Paper mapping

Section-by-section correspondence between D. Mojila, "LLM Orchestration with Geometry-Aware Memory" (IEEE ICBDML 2026, doi:10.1109/ICBDML68582.2026.11544690) and this package. Equation numbers are the paper's.

Paper	What it says	Where it lives here
§I Introduction, §II Related Work	Motivation and context	Home, FAQ
§III-A Geometric Knowledge Representation (eq. 1–3)	Poincaré ball, metric tensor, hyperbolic distance	poincare-memory; used through Scenario.build_memory() and Memory.search(mode="subtree")
§III-B Metacognitive Control Theory (eq. 4)	Control signal C(t) = F(M(t), G(t), R(t))	Orchestrator.run: state = plan and responses, goal = constraints, resources = ControllerPolicy.total_budget
§III-B (eq. 5)	Hierarchical decomposition Q → {q_1, …, q_n}	controller.decompose → Plan.subtasks with depends_on (Controller)
§III-C Information-Theoretic Resource Allocation (eq. 6)	I*(q_i) = argmax_I G(q_i, I) / C(I)	InvestmentScheduler.expected_gain, .allocate (skip rule), .measure_yield, .reinvest (Controller)
§III-D Multi-Objective Response Optimization (eq. 7)	max J(R) = Σ w_i f_i(R) s.t. f_j(R) ≥ ε_j	synthesis.weighted_sum; weights and thresholds are ControllerPolicy.objectives / .constraints (Synthesis)
§IV-A Overview	Four components: controller, memory, ensemble, synthesis	controller.py, poincare_memory.Memory, ensemble.py, synthesis.py, wired by orchestrator.py
§IV-B Metacognitive Controller: query analysis and intent recognition	Intent, domain requirements, complexity	controller.analyze → QueryAnalysis
§IV-B: task decomposition	Sub-tasks with dependencies and execution order	controller.decompose; the dependency-respecting loop in Orchestrator.run
§IV-B: capability matching	Static profiles and performance characteristics	CapabilityProfile.fit, role selection in decompose; dynamic health in Router
§IV-B: resource allocation ("glycolytic investment cycle")	Invest by expected yield, re-invest	InvestmentScheduler.allocate then .reinvest in the refinement loop
§IV-C Geometry-Aware Memory System	Structured representation, temporal context, retrieval, cross-domain integration	Scenario.knowledge → poincare_memory.Memory with a decay policy (ControllerPolicy.decay); Orchestrator._context retrieval order
§IV-D Provider-Agnostic LLM Ensemble: capability profiling	Domain expertise, reasoning, creativity, reliability	CapabilityProfile fields of the same names
§IV-D: throttling detection and mitigation	Response patterns, computation time, information density	ThrottleDetector signals: refusal, latency, length, density (Ensemble)
§IV-D: response standardization	One schema for every model output	schema.AgentResponse, schema.Claim
§IV-D: load balancing and failover	Balance across models, fail over on degradation	Router: candidate order, cooldown, failover, provider-exception handling
§IV-E Response Synthesis: multi-source integration	Fuse partial results	synthesis.fuse_claims
§IV-E: quality assessment	Multiple quality dimensions	the five objectives in synthesis.score_objectives
§IV-E: uncertainty quantification	Calibrated uncertainty estimates	SynthesisResult.uncertainty (a proxy, see Synthesis)
§IV-E: iterative refinement	Re-process when thresholds fail	the refinement loop in Orchestrator.run (refine_rounds)
§V Experimental Evaluation, Table I	Expert-rated accuracy, bits/token, response time, satisfaction	Not reproduced. This package reports facet recall and synthesis scores on its own scenarios (Evaluation)
§V-E Failure Mode Analysis	What goes wrong and how it is handled	Skipped sub-tasks, refused responses, throttle events and error: <type> events are all visible in RunResult
§VI Discussion, §VII Conclusion	Insights, efficiency, limitations	FAQ

Equations as quoted in the paper

Eq. 5 (hierarchical decomposition):

\[Q \rightarrow \{q_1, q_2, \ldots, q_n\}\]

Eq. 6 (optimal investment per sub-task):

\[I^*(q_i) = \arg\max_{I} \frac{G(q_i, I)}{C(I)}\]

Eq. 7 (weighted sum with epsilon constraints):

\[\max_{R}\; J(R) = \sum_{i=1}^{m} w_i f_i(R) \quad \text{s.t.} \quad f_j(R) \geq \epsilon_j,\ \forall j \in \{1, 2, \ldots, m\}\]

The controller and synthesis pages show how each symbol is instantiated in code.