Closure-SDK synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
A numerical memory subsystem, not an agent harness
Closure EA's strongest supported contribution is an implemented chain from accumulated observations and target feedback to changed numerical memory and later consumers. It can revise values, reorganize entries and use the resulting state on subsequent calls. Whether those changes improve future performance remains unestablished. The analysis covers the current Rust Closure EA subsystem within Closure-SDK at commit 9735f48a6ddd00b6f429a91443489531b3a115ef, with applicability cutoff 2026-10-13. SRC-1 supplies inspected implementation and attributed design, not observed execution or causal experiments. No target tests or examples were run. The sibling integrity SDK, geometric database, archived Python runtime, homeostasis demo and browser implementation are excluded; this is not a whole-monorepo assessment (boundary, Source register).
The runtime is synchronous local Rust computation. A caller supplies quaternion carriers—four-component numerical inputs—and, for teaching, expected targets. ThreeCell owns field reads, buffer aging, closure processing, corrective writes and maintenance. The caller owns scheduling, labels and save/load timing. Fixed symbolic rules operate over mutable dense geometry; there is no model-provider endpoint in these traced loops. Source terms such as “brain” and “semantic” do not establish human-like cognition or an LLM integration (RT-CMP-three-cell, RT-RTE-carrier-ingest).
From input to retained field
Ingestion puts a carrier into a live buffer, advances ages, computes error and chooses a channel from input geometry. ZREAD composes values from eligible genome and buffer entries; RESONATE finds a closest address. These are numerical selection operations, not construction of a text prompt. Retained addresses select entries, while their values determine returned content; the two can diverge. LOAD provides a genome reference for geometric verification. Genome hits use their exact index; buffer hits instead use a genome nearest-address fallback. Surprise and non-self Identity or Balance verification can tag observations for later retention (RT-RTE-carrier-ingest, EPI-RTE-verify-check, memory annotations).
Expired tagged buffer entries enter the genome; untagged entries are discarded. This gate is local, not a global requirement that every retained input passed independent validation. Supported structural products, higher-level closure packets, target associations, direct seeding, public mutable access and state reconstruction provide other admission paths. A Carry event can cascade into a higher level and write there directly; FixedPoint does not initiate that cascade (MEM-OBJ-buffer, MEM-RTE-hierarchy-retention, RT-RTE-direct-maintenance, RT-RTE-state-restoration).
The field separates DNA structural entries, Epigenetic event/closure entries and Response context/target associations. Layer-sensitive methods protect DNA values from certain corrections and omit DNA from merge/prune. That is a local update protocol, not immutable storage: mutable references and reconstructed state can bypass it. Later requested computations read retained geometry, and generation follows highest-count stored edges. Buffer lifetime is measured in explicit ticks. No token budget or model-visible context window is established; closure processing removes resolved history windows while retaining unresolved prefixes and derived products, not lossless pointers to discarded inputs (MEM-OBJ-dna, MEM-OBJ-epigenetic, MEM-OBJ-response, MEM-OBJ-closure-histories, RT-RTE-readonly-consumption).
Feedback changes organization without authenticating targets
A teaching caller commits a prediction into one replaceable pending slot, then supplies a target. Evaluation consumes that slot, calculates a geometric discrepancy and a threshold-based correct flag, credits eligible Response entries, writes the captured context/target association and changes sequential edges. Incorrect predictions can weaken an edge; discrepancy contributes maintenance pressure. Both correct and incorrect results proceed through association writing. The target is the caller's answer oracle, not independently authenticated reality (RT-OBJ-pending-prediction, RT-RTE-prediction-feedback, EPI-RTE-feedback-check, EPI-RTE-response-write).
Correct-credit attribution is narrower than update capacity. Teaching callers can label a buffer hit's index as a genome slot. Eligibility comes from Cell C, without establishing that it produced an arbitrary committed prediction. Maintenance remaps genome references but not pending indices. These problems prevent stable producer/slot guarantees; they do not erase the separate context/target write and its later consumers (report-verification limits).
Queries, commits and corrections also accumulate coupling, coalition, salience and coherence statistics. Maintenance consumes these traces to merge, prune and promote memory. Same-channel statistics include entries below ZREAD's value-contribution cutoff, so they are not solely a record of actual selected-value participants. Consolidation combines eligible same-layer entries and removes selected weak mutable entries, with a configurable coalition-survival exception for Response entries. When the coalition gate is enabled, eligible Response merge proposals prioritize co-resonance rather than universally nearest addresses. Force bypasses pressure, not all internal selection criteria (MEM-OBJ-genome-statistics, MEM-RTE-memory-consolidation, report-verification limits).
Promotion uses recurrence, salience, coupling, coalition and conditionally sufficiently sampled coherence, then checks higher-level novelty. Promoted values enter level-1 storage and explicit level-read capabilities; an automatic path back into primary level-0 prediction is not established. The promoted write retains selected support metadata, not the full episode or reasons for promotion. These are operational eligibility and organization changes, not acceptance and subsequent integration of a new explanatory proposition (EPI-RTE-category-check, EPI-RTE-category-disposition, MEM-RTE-memory-consolidation).
Later use, checkpoints and local warrants
evaluate composes a requested sequence and selects over genome plus live buffer without the ingestion channel restriction. generate starts from a genome seed and follows stored edges. Associative recall and OrbitRuntime are shipped consumers, not hypothetical storage clients. Numerical state therefore has a concrete channel to later outputs; exercised dependence and benefit remain unobserved (RT-RTE-readonly-consumption, RT-BAP-field-consumption).
Persistence has separate contracts. BrainState saves genomes, selected histories, buffer, pressure and pending feedback, but restoration resets neuromodulation and transient self-observation/salience. Genome-only reconstruction affords resumed training with fresh transient state and no restored higher-level runtime organization. Vocabulary is saved separately. Temporary-file rename provides best-effort replacement, not demonstrated crash durability, migration or automatic rollback. None of these capabilities establishes a shipped cross-process reuse sequence or equivalence to uninterrupted execution (MEM-OBJ-brain-state, RT-RTE-state-restoration, MEM-RTE-genome-reconstruction, MEM-OBJ-vocabulary).
Local evaluators exist, but their licenses differ. VERIFY establishes geometric closure within tolerance. Feedback establishes agreement with supplied targets. Accuracy code scores selected examples without enforcing held-out independence or automatically admitting a claim. The solver checks nearest-slot identity, not exact equation equality. Integer factor outputs have a conditional arithmetic witness license, while enumeration-based labels do not establish properties recovered from learned carrier content. Arbitrary carriers do not by themselves individuate external propositions. Token normalization and hash-derived addressing likewise do not establish unrestricted semantic preservation; nearest decoding lacks a semantic confidence veto. These findings neither validate nor refute the source's unassessed theorem and prime-zero correspondence claims (EPI-RTE-accuracy-report, EPI-RTE-slot-solver, EPI-RTE-factor-output, EPI-RTE-vocabulary-codec, EPI-CLM-closure-convergence, EPI-CLM-prime-zeros).
What the representative workflow establishes
The shipped neuromodulation experiment connects these operations in one bounded caller. Two independent configured brains receive stable or disrupted contexts for 20 rounds, each with eight context inputs before a pattern. Each round immediately commits and evaluates against a fixed label. Final forced consolidation follows copying the body tones. Returned tuples therefore mix pre-final-maintenance tones with promotion counts including final maintenance. The example prints comparisons and asserts disrupted arousal is higher and coherence lower before its final verified message. Those checks concern tone ordering, not recall accuracy or successful learning. Failure supplies no internal retry or rollback, and the example invokes no persistence. This is synchronous numerical experimentation, not established worker delegation or multi-agent delivery (WF-RTE-regime-comparison).
For the scenario actually supported—caller-led numerical association, corrective feedback and subsequent field reads—the implementation supplies automatic retained updates and real later consumers. For semantic knowledge production, promoted-category improvement of primary recall, or dependable cross-session continuity, the evidence does not establish the corresponding guarantees. Trace-fed adaptation is wired; improved capacity for future action is not demonstrated. No accepted member establishes that criticism of a consumed formulated theory improved future capacity. The runtime supplies no separate supported assessment of theory-builder conditions 1–4, so this synthesis assigns no system-wide result to those conditions. The represented self-discrepancy and body-state mechanisms are insufficient here for a complete reflection finding; synchronous internal computation likewise does not establish an autonomous theory builder. These are independent unresolved properties, not implications of missing performance evidence (RT-RTE-prediction-feedback, epistemic conclusion, profile verification).
Feedback and trace-conditioned maintenance provide an improvement-directed self-change pathway: discrepancy/statistics shape updates to the runtime's own field, and that field has later behavioral force. This supports a dispositional pathway, not exercised self-improvement: no supplied subsequent operation establishes causal dependence on the change, and no outcome establishes successful improvement. The source's convergence claims remain design assertions beyond that contribution (RT-CLM-teaching-convergence, MEM-RTE-memory-consolidation, RT-BAP-field-consumption).
An inspectable before/after run with linked updates and later outputs would strengthen the exercised-dependence finding. Controlled comparisons independently varying correction or consolidation would be needed for component benefit, rather than attributing a whole episode's outcome to one mechanism. Preserving population identity and remapping pending references would change the credit-fidelity assessment. Identifying a shipped promoted-state consumer would change the integration assessment; paired checkpoints and continuation comparisons would change the continuity assessment. Semantic claims require explicit carrier interpretation and evidence-consuming acceptance within a named scope, while theorem claims require audited correspondence to executable prerequisites. No new source reads were made for this synthesis.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Unresolved feedback credit lineage; carried report-verification limit | SRC-1; RT-OBJ-pending-prediction, RT-RTE-prediction-feedback, MEM-RTE-memory-consolidation; verification | Teaching population/index conversion, Cell-C eligibility and consolidation remapping in the frozen Rust subsystem | Stable source-slot and correct-credit attribution, including maintenance before commitment or feedback; separate association writes remain supported | Population-preserving handles, checked producer linkage and pending-reference remapping, with intervening-maintenance traces |
| Vocabulary replacement leaves stale reverse aliases; carried report-verification limit | SRC-1; MEM-OBJ-vocabulary, EPI-RTE-vocabulary-codec; verification | Register and reload of the forward/reverse maps | Replacement consistency, collision handling and fresh-versus-reloaded exact-display equivalence | Explicit alias/collision policy and tested replacement/reload behavior |
| Public distribute_credit omitted from the correction inventory; carried report-verification limit | SRC-1; RT-RTE-direct-maintenance, MEM-OBJ-epigenetic, MEM-OBJ-response, MEM-OBJ-genome-statistics; verification | Verifier inspected Genome's query/channel/coupling-selected non-DNA value/statistics update; bounded search identified no shipped invocation | Complete correction-API enumeration or universal Response-only correction comparison; canonical feedback remains supported | Accepted records tracing this alternative capability and any shipped consumer, without upgrading it to observed learning |
| Merge-order simplification; carried report-verification limit | SRC-1; RT-RTE-direct-maintenance, MEM-RTE-memory-consolidation; verification | Consolidation prioritizes Response co-resonance when its gate is enabled; disabled-gate and Epigenetic cases use address proximity | Universal nearest-pair-first ordering; eligibility/reorganization and the selected gate-disabled example remain supported | Corrected route account distinguishing configurable ordering branches |
| No observed operation or causal intervention | SRC-1; boundary, WF-RTE-regime-comparison, RT-RTE-prediction-feedback, EPI-RTE-accuracy-report | Static Rust implementation and source-native design only | Passing assertions, effective learning, exercised self-improvement, numerical performance, deployment reliability or component benefit | Registered execution artifacts linking updates to later consumption and objective outcomes; independently varied interventions for component attribution |
| Promoted-state downstream integration unresolved | SRC-1; MEM-RTE-hierarchy-retention, MEM-RTE-memory-consolidation | Higher-level writes and explicit level-read capabilities; bounded consumer searches | Automatic promoted-category influence or benefit in primary level-0 prediction | A shipped consumer route selecting promoted content into primary prediction, with subsequent dependence and benefit evidence separately |
| Checkpoint horizons and external durability unresolved | SRC-1; MEM-OBJ-brain-state, MEM-RTE-genome-reconstruction, RT-RTE-state-restoration, MEM-OBJ-vocabulary | Distinct state/genome/vocabulary serializers and constructors; selected transient resets; external filesystem contract | Coupled vocabulary/brain transaction, crash durability, migration, uninterrupted-state equivalence or observed cross-session effectiveness | Paired checkpoint protocol, explicit version/trust contracts, crash tests and traced resumed-versus-uninterrupted comparisons |
| Arbitrary carrier meaning and upstream admission provenance unresolved | SRC-1; MEM-OBJ-response, RT-RTE-direct-maintenance, EPI-RTE-vocabulary-codec, EPI-RTE-response-write | Numerical inputs, supplied labels, public mutation, checkpoint choice and lossy text codec | Target authenticity, unrestricted semantic preservation, manual upstream write control or accepted ampliative knowledge production | Specified encoding/interpretation and label provenance, explicit principal/control evidence, and recorded acceptance against evidence and intended scope |
| Included helper/evaluator alternatives and proofs remain partly uninspected | SRC-1; RT-CMP-three-cell, RT-RTE-readonly-consumption, EPI-RTE-factor-output, EPI-CLM-closure-convergence, EPI-CLM-prime-zeros; profile verification | Selected helper invocations, mathematical outputs and examples, not every internal body, evaluator or proof correspondence | Complete memory/property inventory, whole-boundary pull-only claims, exhaustive evaluator or derivation negatives, all-input arithmetic correctness or theorem/runtime guarantees | Accepted coverage of material alternatives, arithmetic/encoding audit and proof-to-runtime prerequisite correspondence |
| Deliberate subsystem exclusions | SRC-1; boundary | Current Rust Closure EA only; excluded sibling products, legacy runtime, separate homeostasis and browser implementations | Whole-SDK integrity/database, browser equivalence, legacy cognition or deployed-agent conclusions | A separately authorized analysis boundary and frozen evidence for those responsibilities |