Research library
Everything we have published, by subject.
124 articles on the problems a system of record has to answer: what counts as one thing, what a value must carry to be defended, how conflicting sources are weighed, and how a generated answer is checked against its evidence. Each entry says in one line what the article covers.
Identity and canonical records
When two records describe one real thing, when they should only be related, and how to measure either decision.
Entity Resolution Evaluation Beyond Pairwise F1
Pairwise against cluster metrics, sampling design and uncertainty for judging entity resolution, worked through an eight-record example.
Merge or Relate? A Framework for Organization Records
A decision tree for when organization records share one identity and when they need a typed link instead, tested on five synthetic cases.
Provenance, time and declared unknowns
What a value must carry to be defended later: its source, its clocks, and an explicit statement when nothing was found.
Minimum Data Provenance Requirements for Defensible Values
Fifteen minimum provenance fields and five audit questions, mapped to PROV-O, with temporal rules, declared unknowns and validation tests.
W3C PROV for LLM Context Provenance: A Testable Profile
An application profile of W3C PROV for reconstructing the evidence behind LLM context, with mandatory mappings, source-state fields and mutation tests.
Valid Time vs Transaction Time: Which Clock Proves What?
Valid, transaction, observation and publication time distinguished through a delayed correction and four as-of tests.
Bitemporal Data vs Temporal Knowledge Graphs: A Guide
Bitemporal records against temporal knowledge-graph research: snapshot replay versus link prediction, their benchmarks, and when to combine them.
Open World Assumption: Why Unknown Is Not False
Open-world negation, explicit negative knowledge and five epistemic states, with the evaluation protocols that keep an unknown from being scored as false.
Ontologies and data contracts
Testing a shared vocabulary the way code is tested: against the questions it must answer and the consumers it must not break.
Competency Questions: From Business Need to Ontology Test
Competency questions treated as traceability contracts: five question types, each tested in SPARQL and SHACL against fixtures with expected results.
Ontology Crosswalk Acceptance Tests for Mapping Quality
Acceptance tests for a crosswalk between ontologies: predicate semantics, inference closure, consumer queries, evidence packets and version drift.
Safe Ontology Releases Beyond SHACL Conformance
A release gate for ontology owners that goes past SHACL conformance to import closure, query and entailment regression, migration evidence and rollback.
When a Data Contract Needs an Ontology: ODCS 3.2 Test
An acceptance test for choosing between a plain ODCS 3.2 contract, linked definitions, a taxonomy or an ontology, with a worked field crosswalk.
Truth discovery and knowledge fusion
Research on resolving conflicting sources: which methods to trust, how to test them, and what their assumptions hide.
How to Test Truth-Discovery Algorithms Before Choosing
Formal property tests, transparent baselines and benchmark design for comparing truth-discovery algorithms before committing to one.
Comparing Truth Discovery Methods Without Gold Labels
Evaluating truth discovery when no ground truth exists: synthetic stress tests, sparse adjudication, observation likelihood and stability.
Confidence-Aware Truth Discovery for Long-Tail Evidence
How to weigh sources that have made few claims, with Wilson intervals, CATD and RTD analysis and source-count diagnostics.
Multiple Sources of Truth: An Assumption Audit
The assumptions behind multi-truth integration models LTM, MBM and DART, compared on worked claim matrices, including what an omission means.
Truth Discovery with Copying Sources: Benchmark Guide
What benchmarks with copying sources actually test, read assumptions first across multi-method studies, sparse labels and metrics.
Testing Streaming Truth Discovery Under Drift
Separating one-pass computation, an evolving truth and drifting source reliability when testing streaming truth discovery.
Fairness Tests for Truth-Discovery Aggregation
An audit protocol for truth-discovery aggregation across accuracy, disparity, coverage, uncertainty and dataset lineage.
Context engineering and LLM evaluation
Whether a generated answer is supported by its evidence, measured claim by claim rather than by impression.
How to Evaluate LLM Citations at the Claim Level
Evaluating LLM citations by atomic claim, with support and completeness metrics, confidence intervals and perturbation tests.
RAG Fault Isolation: Is the Answer Defensible?
Locating where a retrieval-augmented answer failed across six observable boundaries, with claim-level metrics and controlled evidence tests.
LLM Judge Groundedness Evaluation Calibration Protocol
Calibrating an LLM groundedness judge with controlled contradictions, paired judge-by-generator tests, confidence intervals and human review.
What Does a Fact-Verification Benchmark Test?
What fact-verification benchmarks measure: system boundaries, evidence regimes, metric validity and leakage, and how to choose one.
AI Agent Memory Update: When to Update or Abstain
Testing when an agent's memory should update or abstain, with two clocks, temporal fixtures and stale-answer rates.
Other recent articles
Published since this library was last arranged by subject.
OWL import closure: Can an old graph be reinterpreted?
A 2026 analysis of OWL import closure and historical graph replay, with a two-hop dependency manifest, an entailment comparison, and criteria for version…
Negation in Information Retrieval: Recall vs. Violations
A 2026 research report on negation in information retrieval: NevIR and ExcluIR evidence, recall versus exclusion metrics, and a reproducible collection-level…
Knowledge graph fact verification: a controlled evidence test
A 2026 research report on knowledge graph fact verification, comparing FactKG studies and specifying paired edge-removal controls, evidence audits,…
Wikidata Statements, References, Ranks and Qualifiers
How Wikidata lets conflicting claims coexist: statements, snaks, qualifiers, references and ranks, a real example in JSON and RDF, and enterprise lessons.
W3C PROV Explained: Entities, Activities and Agents
A worked introduction to W3C PROV: the three core types, the core relations, one value traced in PROV-N and PROV-O, and what PROV validity does not prove.
SEC Rule 17a-4: WORM and the Audit-Trail Alternative
How SEC Rule 17a-4(f) lets broker-dealers keep electronic records in WORM form or with a complete time-stamped audit trail, and what each path requires.
Survivorship Rules for Golden Records: Worked Examples
Recency, source priority, completeness and majority vote applied to one fictional supplier record from four sources: what each rule keeps and loses.
Tamper-Evident Logs: Hash Chains, Merkle Trees, CT
How hash chains, Merkle trees and transparency logs make edits to a log detectable, with computed examples, and what integrity proofs cannot tell you.
Why Language Models Hallucinate: Incentives and Abstention
Why LLMs hallucinate: the singleton-rate bound, how accuracy-only grading rewards guessing, where RAG fails, and why unknown must be an allowed answer.
Traceability Evidence: GS1 EPCIS, DSCSA and the EUDR
Three traceability regimes compared by the question each asks of a record: GS1 EPCIS events, US DSCSA package-level tracing, and EUDR due diligence data.
SQL:2011 Temporal Tables in Practice
System-versioned and application-time tables in SQL Server, MariaDB and PostgreSQL: DDL, AS OF queries, retention, and what history does not record.
Sanctions Screening as Entity Resolution: OFAC 50% Rule
Sanctions screening splits into name matching, identity and ownership aggregation. A worked example of the OFAC 50 Percent Rule and the EU approach.
RDF vs Property Graph vs Relational for Knowledge Graphs
RDF, property graph or relational tables for an enterprise knowledge graph, compared on statement provenance, merging, standards and query.
NIST AI RMF Explained: Functions and Data Provenance
NIST AI RMF 1.0 read as an evidence plan: the four functions, the trustworthiness characteristics, where provenance appears, and the records each needs.
NIST AI 600-1: Confabulation and Information Integrity
NIST AI 600-1 lists 12 generative AI risks. How its confabulation, information integrity and value chain actions map to records you can show.
Model Risk for LLM Systems: SR 26-2, SR 11-7, OSFI E-23
SR 26-2 replaced SR 11-7 in April 2026 and excludes generative AI; OSFI E-23 covers AI/ML from May 2027. What that means for LLM and RAG systems in banks.
Model Context Protocol (MCP) Explained
MCP's hosts, clients, servers, tools, resources and prompts under the stateless 2026-07-28 revision, with a JSON-RPC example and its gaps on provenance.
OpenLineage Explained: Runs, Jobs, Datasets and Facets
How the OpenLineage spec models runs, jobs, datasets and facets, with a sample RunEvent, the run cycle, PROV mapping and the limits of job-level lineage.
SOX 404 and IPE: What Auditors Test About Company Reports
How PCAOB AS 1105 and AS 2201 treat information produced by the company: accuracy, completeness, precision, report logic and the controls behind them.
MCP Security: Confused Deputy, Token Passthrough, Poisoning
MCP security against spec 2026-07-28: confused deputy, token passthrough, SSRF, state handles, tool poisoning, trust boundaries, review checklist.
LEI Level 2 Data: Parent Relationships and Who Owns Whom
How GLEIF Level 2 records direct and ultimate accounting-consolidating parents, what reporting exceptions mean, and how to relate entities without merging.
Logging AI Agent Actions: Traces an Auditor Can Replay
A minimum event model for AI agent logs: context items, tool calls, results and human actions, referenced by ID and version, mapped to OpenTelemetry GenAI.
Master Data Management: What a Hub Holds and Cannot Prove
What master data management is, how hubs match and merge records, and why a golden record rarely shows which source said what, when, and who disagreed.
ISO/IEC 42001 Explained: What an AI Management System Is
What ISO/IEC 42001 requires of an AI management system: clause structure, impact assessment, documented information, certification, and audit evidence.
ICH E6(R3) Data Governance: Trial Data Traceability
How ICH E6(R3) section 4 treats data governance: the data life cycle, audit trails, source records, certified copies and computerised systems in trials.
Hallucinated Legal Citations: Lessons from Mata v. Avianca
What the Mata v. Avianca sanctions opinion and two empirical studies show about fake case citations, and three controls that keep them out of a filing.
Delta Lake and Iceberg Time Travel Is Not an Audit Trail
What Delta Lake and Apache Iceberg time travel reproduce, how VACUUM and snapshot expiry bound it, and what to add for actor, reason, source, valid time.
Computer Software Assurance vs CSV: GAMP 5 and Annex 11
How FDA's February 2026 CSA guidance, GAMP 5 and EU Annex 11 scale validation to risk, what evidence data systems need, and where teams over-document.
GDPR Accuracy and LLM Hallucination: What Regulators Said
How GDPR Article 5(1)(d) applies to hallucinated personal data: the EDPB ChatGPT Taskforce, the noyb complaints, ICO guidance, and the controls that help.
Fellegi-Sunter Model: m and u Weights, a Worked Example
A worked Fellegi-Sunter example: m and u probabilities, agreement and disagreement weights, prior, match probability, thresholds and clerical review.
FDA and EMA on AI in Medicines: Credibility and Traceability
FDA's AI credibility framework, context of use and model risk, the EMA reflection paper and the joint 2026 principles, read for data traceability.
Entity Resolution Pipeline: Five Stages, Worked Example
Entity resolution in five stages (standardize, block, compare, classify, cluster): what fails at each, with a worked Fellegi-Sunter scoring example.
Event Sourcing vs Audit Tables vs Change Data Capture
One address change recorded as a domain event, a trigger audit row and a Debezium CDC event: which can reproduce a past state and show who and why.
EU AI Act Dates After the Digital Omnibus (2026/1744)
Which EU AI Act obligations apply when, after Regulation (EU) 2026/1744: high-risk dates, Article 50 marking, GPAI, prohibitions and AI literacy.
EU AI Act Article 50: Transparency and AI Content Marking
EU AI Act Article 50 read clause by clause: chatbot disclosure, machine-readable marking, deepfake labels, and what the 2026 Code of Practice asks for.
EU AI Act Article 12: Record-Keeping and Logging Explained
What Article 12 of the EU AI Act requires high-risk AI systems to log, who keeps the logs under Articles 19 and 26, for how long, and from when.
GDPR Erasure and Rectification in Event-Sourced Systems
How to honour GDPR Articles 16 and 17 in append-only and event-sourced stores: corrections as new events, payload separation, crypto-shredding, backups.
EU AI Act Article 10 and Annex IV: Data Governance
What Article 10 of the EU AI Act requires of training, validation and testing data, how Annex IV documents it, and what changed with the Digital Omnibus.
Context Engineering for AI Agents: A Practical Guide
What context engineering is, how it differs from prompt engineering, and how to budget an agent's context by origin, provenance and keep/compress/drop.
EIOPA AI Opinion: Data Governance and Records for Insurers
What EIOPA's August 2025 Opinion on AI governance asks of EU insurers: scope, data governance, record-keeping, explainability, oversight and third parties.
Data Integrity Guidance Compared: FDA, EMA, MHRA, PIC/S, WHO
Side-by-side comparison of FDA, MHRA, PIC/S, WHO and EU data integrity guidance: definitions, ALCOA lists, audit-trail review, governance, retention.
BCBS 239 Explained: The 14 Risk Data Principles
The 14 BCBS 239 principles for risk data aggregation and reporting, what supervisors found in practice, and the checks a system of record must pass.
Business Records as Evidence: FRE 803(6), 901, 902 and AI
How US federal evidence rules 803(6), 901 and 902(11), (13), (14) apply to database records, hash-verified copies and AI output, with a checklist.
Beneficial Ownership Data: Resolving Who Owns Whom
How FATF, the US CDD rule, OFAC and the EU AML Regulation define beneficial owners, and why entities must be resolved before percentages are computed.
Data Contracts Explained: Producers, Consumers, Evidence
What a data contract holds beyond column types, how ODCS v3.2.0 structures it, what dbt model contracts enforce, and where quality checks fit.
C2PA Content Credentials: Manifests, Claims, Signatures
How C2PA Content Credentials work under spec 2.5: manifests, claims, assertions, hard and soft bindings, trust lists, and what validation proves.
Audit Trail Review in GxP: Capture, Protect, Review
What GxP audit trails must record, how to protect them, and how often to review them, read from 21 CFR 11.10(e), Annex 11, FDA, MHRA, EMA and ICH E6(R3).
Anchoring a Claim to a Passage: Web Annotation Selectors
How W3C Web Annotation selectors (TextQuote, TextPosition, Fragment, Range) and Time State anchor a claim to an exact evidence span that survives edits.
21 CFR Part 11 Explained: Records, Signatures, Audit Trails
A section-by-section reading of FDA's 21 CFR Part 11: scope, closed and open systems, the 11.10(e) audit trail, e-signatures and current FDA guidance.
ALCOA and ALCOA+ Explained: Data Integrity Attributes
The nine ALCOA+ data integrity attributes, as FDA, MHRA, WHO, PIC/S and EMA word them, with typical failures and the checks a record system must pass.
ISO 42001 vs NIST AI RMF vs EU AI Act: One Evidence Base
ISO/IEC 42001, the NIST AI RMF and the EU AI Act compared: what each instrument is, who it binds, the records it asks for, and one evidence register.
Data Provenance vs Data Lineage vs Audit Trail
Lineage, provenance and audit trails answer three different questions about the same value. Definitions, a worked example, and which rules ask for which.
Chain of Custody for Digital Evidence, Collection to Court
How a digital chain of custody works: ISO/IEC 27037 stages, hash values, the custody log, US authentication under FRE 901 and 902(13)-(14), and cloud data.
Best Table Format for LLM Prompts: Does Serialization Matter?
Data report on the best table format for LLM prompts: HTML, Markdown, CSV, JSON, LaTeX and images compared on header scope, row identity and token cost, with…
Query Provenance: Which Source Rows Justify a Derived Value?
A 2026 worked test of query provenance: witness sets for a join, projection and SUM, where, why and how provenance, PROV-O mapping, and an acceptance protocol…
AI Agent Handoff Testing: Did the Evidence Reach the Agent?
A 2026 research report on AI agent handoff testing: source-span obligations, verified fault injection, recipient-visible evidence, paired scoring and…
Cross-Lingual Fact-Checking: Auditing Claims vs. Evidence
Stage-by-stage audit of cross-lingual fact checking in 2026: X-Fact, XFEVER and CONCRETE compared without pooling, matched retrieval budgets, Wilson…
Entity Resolution Blocking Recall: Audit for Hidden Matches
A 2026 acceptance test for entity resolution blocking recall: compare candidate pairs with adjudicated true pairs, compute pairs completeness and reduction…
RAG citation provenance: did generated context become evidence?
A 2026 evaluation report on RAG citation provenance: distinguish HyDE hypotheses and generated prefixes from original evidence, build paired exposure tests,…
When is evidence sufficient for fact checking? Omission tests
A 2026 research protocol for deciding when evidence licenses a fact-checking verdict, using deletion and restoration tests, abstention metrics, source…
Data Provenance Privacy: A Recipient Disclosure Test
A 2026 guide to data provenance privacy: evaluate recipient permissions, redact citation metadata, count exposed graph paths, and decide when evidence is too…
RAG chunking sensitivity: can boundaries reverse answers?
A 2026 research report on RAG chunking sensitivity, with a paired boundary-phase protocol, complete-support annotations, fixed-token comparisons and…
Privacy-Preserving Truth Discovery: A Utility Audit
A 2026 research audit of privacy-preserving truth discovery: protected units, local privacy budgets, source-weight reversals, secure aggregation, and a…
RAG Query Rewriting: Does It Preserve the Question?
Evaluation protocol for RAG query rewriting in 2026: intent contracts, paired qualifier fixtures, constraint preservation vs regression, McNemar tests, and…
Data Fusion Error Propagation Benchmark: Upstream Error Audit
Evaluation protocol and literature critique of data fusion error propagation in MaDI-Bench: 15 and 20 tested cells, the Products missing-record bound, and an…
Cached answer validity test: is a cached answer still valid?
A 2026 reuse-validity test for cached LLM answers: crossed request-by-evidence matrix, false-reuse metrics with clustered intervals, and a decision ledger,…
Entity Resolution Bridge Records: A Cluster Acceptance Test
A 2026 practitioner test for entity resolution bridge records: four acceptance questions, a bridge-removal perturbation test, a sign-off worksheet, and…
LLM conflicting sources: a counterfactual fusion test
A 2026 evidence review of LLM source conflict resolution, reconstructing tabular-fusion F1 results and specifying entity masking, source-dependence tests,…
Conformal RAG Guarantees: A Loss-and-Denominator Audit
A 2026 audit of conformal RAG guarantees: compare set coverage, bounded loss and selective error, inspect SAFER's finite budget, and reproduce a scoped…
Data Provenance Tracking: Whose Claim Is This Evidence?
A 2026 analysis of data provenance tracking across organizations: a six-check handoff test for W3C PROV bundles, a three-case table, and a completeness…
How to resolve conflicting numerical claims: normalize first
A 2026 methods guide to reconciling numerical claims: unit and time alignment, CRH loss functions, versioned MaDI-Bench evidence, and a reproducible fixture…
Entity Resolution Human Review: Which Decisions Need It?
A 2026 guide to entity resolution human review: accept, review, or hold unresolved. Includes a decision card, fixed-capacity harm calculations, reviewer…
AI Agent Memory Deletion Verification: Counterfactual Tests
A 2026 protocol for verifying AI agent memory deletion with never-exposed controls, multi-channel disclosure probes, behavioral residuals, retained utility…
Source selection in data fusion: evaluating marginal evidence
A 2026 evaluation protocol for source selection in data fusion: compare held-out subsets, separate coverage from verified accuracy, test source dependence,…
Temporal knowledge graph benchmarks: A data leakage audit
A 2026 audit of temporal knowledge graph benchmark data leakage: time-stripped overlap, inverse shortcuts, forecasting history, ranking protocols and a…
Permission-aware RAG evaluation: Can you defend the answer?
A 2026 evaluation protocol for reader-eligible RAG evidence, with matched permission cases, claim-level provenance cards, and separate utility, disclosure,…
Knowledge graph truth maintenance: retracting conclusions
A 2026 guide to knowledge graph truth maintenance: test assertion withdrawal, alternative support, inferred-triple retraction, provenance evidence, and…
JSON-LD context changes: When does data meaning change?
A 2026 guide to testing JSON-LD context changes through expansion diffs, frozen remote dependencies, integrity checks, consumer contracts, and versioned…
Knowledge graph identifier versioning after corrections
A 2026 guide to knowledge graph identifier versioning: stable referents, correction responses, retirement and custody transfer, with a lifecycle matrix,…
Constraint-based AI evaluation: When exceptions matter
A 2026 research guide to constraint-based AI evaluation, with independent holdouts, rule-error and exception stress tests, provenance checks, and a worked…
Data Extraction Error Analysis: Source or Extractor?
A 2026 research report on data extraction error analysis: separate source and extractor errors with a provenance worksheet, identical-holdout ablations,…
How to evaluate data sources for a knowledge graph
A 2026 guide to knowledge graph source admission: permission, provenance, versioning, sampled coverage and freshness, with an evidence card and accept, defer…
Truth discovery with hierarchical claims: A specificity test
A 2026 research protocol for hierarchical truth discovery: a five-relation conflict matrix, a disclosed scoring fixture, and tests of specificity, source…
Long context LLM robustness: Does evidence position matter?
A 2026 research report on long context LLM robustness: evidence-position tests, separate answer and citation verdicts, frozen source versions, and paired…
OWL vs SHACL: Entailment or Constraint? A Test Card
OWL vs SHACL in 2026: classify requirements by entailment or graph acceptance, test missing and inconsistent values, and preserve a reproducible packet with…
Multi-Hop RAG Citation Completeness: An Evidence-Set Test
A 2026 evidence-set test for multi-hop RAG citation completeness: distinguish answer correctness, joint support and missing hops, preserve source versions,…
SKOS Concept vs OWL Class: A Modeling Decision Test
A 2026 modeling test for SKOS concepts versus OWL classes, covering hierarchy semantics, punning, RDF fixtures, validation obligations and migration…
Source Reliability by Topic: Testing a Single Trust Score
A 2026 analysis of source reliability by topic: FaitCrowd and DART evidence, global versus pooled models, and a held-out benchmark for error, calibration,…
RAG data poisoning: can poisoned sources become facts?
A 2026 integrity protocol for RAG data poisoning: versioned source fixtures, separate retrieval and citation scores, paired controls, confidence intervals,…
How to Preserve Evidence When a Cited Source Changes
A 2026 guide to preserving value-level source evidence, separating observation and archive times, and rechecking changed, corrected, missing, or blocked…
Knowledge graph completion evaluation protocol
A 2026 acceptance protocol for predicted missing triples, comparing CoDEx, InferWiki, FB15K-237, hard negatives, adjudication, MRR sensitivity, and…
How to Validate Knowledge Graph Embedding Predictions
A 2026 protocol for validating knowledge graph embedding predictions against original source spans, identity and time boundaries, then measuring…
What Does a Signed RDF Export Prove? A Verifier's Guide
A 2026 verifier guide to signed RDF exports: define the graph boundary, reproduce canonicalization and signature checks, inspect issuer authorization, and…
Sybil Attacks in Truth Discovery: A Stress Test
A 2026 research guide to Sybil attacks in truth discovery, comparing copied-source and controlled-identity threats, TDSSA evidence, factorial stress tests,…
RAG evaluation dataset leakage audit as sources change
A 2026 audit guide for RAG evaluation sets: trace question and source lineage, test answer and training exposure, compare fixed with rolling corpora, and…
How to deprecate and replace ontology terms safely
A 2026 decision guide for ontology stewards: test changed referents, choose exact or candidate replacements, publish a migration notice, and measure consumer…
Truth existence in data fusion: when all sources err
A 2026 research guide to no-truth questions in data fusion, TEM source-quality assumptions, SF2013 F1 denominators, and a held-out test for candidate omission…
LLM Context Compression Evidence Loss: An Acceptance Test
A 2026 guide to testing LLM context compression for evidence loss, with paired source bundles, exact-span scoring, citation-grounding data, and release rules…
Knowledge graph quality assessment for decisions
A 2026 guide to assessing whether a knowledge graph can support a named decision, with separate tests for correctness, coverage, freshness, provenance,…
RDF 1.2 provenance: triple terms, reifiers, named graphs
A 2026 migration analysis of RDF 1.2 provenance, including six assertion-safety tests, a 13-quad two-library round trip, and guidance on reifiers versus named…
Google Knowledge Vault: What Survived at Web Scale
A 2026 evidence-led analysis of Google Knowledge Vault, its 1.6 billion unique triples, POPACCU+ error audit, provenance assumptions, and a defensible…
The list is read from the published articles and refreshed hourly, so a new article appears here without waiting for this page to be edited. Articles carry their own sources and dates; where one of them is wrong, write to [email protected] and it will be corrected with the date of the correction.