LLMs Interpret, Embeddings Organize, Graphs Emerge: Agent-Driven Compilation of Scientific Knowledge

1 Introduction

• Scientific work must organize methods, claims, terminology, open questions, and connections across a growing, heterogeneous literature.Scientific agents and retrieval-augmented systems can gather evidence for individual tasks, but repeated workflows also need infrastructure that preserves and reorganizes interpretation for later tasks [7].Source: §1 · p. 1
• The architectural gap is a persistent, source-linked transition whose document-local changes, construction history, and navigation surfaces can be inspected, replayed, and measured.Knowledge graphs, LLM–KG systems, embedding methods, and dynamic-network studies provide established components, but the paper identifies their integration into this transition as the remaining question.Source: §1 · p. 1
• ASKS compiles each source into a human-readable Wiki view and machine-facing semantics, validates a document-local GraphDelta, then uses embeddings and explicit graph rules to update persistent state.Scientific knowledge compilation is defined as a repeated, provenance-preserving transformation of records into revisable representations and persistent relational state; higher-level organization emerges through repeated local transitions.Source: §1 · pp. 1–2
• The study examines formation and persistence across 56 chronologically ordered papers using node reuse, multi-source support, membership churn, Hub birth, persistence, lineage, and portrait coverage.Order reconstruction, cross-domain transfer, and comparison with alternative graph-construction methods are separate robustness and comparative evaluations.Source: §1 · p. 2

2 ASKS: System Model

• ASKS preserves each source while maintaining reusable compiled knowledge, including Wiki pages, graph relations, and Hubs that organize navigation.A separate research state holds open questions and provisional interpretations under review.Source: §2 · p. 2

2.3 Ingest as the State-Transition Boundary

• Each ingest validates source-local semantics, constructs a transient GraphDelta, and fuses the update transactionally, restoring the previous graph if the complete update fails.Mechanical checks, savepoints, origin recording, and snapshots make changes inspectable and support later re-ingest or replay.Source: §2 · p. 2, §2.3 · p. 4

2.1 Source Record and the Two Compiled Surfaces

• The preserved source yields sibling Wiki and graph views: the Wiki supports reading, while the separately stored graph connects source packages, concepts, propositions, people, and Hubs.Both products link back to the source and can evolve under distinct validation contracts.Source: §2.1 · p. 3, Figure 1 · p. 3

2.2 The Three Computational Roles

• LLMs interpret sources into local semantics, embeddings organize integration with accumulated knowledge, and graph organization emerges across repeated ingests.Embedding similarity supports identity disambiguation, proposition alignment, paper routing, node-to-Hub membership, and Hub lifecycle candidates; coded gates and graph rules constrain resulting changes.Source: §2.2 · pp. 3–4

2.4 Authority and Interpretation

• Graph structure and embedding scores support navigation and computational organization, but factual claims remain grounded in source evidence and scientific judgment.Human judgment is concentrated in consequential ambiguity, Hub scope, interpreting splits or merges, and deciding whether emergent organization merits investigation.Source: §2.4 · p. 4

3.1 Encoding: One Source, Two Complementary Outputs

• Each source yields a readable Wiki view and separate machine-facing semantic slots, with distinct validation contracts and the received source retained as the factual reference.The Wiki uses a deterministic skeleton with LLM-written navigation and content; semantic slots are normalized and checked separately before graph construction.Source: §3.1 · p. 4

3.2 GraphDelta: The Document-Local Intermediate Representation

• GraphDelta is an inspectable, document-local proposal that is checked before attachment, with ambiguous identity matches abstaining rather than forcing persistent reuse.Identity resolution prioritizes canonical IDs and aliases before embedding-assisted tests; topical routing separately maps paper positioning to active Hub scopes.Source: §3.2 · p. 5

3.3 Transactional Fusion and Provenance-Preserving Reuse

• Transactional fusion either commits a complete document update or restores the prior graph state, while identical edges accumulate source lineage and compiled state trajectories remain replayable.The page, raw package, and source edge are checked after fusion; hard-check violations trigger rollback. Cross-source support is recorded in edge origin records without duplicating canonical edges.Source: §3.3 · pp. 5–6

3.4 Embedding-Driven Organization

• Embedding geometry guides identity and proposition alignment, Hub membership, and Hub birth, overload, split, and merge, while hysteresis and structural checks constrain reorganization.Proposition alignment may reuse a highly similar claim, relate an intermediate match while retaining both nodes, or leave lower-similarity claims independent; high-consequence Hub changes require an agent’s semantic commitment.Source: §3.4 · p. 6

3.5 Lineage as Structural Memory

• Lineage preserves hierarchy across ingests: hysteresis stabilizes membership, ancestry keeps kin branches distinct, and lineage-preserving merges retain retired Hub history through redirects.A Hub with more than twenty active members first attempts redistribution to existing children; merge analysis traces parent–child ancestry up to three generations.Source: §3.5 · p. 6

3.6 Compilation as a Replayable and Revisable State Transition

• Compilation is an ordered, replayable state transition: sources are interpreted and validated before their visible deltas are fused and the affected graph reorganized.Re-ingest revisits an existing source, while recompilation rebuilds derived state from recorded sources and dependencies; rebuildable membership and lifecycle operations preserve lineage and history.Source: §3.6 · p. 6

4.1 Retrieval over the Same Semantic Geometry

• Semantic retrieval proposes related keywords, propositions, pages, and source regions; graph relations, Hubs, and neighborhoods add complementary structure.Compiled pages also serve as reading indexes: agents can follow leads and headings to stable source locators, making the scope of reading explicit.Source: §4.1 · p. 7

4.2 Support Traces and Provisional Research State

• Support traces link claims to source regions, record dependencies among composite judgments, and capture open questions, conflicts, and explored paths when available.Because traces derive from retrieval and reading operations, they provide process-level provenance rather than only post-hoc attribution.Source: §4.2 · p. 7
• Composite judgments remain provisional, and an explicit construction or review operation can promote a useful hypothesis into a permitted persistent representation.Substrate-level search describes corpus coverage, while broader scientific status draws on external review; query responses retain a record of what was retrieved and read, and which questions remain open.Source: §4.2 · p. 7

5 Graphs Emerge: From 56 Papers to an Author Research Portrait

• Repeated local updates produce persistent higher-level navigation, with node reuse and creation, Hub membership, and hierarchy operating at distinct organizational levels.The graph’s emergent organization is navigation along the compiler’s state trajectory, not a claim that the graph itself establishes scientific novelty.Source: §5 · p. 7

5.1 How Local Updates Become Global Structure

• Semantic and structural affinity guide Hub membership, while entry-versus-retention thresholds and lineage rules provide temporal and structural memory during refinement.Capacity and lineage rules can form Hubs from coherent unassigned regions while preserving organizational history.Source: §5.1 · p. 8

5.2 What We Mean by “Graphs Emerge”

• The demonstration evaluates formation along one frozen chronological trajectory; order robustness, causal decomposition, and domain interpretation are distinct subsequent evaluation layers.The corpus comprises 56 independently published works from one research program, compiled in fixed chronological order from frozen source-local interpretations.Source: §5.2 · p. 8, §5.3 · p. 8, §5.4 · p. 12

5.3 Worked Longitudinal Demonstration

• Rcum(56) = 0.1885, mean R(t) = 0.2043, and final M(t) = 0.06764 show that canonical reuse and multi-source consolidation remain limited as the vocabulary expands.Reuse and consolidation have slightly declining descriptive trends, while cross-source support grows in absolute terms; low membership churn persists throughout the chronological trajectory.Source: §5.3 · p. 9, Figure 2 · p. 10
• The final Hub lineage depicts a tensor-network-centered research tree, with persistent representation and tensor-renormalization branches and later branches into quantum many-body, TN machine learning, and quantum-AI-oriented work.The portrait’s macro-domain labels are post-hoc display interpretations; its graph structure and source-support counts come from the frozen lifecycle record.Source: §5.3 · p. 9, Figure 4 · p. 11
• Psurvive = 1 and Alife = 1 identify a persistent birth-driven tree; Bmulti = 0.889, Lbranch = 0.222, Qmap = 0.893, and C̄ = 0.00467 quantify its support, refinement, coverage, and churn.The run committed 18 Hub births; all remained active at G56, while split, merge, and retirement counts were zero. The mapping covers 50 of 56 papers, with six explicit abstentions.Source: §5.3 · pp. 9–10, Table 2 · p. 10, Figure 3 · p. 11

5.4 What the Demonstration Shows

• The author-corpus map shows structural continuity and topical branching, but Hub organization is a compiler output rather than scientific authority; inspection and interpretation remain anchored in source records.The demonstrated claim is limited to visibility of a longitudinal trajectory in graph state; the underlying records and their assessment retain scientific authority.Source: §5.4 · p. 12

• ASKS preserves source records, validates Wiki and semantic outputs separately, and uses document-local deltas with rollback for hard fusion violations.Reused relations accumulate source origins, research-state objects remain provisional, and idempotence requires repeated operations with the same inputs to produce the same graph state.Source: §6 · p. 12
• Replay binds the source manifest, local model outputs, embedding identifier, configuration, operation versions, and ordered construction calls.The paper compares replay to rerunning an experimental protocol.Source: §6 · p. 12
• The author-corpus demonstration uses frozen local bundles, an isolated clean-state graph, hash-bound Hub decisions, step reports, and snapshot receipts.Supplementary Sections A–E provide the complete engineering record.Source: §6 · p. 12
• Agents automate mechanically checkable, rollback-safe operations, while changes assigning scientific meaning, merging identities, or approving new Hub scope require an explicit gate.Concise summaries expose source-driven map changes and available evidence paths; detailed vector scores and engineering records remain available for audit in the supplement.Source: §6 · p. 12

7 Conclusion

• ASKS implements source-to-graph compilation through source-local LLM encoding, validated document-local GraphDeltas, embedding-based organization, and deterministic graph rules.Accepted state transitions retain source links and construction history for replay and revision.Source: §7 · p. 12
• Across 56 chronologically compiled papers, ASKS yielded a source-traceable tensor-network-centered research portrait, stable low-churn Hub organization, and predominantly additive canonical-node growth.The portrait has explicit mapping coverage and later branches into several application domains; higher-level organization stayed stable along the frozen trajectory.Source: §7 · p. 12
• ASKS treats graph emergence as source-linked organization, while scientific novelty remains a separate review judgment.Compiled views and navigation paths remain traceable to preserved scientific records as graph state evolves.Source: §7 · p. 12

Code and Data Availability

• Ran-ASKS v0.2.0 includes ASKS source and frozen paper artifact 1.0.0, containing 56 sanitized Wiki pages, 18 Hub pages, the final graph, corpus manifest, metrics, portrait tables, validation records, provenance, and checksums.The release distributes compiled and sanitized records, not primary PDFs, parsed Raw text, credentials, or private production knowledge-base state.Source: Code and Data Availability · p. 13
• The software is licensed under PolyForm Noncommercial License 1.0.0, while the frozen paper data and compiled Wiki/graph artifact use CC BY-NC 4.0.Source: Code and Data Availability · p. 13
• Reported results remain bound to v0.2.0 and artifact 1.0.0, although subsequent ASKS development continues on the repository’s main branch.Source: Code and Data Availability · p. 13

A Frozen Demonstration Boundary and Reproducibility Record

• Appendix A is titled “A Frozen Demonstration Boundary and Reproducibility Record.”Source: Supplementary Contents · p. 16

B Frozen Organization Rules

• Appendix B covers frozen organization rules.Source: Supplementary Contents · p. 16

C Complete Ingest State Machine and Transaction Boundary

• Appendix C covers the complete ingest state machine and transaction boundary.Source: Supplementary Contents · p. 16

D Re-Ingest, Recompilation, and Implementation Boundaries

• Appendix D covers re-ingest, recompilation, and implementation boundaries.Source: Supplementary Contents · p. 16

E Replay, Impact Analysis, and Construction Memory

• Appendix E covers replay, impact analysis, and construction memory.Source: Supplementary Contents · p. 16

F Frozen Publication Manifest

• Appendix F is a frozen publication manifest.Source: Supplementary Contents · p. 16

A Frozen Demonstration Boundary and Reproducibility Record

• The frozen corpus comprises 56 independently published works selected from 65 candidates, with three excluded and six related-only records; every included work had a manually verified complete canonical PDF.Formal journal issue year determines chronological grouping, and machine-readable derivatives were hash-bound before semantic compilation.Source: A Frozen Demonstration Boundary and… · p. 16
• Construction froze 56 audited source-local bundles in Phase L before Phase G fused them from empty state G0 in manifest order, grouped by formal publication year.The fusion lock binds the Phase-L run lock, audit reports, bundle hashes, embedding identifier, threshold configuration, graph schema, Hub-gate contract, code hashes, runtime, and isolated write paths.Source: A Frozen Demonstration Boundary and… · p. 16
• All 56 fusion steps and 57 snapshots passed the reported integrity and validation checks, establishing the construction trajectory’s identity and internal consistency within the frozen corpus scope.The 57 snapshot databases passed read-only integrity checks, 170 artifact receipts matched their recorded SHA-256 values, and all 56 step reports passed graph validation.Source: A Frozen Demonstration Boundary and… · p. 16
• Excluding D047 shifts Rcum(56) from 0.18854 to 0.20718, final M from 0.06764 to 0.07663, and final C from 0.01261 to 0.01432, without changing the observed Hub lifecycle.The comparison removes D047’s direct contribution while keeping downstream fusion outcomes fixed.Source: A Frozen Demonstration Boundary and… · p. 16, Figure S1 · p. 17

B Frozen Organization Rules

• The principal numerical organization gates were frozen before Phase G, with identity, lineage, and high-consequence graph writes requiring both a score crossing and additional evidence.These are implementation parameters of the demonstrated compiler.Source: B Frozen Organization Rules · p. 17

C Complete Ingest State Machine and Transaction Boundary

• Each ingest verifies and preprocesses a source, emits separate Wiki and semantic outputs, validates slots, builds a document-local GraphDelta, then fuses it in a savepoint while recording source origins.Fusion applies the planned delta and records new source origins on a single reusable edge.Source: C Complete Ingest State Machine and… · p. 17
• After fusion, the compiler aligns propositions, refreshes Hub membership, applies specificity and overload rules, revisits abbreviations, tests Hub birth, split, or merge under lineage constraints, and finalizes graph validation, reports, indexes, and logs.Hub birth, split, and merge occur downstream of source validation over a provenance-preserving graph.Source: C Complete Ingest State Machine and… · p. 18
• Before fusion, sources must resolve to addressable packages; ambiguous boundary mentions trigger abstention, and hard-postcondition violations roll back the full transaction to the prior state.Exact duplicates are removed, and persistent reuse requires sufficient identity evidence; identical edges reuse canonical topology while adding the new page to the origin record.Source: C Complete Ingest State Machine and… · p. 18

D Re-Ingest, Recompilation, and Implementation Boundaries

• Re-ingest reconstructs source interpretation, while recompilation rebuilds derived state from recorded sources and dependencies without erasing construction history.Hub membership is rebuildable; splitting preserves the parent, and merging preserves the retired Hub through a redirect.Source: D Re-Ingest, Recompilation, and… · p. 18
• During embedding-service outages, lexical evidence resolves identity and current Hub memberships remain stable; active compilation uses local attachment and Hub rules.The historical production full-graph keyword-similarity sweep is separate from the isolated author-corpus trajectory, which uses active compiler outputs.Source: D Re-Ingest, Recompilation, and… · p. 18
• Idempotent re-ingest with identical inputs reproduces the same nodes, edges, origins, and memberships.Changes to the source bundle, model output, embedding identifier, schema, thresholds, or relevant code hashes require a new lock and clean derived state.Source: D Re-Ingest, Recompilation, and… · p. 18
• Production graph governance adapts Semantica’s declarative-constraint, provenance, and temporal-validity patterns [S4] within a single-store boundary.These mechanisms strengthen validation and historical accountability while retaining the Raw–Wiki–graph evidence architecture.Source: D Re-Ingest, Recompilation, and… · p. 18

E Replay, Impact Analysis, and Construction Memory

• Frozen artifacts and isolated snapshots define complete replay boundaries, while selected production sessions can serve as golden replays before system changes.The changes include prompts, models, schemas, and graph code.Source: E Replay, Impact Analysis, and… · pp. 18–19
• An engineering dependency graph links capabilities and contracts to affected code, validation rules, and graph surfaces, enabling impact analysis and targeted regression selection.It is physically separated from the scientific graph, which organizes source-linked knowledge; session logs record execution diagnostics.Source: E Replay, Impact Analysis, and… · p. 19
• Work states carry task context, capabilities provide temporary instructions and policies, and tools execute bounded functions through explicit arguments, guards, and structured results.Their typed entry points distinguish persistent context, composable guidance, and execution; the engineering dependency graph supplies shared discovery and impact context.Source: E Replay, Impact Analysis, and… · p. 19
• Guarded workflows enforce evidence, path, privacy, and transaction constraints, while content-addressed checkpoints reuse completed units only when source, version, model, parameters, and permissions match.Partial units are retried, changed bindings create new runs, and receipts store status, hashes, and compact validation results while credentials stay in runtime configuration.Source: E Replay, Impact Analysis, and… · p. 19
• Decision records preserve why engineering choices changed, while explicit gates bound high-consequence writes even when agents propose candidate actions.Examples include predicate normalization, Hub scope approval, parser changes, and tightened validation rules.Source: E Replay, Impact Analysis, and… · p. 19

F Frozen Publication Manifest

• The frozen manifest fixes the chronological order of 56 works, including within-year ordering and D-label fusion sequence, and records bibliographic DOI corrections.Dates follow the formal-publication convention; a dash marks an empty verified DOI field.Source: F Frozen Publication Manifest · p. 19, Table 5 · p. 20

.com/deepseek-ai/deepseek-harness

• The supplementary references list resources on document extraction, graph infrastructure, models, sparse attention, and embeddings, including entries marked [S3]–[S7].The bibliography also includes the spatiotemporal composability reference [S2].Source: .com/deepseek-ai/deepseek-harness · p. 21