. This is source evidence for a persistence contract and an evaluator, not evidence that the test passed during this run (SRC-1, closure_ea/src/three_cell.rs, closure_ea/src/lib.rs).

let genome = brain.hierarchy_mut().genome_at_mut(0); for k in 0..N_PAIRS { genome.learn_response(&inputs[k], &targets[k]); } --- closure_ea/examples/exp_associative_memory.rs

pub fn load_state_from_file(path: &str) -> Result> { let json = std::fs::read_to_string(path)?; let state: BrainState = serde_json::from_str(&json)?; Ok(Self::from_brain_state(state)) } --- closure_ea/src/three_cell.rs

Exclusions and limits

These are deliberate scope exclusions, not inaccessible source. Included module bodies and examples not yet read are not yet inspected, not access gaps. There is no established source-access gap at this pin. The Rust crate declares sha2, serde and serde_json dependencies (SRC-1, closure_ea/Cargo.toml); their implementations, the compiler, filesystem guarantees and any enclosing application's scheduling are external and not attributed to Closure EA. A caller is responsible for supplying carriers and target feedback and choosing when to save or reload; the included crate methods and shipped callers establish only their own responsibilities.

The frozen repository is https://github.com/faltz009/Closure-SDK at 9735f48a6ddd00b6f429a91443489531b3a115ef. HEAD was checked against the supplied acquisition pin. The applicability cutoff is 2026-10-13, the supplied run date; it is not an assertion that later repository revisions behave identically. Overall evidence is code-grounded, supplemented by source-native doctrine/design. No target code, tests, examples, model calls or services were executed. No execution traces or causal experiments are registered. Source inspection therefore cannot establish measured correctness, learning effectiveness, runtime performance or deployment reliability.

Source register