PaperlayerView exact revision on arXiv

Source-linked AI summary

CTIFoundry: An Agent-Native Corpus Scaffold for Cyber Threat Intelligence

Yutong Cheng, Changze Li, Qian Cui, Wei Ding, Lingzhi Wang, Yan Chen, Peng Gao

arXiv:2608.18613v1cs.AIcs.CR

TL;DR

CTI agents still investigate corpora packaged as opaque RAG chunks, leaving aliases, official cross-references, and span-level provenance inaccessible. CTIFoundry materializes those structures and exposes them through typed tools and procedural skills without modifying the agent loop. On CTIConnect, this action-surface swap improves overall F1 by +0.19 to +0.28 across a four-model panel.

  • Problem

    CTI corpora remain packaged as opaque RAG chunks that do not resolve aliases, expose authoritative cross-references, or preserve span-level provenance for agent investigation.

  • Method

    CTIFoundry materializes a typed ontology graph, span-grounded canonical entities, and hybrid retrieval surfaces, then exposes them through typed tools and procedural skills on a fixed agent harness.

  • Results

    +0.19 to +0.28 overall F1 across a four-model, two-provider panel when only the action surface changes on CTIConnect.

  • Takeaways & Limitations

    The ablation indicates that structure enables the right investigation, procedure makes it reliable, and their combination is super-additive.

  • Takeaways & Limitations

    The evaluation compares configurations under one judge, with auditor rescoring indicating a 0.06 resolution for reading MDS gaps.

Abstract

from arXiv · show

Cyber threat intelligence (CTI) is increasingly consumed not by human analysts but by LLM agents that compose multi-step investigations at query time. The harness side of this shift has matured rapidly (planning loops, tool protocols, context management), but the corpus side has not: threat reports and vulnerability databases are still packaged for retrieval-augmented generation, as opaque chunks behind an embedding index. We argue that this substrate, not model capability, is the bottleneck on agentic CTI investigation, and present CTIFoundry, an agent-native corpus scaffold. At build time, CTIFoundry materializes the latent structure of a CTI corpus: a deterministic ontology graph over four authoritative knowledge bases (CVE, CWE, CAPEC, ATT&CK) whose official cross-references become typed, traversable edges; a span-grounded report layer whose canonical, alias-resolved cross-vendor entities index provenance-carrying chunks; and hybrid dense+lexical retrieval surfaces. At query time this structure is exposed through seven typed tools and three procedural skills mounted on a stock open-source agent harness. On the public CTIConnect benchmark, swapping only the action surface lifts the identically-harnessed agent by +0.19 to +0.28 overall F1 across a four-model, two-provider panel: a small model on CTIFoundry surpasses a flagship on the flat substrate, and the gain is not bought with search effort, since on both Claude models the scaffolded agent is more accurate at roughly half the tool calls. An ablation attributes it: typed structure carries the larger share, procedural skills convert structure into discipline, and the two compose super-additively, because skills bind only to structure that exists.

1 INTRODUCTION

CTI investigation requires multi-step alias resolution and linking behavioral descriptions to authoritative taxonomies, but RAG-style corpora discard this structure. CTIFoundry materializes that structure and, under controlled evaluation, improves agent performance by changing only the corpus action surface.

  • 1 INTRODUCTION: CTI investigations must resolve vendor aliases to canonical entities and link behavioral descriptions to authoritative taxonomies through official cross-references.Examples include Lazarus, Hidden Cobra, and APT38, alongside CVE, CWE, CAPEC, and ATT&CK.
  • 1 INTRODUCTION: RAG packaging leaves agents with opaque chunks behind similarity search, sharding aliases, burying cross-references in record blobs, and obscuring derived claims.The substrate inherits these limitations even when more capable agents are deployed in the domain.
  • 1 INTRODUCTION: CTIFoundry materializes latent corpus structure as typed, validated artifacts: a zero-fabrication ontology graph and a span-grounded report layer with exact provenance and typed entity resolution.The graph contains authoritative knowledge-base entries and official cross-reference edges.
  • 1 INTRODUCTION: 0.610 to 0.829 overall F1 for gpt-5.4 and 0.470 to 0.745 for claude-haiku-4-5 after swapping only the action surface under identical experimental settings.Both arms used mini-swe-agent, with identical loop, step budget, temperature, and model; the baseline used bash over flat files.
  • 1 INTRODUCTION: The substrate, rather than agent capability, is the binding constraint because iteration over a flat substrate cannot recover structure discarded by corpus packaging.This conclusion follows from holding the third-party harness fixed while varying only the action surface.

2 BACKGROUND AND MOTIVATION

CTI knowledge combines authoritative, cross-referenced taxonomies with vendor-specific narrative reports, creating challenges that textual similarity alone cannot resolve. CTIFoundry addresses these challenges by materializing structure and provenance, combining retrieval modes, and packaging analyst procedures with the interface.

  • Operational CTI knowledge: Four taxonomies—CVE, CWE, CAPEC, and ATT&CK—form a cross-referenced reference backbone whose typed edges answer many analyst questions authoritatively.Examples include CVE-to-CWE, CWE-to-CAPEC, and mappings to ATT&CK techniques.
  • Operational CTI knowledge: Vendor reports add a narrative layer whose threat actors, malware families, and campaigns use vendor-specific names, complicating cross-vendor intelligence.Canonical aliases are therefore needed to resolve equivalent entities across sources.
  • Benchmark: 1,859 expert-verified questions and 321 report summaries comprise CTIConnect’s nine-task benchmark across entity linking, entity attribution, and multi-document synthesis.The benchmark covers the four knowledge bases and evaluates LLMs with retrieval access in the RAG setting.
  • Design requirements: CTIFoundry’s design materializes aliases and official references as typed, traversable structure; grounds assertions in exact vendor-attributed spans; and combines dense and lexical retrieval.These correspond to the scaffold’s deterministic ontology graph, canonical entity layer, span-grounded chunks, and hybrid dense+BM25 surfaces.
  • Design requirements: Per-task-family procedural skills accompany self-described typed tools to guide tool ordering, edge prioritization, and candidate verification.The build-time layers address structure, provenance, and retrieval, while the query-time skill layer supplies analyst procedure.

3 CTIFOUNDRY

CTIFoundry is an agent-native scaffold that derives a validated, indexed corpus representation and action surface without modifying the fixed agent loop. It combines typed ontology and report structure, provenance-carrying chunks, hybrid retrieval, seven typed tools, and three procedural skills.

  • Scaffold definition: CTIFoundry derives a validated, indexed representation and action surface over a CTI corpus while keeping the agent loop fixed.The scaffold is designed to maximize investigation accuracy without modifying the agent.
  • Corpus and ontology: The scaffold represents four authoritative CTI taxonomies as an ontology graph with typed official cross-reference edges.The corpus combines vendor reports with CVE, CWE, CAPEC, and ATT&CK records carrying canonical identifiers and cross-references.
  • Report layer: Its report layer indexes provenance-carrying document spans and canonical, alias-resolved entities linked to their supporting chunks.The derived representation includes character offsets, vendor-attributed alias sets, optional knowledge-base grounding, and an entity-to-chunk index.
  • Retrieval surfaces: Retrieval uses dense and lexical indexes, with paraphrase matching delegated to embeddings and knowledge-base search combining dense and BM25 rankings.The retrieval surface also supports caller-supplied constraints such as must terms and conjunctive filters.
  • Query-time surface: Seven typed tools expose distinct scaffold capabilities, while three procedural skills encode task-specific investigative discipline.The tools cover resolution, record fetching, ontology traversal, search, chunk search, entity-indexed collection, and document reading; skills constrain tool use through prescribed procedures.

4 EVALUATION

On CTIConnect, CTIFoundry improves an identically harnessed agent through validated typed structure and procedural skills, raising accuracy across models while reducing search effort. The gains concentrate where authoritative structure exists and combine super-additively.

  • Evaluation setup: CTIConnect evaluates build soundness, accuracy, search effort, and the separate contributions of typed structure and procedural skills across 1,859 expert-verified questions.The main set contains 691 controlled-comparison questions, while the scale set contains 1,168 cost and volume questions.
  • Build validity: 6,044 ontology nodes, 7,097 official edges, 688 provenance-carrying chunks, and 4,868 canonical entities were materialized, with zero validator violations.The validator checks fabricated identifiers, orphan edges, and byte-exact span violations.
  • Accuracy: +0.219 overall F1 for gpt-5.4 (0.610→0.829), +0.190 for gpt-5.4-mini, +0.222 for claude-sonnet-4-6, and +0.275 for claude-haiku-4-5.The substrate swap uses identical harness, model, budget, and temperature within each comparison.
  • Search effort: 12.33→5.09 calls per question for claude-haiku-4-5 and 8.75→4.39 for claude-sonnet-4-6, while accuracy rises by +0.275 and +0.222 F1, respectively.These reductions show that the scaffold improves accuracy without requiring greater search effort.
  • Ablation: +0.062 from procedure alone, +0.136 from structure alone, and +0.219 together (0.610→0.746→0.829), demonstrating super-additive composition.Procedural skills transfer discipline to bash, but their prescriptions depend on actions available through the typed surface.

5 RELATED WORK

Related work has pursued structured CTI through indicator mining, behavior extraction, LLM-based construction, and taxonomy graphs, while agent research has advanced from reason–act loops to deployable and self-evolving harnesses. CTI structuring is characterized as producing a static analytic asset because extraction remains the endpoint.

  • Structuring CTI: CTI structuring spans indicator mining, bespoke behavior extraction, LLM-based construction, and unified taxonomy graphs.These approaches treat extraction as the endpoint, yielding a static analytic asset.
  • Agent harnesses and scaffolds: Agent harnesses have progressed from reason–act loops and learned tool invocation to purpose-built interfaces, deployable skills, and self-evolving agents.The cited work includes interfaces, skills, and agents that rewrite those skills for themselves.

6 CONCLUSION … D The Seven Typed Tools

CTIFoundry is presented as an agent-native scaffold that materializes traversable CTI structure at build time and exposes it through typed tools and procedural skills. The provided material also documents the work’s public-data scope, reproducibility provisions, and appendix contents.

  • 6 CONCLUSION: CTIFoundry materializes authoritative cross-reference edges, canonical cross-vendor entities with span-level provenance, and dual dense+lexical retrieval surfaces for multi-step CTI investigation.These structures are exposed through typed tools and procedural skills on a stock agent harness.
  • AI USE STATEMENT: The AI use statement reports generative AI use as an object of study, a system component, and support for writing and implementation.The models named in §4.1 served as build-pipeline operators and evaluated agents; assistance covered copy-editing, LaTeX formatting, and coding.
  • ETHICS STATEMENT: The ethics statement limits the study to public CTI taxonomies and public CTIConnect vendor-report summaries, excluding human subjects, private data, and proprietary victim data.It also states that CTIFoundry reorganizes public defensive reference material without creating new attack capability.
  • REPRODUCIBILITY STATEMENT: The reproducibility statement identifies the scaffold construction, validation invariants, seven typed tools, three procedural skills, evaluation protocol, and shared experimental conditions.The protocol covers the mini-swe-agent harness, step budget, temperature, model versions, and scoring for nine CTIConnect tasks.
  • A The Build Pipeline in Detail: The appendix contents list a report-layer section organized into four stages.The listed section is A.1 Report Layer: the Four Stages.
  • A The Build Pipeline in Detail: The appendix contents include a section on the embedding model and indexes.This appears as A.2 Embedding Model and Indexes.
  • A The Build Pipeline in Detail: The appendix contents include a section on deterministic-first entity resolution.This appears as A.3 Deterministic-First Entity Resolution.

E Additional Experimental Results … D THE SEVEN TYPED TOOLS

CTIFoundry’s frozen build materializes provenance-grounded report structure, deterministic entity resolution, and typed knowledge-base links for agentic CTI investigation. The appendices document this pipeline, its benchmark motivation, retrieval surfaces, prompts, skills, and extended experiments.

  • A THE BUILD PIPELINE IN DETAIL: The frozen pipeline runs once over 321 vendor reports with gpt-5.4-mini, producing byte-identical artifacts for every query-time configuration.This build-time freeze separates corpus construction from query-time experimentation.
  • A.1 REPORT LAYER: THE FOUR STAGES: 688 chunks preserve clause boundaries, 200–900-character granularity, and exact source offsets as the shared units of content and provenance.The chunk size balances navigable evidence grounding against the coarseness of whole reports.
  • A.1 REPORT LAYER: THE FOUR STAGES: 4,868 canonical entities result from deterministic-first merging over grounded identifiers, TTPs, normalized forms, and within-chunk coreference across eight typed mention classes.The report layer uses STIX-aligned types including actors, malware, tools, techniques, vulnerabilities, campaigns, identities, and indicators.
  • A.2 EMBEDDING MODEL AND INDEXES: Hybrid embedding indexes delegate paraphrase matching to text-embedding-3-large while indexing report chunks and each knowledge base separately.The embedding model is disk-cached, complementing the scaffold’s typed and lexical structure.
  • A.3 DETERMINISTIC-FIRST ENTITY RESOLUTION: Only exact, span-verified alias evidence enters entity unions, while weak possible matches remain visible soft links rather than being merged.This invariant prevents transitive mega-clusters and preserves uncertainty for the agent.
  • B THE CTI ECOSYSTEM: The corpus combines four cross-referencing authoritative taxonomies with vendor reports whose equivalent entities often have three or more names, creating cross-source semantic gaps.The taxonomies are CVE, CWE, CAPEC, and MITRE ATT&CK; reports provide the narrative layer.
  • D THE SEVEN TYPED TOOLS: The seven typed tools are designed around non-overlap, self-description, and structure before similarity, with capability and cost classes exposed in each tool description.The exact tool descriptions are reproduced in the prompt-template appendix.

E ADDITIONAL EXPERIMENTAL RESULTS · E.1 RQ1: EXTRACTION QUALITY ACROSS OPERATOR MODELS · E.2 SCALABILITY AT 1.7× THE QUESTION VOLUME

Section E evaluates extraction quality across six operator models using typed-mention and TTP-grounding metrics, then tests whether CTIFoundry accuracy or cost degrades when question volume increases to 1.7×.

  • E.1 RQ1: EXTRACTION QUALITY ACROSS OPERATOR MODELS: E.1 evaluates six extraction models against a stratified human-audited gold set of approximately 60 chunks covering all eight entity types.A type error counts as both a miss and a false positive.
  • E.1 RQ1: EXTRACTION QUALITY ACROSS OPERATOR MODELS: E.1 measures typed-mention extraction quality and TTP grounding quality as separate metrics.The TTP set includes groundable behavior statements with explicit T-ids plus approximately 40 vague or defender-side phrases requiring abstention.
  • E.1 RQ1: EXTRACTION QUALITY ACROSS OPERATOR MODELS: E.1 defines TTP-grounding evaluation on gold chunks whose explicit T-ids are stripped before extraction.The augmented phrases test whether operators correctly abstain when no TTP should be returned.
  • E.1 RQ1: EXTRACTION QUALITY ACROSS OPERATOR MODELS: The action surface shipped to the agent exposes seven typed tools, each corresponding to a non-overlapping scaffold capability.The tools include usage guidance and cost information.
  • E.2 SCALABILITY AT 1.7× THE QUESTION VOLUME: E.2 evaluates CTIFoundry accuracy on a scale set with gpt-5.4, while the base arm is not re-run at scale.The scale-set result is therefore a CTIFoundry-only measurement rather than a controlled base-versus-scaffold comparison.
  • E.2 SCALABILITY AT 1.7× THE QUESTION VOLUME: At 1.7× the question volume, E.2 measures whether CTIFoundry accuracy or cost degrades.The controlled comparison is established on the main set, with scalability evaluated only for CTIFoundry.

E.3 RQ4: SCALABILITY … F.2 FROM RETRIEVAL PIPELINES TO AGENTIC SEARCH

At scale, CTIFoundry preserves cost and accuracy while its qualitative case study shows that typed operations—not additional search effort—resolve an entity-linking failure. The related work positions CTIFoundry as a consumption-side design that complements extraction, static knowledge graphs, and retrieval or agentic-search systems.

  • E.3 RQ4: SCALABILITY: 1.7× question volume preserves linear cost, ≈2.6 cents and ≈7 seconds per investigation, while forward EL remains at its main-set ceiling.Reported forward EL values are RCM 0.988, ATD 0.957, and ESD 0.834.
  • E.4 A QUALITATIVE CASE STUDY: Both arms use four calls and ≈8 seconds on entity-linking item rcm-005, indicating that the gap is direction rather than effort.The figure compares both arms on gpt-5.4.
  • E.5 RQ6: A QUALITATIVE CASE STUDY: The base agent finds the source CVE immediately but fails because the flat substrate lacks an operation for using it, committing to a wrong answer on a real authoritative edge.The case study states that semantic similarity to the target does not identify which official relationship to use, though the supplied passage ends before completing that claim.
  • F.1 CTI KNOWLEDGE EXTRACTION AND REPRESENTATION: Prior CTI extraction evolved from shallow indicators to behavioral and campaign-stage representations, while much of the resulting structure served human inspection or downstream classification.The passage names iACE, TTPDrill, ChainSmith, and EXTRACTOR across the first two generations.
  • F.1 CTI KNOWLEDGE EXTRACTION AND REPRESENTATION: CTIFoundry complements extraction research by taking extraction quality as input and contributing the consumption-side design for agents.Its report layer extracts typed entities and groundings but no relational triples because the consumer reads provenance text rather than reasoning over extracted edges.
  • F.1 CTI KNOWLEDGE EXTRACTION AND REPRESENTATION: Unified CVE, CWE, CAPEC, and ATT&CK graphs such as BRON are characterized as static analytic assets consumed through human queries outside retrieval or generation loops.The supplied passage also mentions STIX and follow-on work densifying mappings, but ends before completing its comparison.
  • F.2 FROM RETRIEVAL PIPELINES TO AGENTIC SEARCH: Classic RAG retrieves top-k chunks by embedding similarity, while later pipelines interpose derived structure or fuse lexical and dense evidence; CTIFoundry adopts the latter for its KB surface.The passage names GraphRAG, LightRAG, RAPTOR, HippoRAG, and the lexical-dense fusion line.

F.3 LLM-AUGMENTED DATA MANAGEMENT AND KNOWLEDGE-BASE CONSTRUCTION

CTIFoundry applies LLM-augmented data-management and knowledge-base-construction methods to build an offline CTI substrate rather than answer queries directly. Its construction emphasizes provenance, deterministic safeguards, deterministic-first entity resolution, and zero-fabrication validation.

  • Related work: Semantic-operator systems move LLM operators into data-processing loops, optimizing individual operators or pipelines for accuracy, cost, or document-processing quality.LOTUS provides per-operator optimization under accuracy guarantees; Palimpzest and Abacus frame plan selection as cost–quality optimization; DocETL rewrites and validates document pipelines agentically.
  • Related work: ZenDB builds semantic indexes for document analytics, while LLM agents for data-management tasks are emerging as an architectural paradigm.These systems extend the database community’s response to unstructured corpora and connect to the older discipline of knowledge-base construction from dark data.
  • CTIFoundry: CTIFoundry runs once offline to produce a substrate with span-grounded provenance, deterministic guards, deterministic-first entity resolution, and a blocking validator for zero fabrication.Its safeguards include identifier regexes, snapshot validation, Algorithm 1, and enforcement of the zero-fabrication invariant.

F.4 AGENT INTERFACES, SKILLS, AND CONTEXT ENGINEERING

Prior agent-interface and skill research treats the environment as given, whereas CTIFoundry binds both to a validated corpus-side data substrate. Controlled experiments show that tools and skills compose super-additively, while skill adherence depends on prompt placement and model size.

  • Prior Work: Prior work shows purpose-built agent–computer interfaces can outperform raw shells at a fixed model, while MCP standardizes tool protocols.This methodology builds on reason–act loops, learned tool invocation, MCP, and SWE-agent’s repository interface.
  • Measured Composition: +0.219 is the combined gain from skills and tools, exceeding skills alone (+0.062) and tools alone (+0.136).The result supports the claim that procedural advice binds only to structure that exists in the corpus.
  • Skill Deployment: Smaller models under-follow identical skill text in the system prompt but follow it reliably when placed in the user turn.This is a deployment finding from the same experiments and concerns skill engineering rather than corpus structure.

F.5 CTI AND SECURITY BENCHMARKS ON LLM · G PROMPT TEMPLATES

CTIConnect provides the paper’s corpus-grounded, expert-verified benchmark setting, while CTIFoundry reframes its cross-source bottleneck as a corpus-substrate problem. The appendices specify the prompts and validation procedures that operationalize this comparison, from typed-tool versus flat-file agents to span-grounded extraction and claim-coverage judging.

  • F.5 CTI AND SECURITY BENCHMARKS ON LLM: CTIConnect spans entity linking, attribution, and multi-document synthesis, and exposes chunk-and-embed RAG’s weakness on cross-source tasks.The benchmark is adopted with its released corpus and tasks unchanged, while every baseline is re-measured under a fixed harness.
  • G.1 AGENT SYSTEM PROMPTS: THE TWO ARMS: The benchmark comparison holds the harness, model, step budget, and temperature constant, varying only the system prompt and action surface.The CTIFoundry arm uses seven typed tools, whereas the base arm uses one bash tool over flat corpus files.
  • G.1 AGENT SYSTEM PROMPTS: THE TWO ARMS: The CTIFoundry system prompt instructs agents to resolve canonical entities first, prefer authoritative structure, state exact identifiers, and submit a final answer.The base arm instead investigates a flat corpus through one bash command per turn and uses an explicit completion marker before submission.
  • G.2 BUILD TIME: SPAN-GROUNDED EXTRACTION: Span-grounded extraction requires every emitted surface to be a verbatim substring of its attributed span, enabling validator-checkable character-offset provenance.The rules also require completeness, explicit span-asserted relations, exhaustive alias capture, and omission of unanchored claims.
  • G.3 BUILD TIME: TTP EXTRACTION: ATT&CK extraction restricts outputs to attacker behaviors, excludes defenses and generic outcomes, permits concise outcome-phrased behaviors, and forbids invented identifiers.Validation checks every emitted identifier against the materialized ATT&CK node set, supporting a zero-fabrication invariant.
  • G.4 BUILD TIME: ENTITY-RESOLUTION ADJUDICATION: Entity-resolution adjudication keeps near-miss and shared-sponsor actors distinct, accepting authoritative naming relationships through a three-valued kind field.Deterministic signals resolve most mentions; the judge handles the residue and distinguishes exact, claimed-same, and possibly-same relationships.
  • G.5 BUILD TIME: TRIPLE VALIDATION: Triple validation admits only a single evidence sentence and scores predicate explicitness, entity scope, and CTI semantic validity, with type-correction fields for repair.This separates asserted relations from co-occurrence while preserving edges whose entity typing can be corrected.
  • G.6 EVALUATION: MULTI-DOCUMENT SYNTHESIS JUDGE: The MDS judge decomposes reference and prediction answers into atomic claims, uses many-to-many semantically tolerant coverage matching, and applies the same procedure to both arms.Its measured resolution is 0.06, below which an MDS gap is not treated as real.

H PROCEDURAL SKILL PLAYBOOKS … H.5 MULTI-DOCUMENT SYNTHESIS (CSC, TAP, MLA)

The playbooks statically route CTI tasks to specialized procedures that exploit authoritative edges, verified retrieval, strict identifier output, and cross-report synthesis. Together, they impose task-specific discipline for linking, attribution, and multi-document coverage.

  • H PROCEDURAL SKILL PLAYBOOKS: Five markdown playbooks statically route rcm, atd, esd, wim, ata, vca, csc, tap, and mla to specialized procedures without model calls or per-question adaptation.The w/o skill ablation removes these files, while the w/o tools arm retains them but makes their prescriptions inert over bash.
  • H.1 ENTITY LINKING (RCM, ATD, ESD): Entity linking searches the source taxonomy first, traverses an authoritative cross-reference edge, and verifies the resulting target rather than directly searching the target taxonomy.For reverse weakness-to-CVE questions, candidates are confirmed against descriptions and the CWE’s incoming has_weakness edge.
  • H.2 ENTITY LINKING, REVERSE DIRECTION (WIM): WIM receives a separate playbook because it reverses the family direction: locate the CWE, enumerate its incoming CVE edge, then choose a matching CVE.The scaffold does not approach ceiling on WIM, its one entity-linking subtask with reversed direction.
  • H.3 ATTRIBUTION TO ATT&CK TECHNIQUES (ATA): ATA decomposes narratives into atomic behaviors, searches ATT&CK in its own terminology, and emits exactly one T-id because an extra identifier costs precision.The correct technique may rank 5-10 rather than in the top three, so the playbook prioritizes precise grounding over recall.
  • H.4 ATTRIBUTION TO CWE WEAKNESSES (VCA): VCA decomposes vulnerability narratives, searches only CWE in canonical terminology, validates candidates, and emits only the best-supported CWE identifier.It rejects ATT&CK identifiers and requires commitment even when the passage is too thin for confident grounding.
  • H.5 MULTI-DOCUMENT SYNTHESIS (CSC, TAP, MLA): One synthesis playbook serves CSC, TAP, and MLA, emphasizing coverage: the MDS agent relies on read report and the entity index, never ontology tools.This matches the per-tool profile reported in Figure 2b.
  • H.5 MULTI-DOCUMENT SYNTHESIS (CSC, TAP, MLA): For multi-report synthesis, the procedure reads every listed report, resolves cross-vendor aliases, preserves stated dates, attributes claims, and reports disagreements without averaging.Its answer format uses compact bullets addressing exactly the requested aspects.
  • H.5 MULTI-DOCUMENT SYNTHESIS (CSC, TAP, MLA): Synthesis decision rules collect all reports for exhaustive questions, use earliest stated dates for timelines, resolve variants for lineage, and group claims by vendor for corroboration.Claims supported by multiple vendors are treated as stronger than single-source claims.
Loading 2608.18613v1…