Project & Research Roadmap¶

This document outlines the strategic engineering and research roadmap for Episteme.

The project unifies computational epistemic research (formal TheoryNet models, structuralist philosophy of science, and epistemetrics) with a production-grade software engine (episteme-pipeline, epistemetrics, and episteme-studio).

Horizon 1: Platform Stabilization & Core Engine¶

Current Focus | Active Development

Focuses on completing the transition to a non-linear stage-execution engine, operationalizing the automated evaluation harness, and hardening the user workbench.

DAG Stage Execution Engine (_execute_via_plan)¶

Refactors pipeline execution from a sequential 1-based ordinal loop into an arbitrary Directed Acyclic Graph (DAG) stage execution plan (_build_execution_plan, _resolve_phase_order, _execute_via_plan).

Key Capabilities & Milestones:

• Supports non-linear, branching, and conditionally skipped phase topologies.
• Preserves deterministic RunManifest tracking and artifact-envelope provenance across concurrent branches.
• Enables injecting analytical passes at arbitrary graph boundaries rather than appending solely at the tail.

Evaluation Harness Execution & Scorer Wiring¶

Wires intrinsic and extrinsic scorers into the EvaluationRun orchestrator (pipeline/evaluation/).

Key Capabilities & Milestones:

• Connects Graph BERTScore (G-BS) and Optimal Edit Paths (OEP) with injected embedding functions for soft semantic matching.
• Binds extrinsic ranking metrics (MRR, nDCG@k, Hits@k, AP) to evaluation datasets.
• Formats evaluation results into standardized EvaluationReport Pydantic models and automated Markdown summaries.

Neo4j Structural Reasoning & Constraint Validation¶

Integrates GraphValidator checks directly into the pipeline lifecycle.

Key Capabilities & Milestones:

• Enforces polarity-aware disjointness (preventing simultaneous positive and negative edges to identical targets).
• Validates L3_COMPONENT_TYPES and structural constraints natively in Neo4j via async Cypher.
• Emits EvaluationCompleted and ValidationViolationDetected domain events to the event bus.

Episteme Studio Workbench v0.2¶

Hardens the run-oriented workbench for researchers and reviewers.

Key Capabilities & Milestones:

• Interactive tri-layer graph explorer (L1 Provenance, L2 Knowledge Graph, L3 TheoryNet).
• Visual run-to-run identity diffing (diff_artifacts) showing added, modified, and stale entities.
• Interactive candidate triple triage and tenability blur inspection (\(\delta^*\)).

Horizon 2: Scalability, Benchmarking & Dialectical Modeling¶

Medium-Term | Under Evaluation

Focuses on validating extraction quality against external academic baselines, scaling to large-scale literature corpora, and deepening gradual argumentation semantics.

Authentic Human Baseline & LLM-as-a-Judge Calibration¶

Establishes a human-expert baseline on authentic scientific texts before deploying automated LLM judges.

Key Capabilities & Milestones:

• Curate and verify gold-standard annotations from authentic corpus runs (planwirtschaft.md, teachers_expectancies.md).
• Calculate inter-annotator agreement (Krippendorff's \(\alpha\)) across domain-expert evaluation rubrics.
• Mathematically correlate human quality judgments with LLM-as-a-judge scores, enforcing strict calibration before deployment.

Standardized Public Benchmark Adapters¶

Standardizes evaluation against recognized NLP and argument mining benchmarks.

Key Capabilities & Milestones:

• Arg-Microtexts Adapter: Benchmark Layer 3 argument discourse units and support/attack relation classification.
• SciERC Adapter: Benchmark Layer 2 scientific entity recognition and relation extraction.
• SciFact Adapter: Benchmark extrinsic claim verification and evidence retrieval.

Weak Supervision for Logical Patterns¶

Bootstraps extraction classifiers without expensive manual labeling.

Key Capabilities & Milestones:

• Formulate high-precision Corpus Query Processor (CQP) patterns for discourse markers and dialectical connectives.
• Train few-shot classifiers (e.g., SetFit) using query matches as silver-standard training data.

Advanced Gradual Semantics Solvers (QBAF & TheoryNet)¶

Implements high-performance iterative convergence algorithms for dense, cyclic argumentation networks. Governed by ADR 0011.

Key Capabilities & Milestones:

• Implements gradual argumentation semantics (Eulerian, quadratic energy, and categorization models).
• Solves for equilibrium justification degrees (\(\tau\)) and empirical content ratios (\(B \cap P\)) in TheoryNet (\(\text{TF} = (\text{At}, R)\)).

Streaming & Large Corpus Ingestion¶

Enables memory-bounded ingestion of multi-chapter scientific treatises and book-length corpora.

Key Capabilities & Milestones:

• Incremental, streaming chunk generation with ToC boundary preservation.
• Windowed episodic memory eviction to prevent RAM exhaustion during multi-hundred-page processing.

Pluggable Vector Store Adapters¶

Supports specialized vector backends alongside Neo4j.

Key Capabilities & Milestones:

• Reference adapters for Qdrant, Milvus, and pgvector implementing GraphReader / vector search protocols.
• Decouples dense similarity candidate blocking from property graph traversal.

TypeSafe / Jev Typed Classification Engine (JevClassifier)¶

Integrates TypeSafe's Jev as a non-generative, typed classification adapter across the pipeline's classification boundaries.

Unlike generative LLMs, Jev directly emits typed values, confidence scores, and calibrated probabilities with sub-second response times (70–500ms) at 40–200× lower cost and latency.

Key Capabilities & Milestones:

• JevACCClassifier: High-throughput argument component classification (CLAIM, PREMISE, CONCLUSION) in Phase 4.
• JevARCClassifier: Defeasible relation stance classification (SUPPORTS, ATTACKS) bypassing heavy LLM prompt envelopes.
• JevRelationReranker: Drop-in replacement for CrossEncoder relation scoring in Phase 3 Global Relations.
• Evaluates throughput speedups and cost reductions on book-length corpora against frontier generative LLM baselines.

Horizon 3: Epistemic Frontiers & Scientific Research¶

Long-Term | Under Consideration / Research

Explores cutting-edge intersections of structuralist philosophy of science, geometric deep learning, and neuro-symbolic reasoning. Anchored in Pan et al. (2024) and related literature. Detailed analysis available in Literature & Future Directions.

Hypercomplex & Quaternion Embeddings (\(\mathbb{H}^d\))¶

Standard Euclidean embeddings struggle to represent complex non-commutative relational structures without large parameter inflation. Using 4D hypercomplex space (\(\mathbb{H}^d\), Hamilton products \(q = a + b\mathbf{i} + c\mathbf{j} + d\mathbf{k}\)), word-, sentence-, and document-level representations can be unified with graph structure embeddings to perform expressive multi-scale relational rotations in TheoryNet topologies.

Contrastive Learning for Candidate Relation Retrieval¶

Raw LLM text embeddings often capture topical similarity rather than relational truth. Training Siamese encoders via contrastive learning (InfoNCE loss) discriminates plausible from implausible candidate triples, significantly boosting candidate retrieval precision in Phase 3 Global Relation Discovery.

Model-Agnostic Knowledge Loss Functions¶

Incorporates a model-agnostic loss function that leverages LLM textual guidance to score triple plausibility directly in Neo4j without requiring full model fine-tuning.

Reasoning on Graphs & Path Generation (RoG)¶

Neuro-symbolic reasoning over theory graphs via two-stage planning and retrieval: a planning module generates relation paths grounded in the graph as faithful reasoning plans, followed by multi-hop constrained LLM inference.

Cross-Paradigm Theory Synthesis & Incommensurability Analysis¶

Extends TheoreticalEnrichmentRunner to model competing scientific paradigms, formalizing cross-theory constraint edges (\(TS_{\text{edge}}\)), admissible blurs (\(\delta^*\)), and Kuhn-incommensurability boundaries.