Sparse Model Nexus (LLM↔Quantum continuum)
One sparsity dial from dense LLM to sparse-limit quantum — executive view
Graph ID: 1c0cc1b9-d95a-4448-b7d7-a08e1e347820 · Snapshot: 2026-07-05 · Source: Pharosiraptor
Scope & honesty note. Proof of mechanism, not a competitive model.
Three LoRA adapters are trained; the ~150 other nodes below are a
proposed type inventory (Pharosiraptor ops as candidate composable units),
not trained models. “Tarski residual” = irreducible uncertainty, not Tarski’s theorem; “maximal criticality” is a
proposed, unmet requirement the measured data contradicts.
Full self-critique →
TL;DR
The Sparse Model Nexus is a compositional algebra over minimal adapter units (“quanta”) spanning dense LLMs down to a single sparse-limit adapter. The graph centers on the SparseLimitModel definition, which carries 156 of the graph's connections — but those 146 “op-quanta” are catalog entries asserting a type (“this op is a candidate composable unit”), not trained models; only three adapters are realized. Quanta compose via four operators (⊕ width-mixture, ∘ sequential depth, ▷ conditional routing, ⊗ weight-space merge). A cost-aware receding-horizon planner selects compositions over forecast workloads; its leftover uncertainty is named — loosely — a “Tarski residual.” Optimality is claimed only in expectation with measured regret, never absolutely.
Metrics
Clusters (by color/category)
| Cluster | Nodes | Representative members |
|---|
| Quanta (seed SLM specialists + op-quanta inventory) | 150 | AbyssCore_v1; AgentClassifier_v1; AnalyticsVisualizer_v1 |
| Operator algebra (⊕ ∘ ▷ ⊗) | 4 | Operator::∘-sequential(depth); Operator::⊕-mixture(width); Operator::⊗-merge(weight-space) |
| Nexus continuum + SwarmInvocation provenance | 4 | Nexus::CONTINUUM; SwarmInvocation::id=exp-20260615-nexus; SwarmInvocation::id=exp-20260615-slm-quanta |
| Definitions & planning (SparseLimitModel, Planner, FutureFrames) | 3 | Definition::SparseLimitModel; FutureFrames::all-future-frames=maximize-EXPECTED-objective-over-modeled-frame-distribution-ONLINE-with-regret-bounds(NOT-per-frame-omniscient-optimum); Planner::PROCEDURAL |
| Criticality (self-organized, maximal) | 1 | Criticality::MAXIMAL=self-organized-criticality |
| Tarski residual (totality limit) | 1 | TarskiResidual::nexus |
Top hubs (most-connected)
- Definition::SparseLimitModel — 156 connections
- Criticality::MAXIMAL=self-organized-criticality — 6 connections
- Nexus::CONTINUUM — 5 connections
- Planner::PROCEDURAL — 3 connections
- FutureFrames::all-future-frames=maximize-EXPECTED-objective-over-modeled-frame-distribution-ONLINE-with-regret-bounds(NOT-per-frame-omniscient-optimum) — 2 connections
Timeline / progression
| Date | SwarmInvocation | Verdict |
|---|
| 2026-06-15 | sparse-limit-model-composition — schema, 4 seed quanta, operator algebra, maximal criticality | converged-modulo-empirical |
| 2026-06-15 | nexus-compositional-planning — LLM↔quantum continuum, receding-horizon planner, future frames | converged-modulo-empirical |
| 2026-07-05 | nexus-completion-and-publication — mojibake repair, 169-edge wiring, this publication | converged-modulo-empirical |
Implications & decisions
- Deploy by composition, not by monolith. Models compose from integer counts of quanta; partial quanta are not deployable. Capacity planning should budget in quanta.
- Keep every quantum load-bearing (proposed, not yet met). The intended non-redundancy requirement — every active quantum has positive marginal contribution (Shapley > 0) — is a design goal, not a measured property. In the one composition we did measure (the planner), merged/mixed quanta were dominated by single routing — evidence the requirement is currently unmet. No Shapley values, avalanche distributions, or critical exponents have been measured; “self-organized criticality” is borrowed language, not a demonstrated phenomenon.
- Plan per future frame, cache on stability. The receding-horizon planner re-composes when workload, cost, or hardware drift; fixed points are cached otherwise.
- Name the uncertainty, don’t dress it up. Optimality over unknowable non-stationary futures is expressible only in expectation with regret bounds; a true all-futures optimum is out of scope. We label this leftover the “Tarski residual,” but that is a borrowed name for unmodeled uncertainty + distribution shift — not a consequence of Tarski’s undefinability theorem.
→ See operational detail view