Skip to content

732 papers · 1,266 curated relations · one static dataset

The Research Network

The literature under this site is one vault: 2,646 records collected for the AI-operations, governance and distributed-systems work published here. 2,347 paper pairs carry a written relation note — and this page draws the 732 papers those notes connect. Open an edge to read its note; the note names the agent that drafted it, because none of them are claimed to be human-curated.

Across areas, 556 drawn relations touch the foundations layer; the strongest program-to-program link is Applied AI in Operations ↔ Distributed Systems (15 relations). The whole graph is a committed JSON extract with a provenance manifest — no database, no runtime requests, every figure re-runnable from the script that built it.

Loading the full graph…

What is in the dataset

2,646

papers in the vault — 2,636 carry a VERIFIED status

1,266

drawn relations (623 semantic · 643 extracted citations) from 2,347 curated vault links, each with the drafting model attached

3

drawn relations are marked refutes — evidence against a specific finding, not a vibe

13/23

authored records have no curated relation yet. They stay in the index below instead of being quietly dropped

Four questions, four evidence chains

A curated reading path: a problem, then the literature that bears on it, then the production study that answers it with numbers. The hop notes are vault relation notes (agent-drafted); the sequence itself is hand-picked, and no claim is made that consecutive records share a vault edge.

Can you prove your AI system is governed, not just documented?

The ISO/IEC 42001 certification work, read against the auditing and model-reporting literature it stands on. Each hop has a production study with numbers behind it.

  1. Auditing large language models: a three-layered approachMökander et al. · 2023 · peer-reviewed · opens in the graph
  2. Toward Trustworthy AI Development: Mechanisms for Supporting Verifiable ClaimsBrundage et al. · 2020 · preprint · opens in the graph↳ Extends Brundage et al.'s institutional verification mechanisms into a three-layered audit architecture covering governance, model, and application audits for large language models.
  3. Model Cards for Model ReportingMitchell et al. · 2019 · peer-reviewed · opens in the graph↳ Pairs institutional audit trails and red-teaming claims with standardized Model Card performance and Subgroup reporting artifacts.
  4. Datasheets for DatasetsGebru et al. · 2021 · peer-reviewed · opens in the graph↳ Model cards document models; datasheets document datasets; together they cover ISO 42001 A.6.2.7 and A.7 documentation evidence.
  5. Putting an Agent on the Incident Queue: What I Measured Before I Trusted It.Vidigal · 2026 · retracted claim · opens in the graph · retracted claimRead the study →↳ Operationalizes dataset provenance, collection-process disclosure, and failure-mode documentation on live incident-queue telemetry under ISO/IEC 42001 and 5259 controls.

Standards in scope: ISO/IEC 42001 (A.7) · ISO/IEC 5259-1…5 · EN 18286 · EU AI Act Arts. 12 & 14

Study paths: /ai-operations/09-implementing-iso-42001-governance-and-lifecycle/ · /ai-operations/10-iso-42001-annex-a4-a7-tool-sandboxing-and-data-quality/ · /ai-operations/11-iso-42001-clause-8-and-human-oversight/

When an agent says it is confident, should you believe it?

Verbalized-confidence evaluation, prior-shift correction, and the 787-run audit — including the finding we retracted when it failed to replicate. That retraction is on this page, struck through.

  1. Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMsXiong et al. · 2024 · preprint · opens in the graph
  2. Adjusting the Outputs of a Classifier to New a Priori Probabilities: A Simple ProcedureSaerens et al. · 2002 · peer-reviewed · opens in the graph↳ Contrasts uncalibrated verbalized self-confidence elicited from LLMs against likelihood-based EM prior-shift correction on class-imbalanced distributions.
  3. From TextBlob to LLM Agents: Sentiment Model Selection for B2B Technical Support with CSAT Ground TruthVidigal · 2026 · peer-reviewed · opens in the graphPeer-reviewed · ACL 2026 Industry Track↳ Applies exact prior-shift log-odds calibration (-2.9704) from Saerens et al. (2002) to production B2B support sentiment classification with verified CSAT ground truth.
  4. Putting an Agent on the Incident Queue: What I Measured Before I Trusted It.Vidigal · 2026 · retracted claim · opens in the graph · retracted claimRead the study →↳ Pairs calibrated B2B customer-support classification (ACL 2026) with empirical failure-mode auditing of autonomous incident-queue triage.

Standards in scope: EU AI Act Art. 14 (human oversight) · ISO/IEC 42001 Cl. 8

Study paths: /ai-operations/01-agent-on-the-incident-queue/ · /ai-operations/11-iso-42001-clause-8-and-human-oversight/

Can agents read production data without the data leaving?

PII-detection benchmarks and prefix-preserving anonymization, read against a monotonic egress firewall: 99.4% secret recall at 29 ms with a local 4B model, replayed over 323 production tool outputs.

  1. PII-TRACE: A Benchmark for Context-Aware PII Detection in Multi-Turn LLM ConversationsZhang et al. · 2026 · preprint · opens in the graph
  2. Prefix-preserving IP address anonymization: measurement-based security evaluation and a new cryptography-based schemeXu et al. · 2002 · peer-reviewed · opens in the graph↳ Combines structured PII and secret entity detection in telemetry text with Crypto-PAn prefix-preserving IPv4/IPv6 network address anonymization.
  3. Four Frontier Models on the Same Support Queue: What Separates Them, and What Does Not.Vidigal · 2026 · preprint · opens in the graphRead the study →↳ Enforces deterministic source-side redaction and monotonic local-model privacy filtering before support-queue payloads reach external frontier model APIs.

Standards in scope: GDPR · EU AI Act Art. 15(5) · ISO/IEC 42001 A.9.2

Study paths: /ai-operations/12-local-4b-monotonic-privacy-firewall/ · /ai-operations/10-iso-42001-annex-a4-a7-tool-sandboxing-and-data-quality/

When does a fixed-cost GPU beat per-token APIs?

Revolut's published 2026 benchmarks, Jevons' paradox, and rigorous benchmarking methodology, read against our own unit-economics study: high-rejection verification loops at $3,945/yr TCO, 7.8×–242.6× cheaper.

  1. Shipping Like a Startup, Regulated Like a Bank: How Revolut Cut AI Coding Costs 90% With Self-Hosted Open ModelsAbitbol · 2026 · external reference · opens in the graph
  2. Jevons' paradoxAlcott · 2005 · peer-reviewed · opens in the graph↳ Illustrates Jevons' efficiency paradox in enterprise AI: lower per-token inference cost induces high-frequency agentic verification loops that shift the break-even point toward fixed-cost local GPUs.
  3. Scientific benchmarking of parallel computing systemsHoefler & Belli · 2015 · peer-reviewed · opens in the graph↳ Provides the statistical benchmarking discipline (non-parametric quantiles, full-loop cost accounting) needed to measure whether unit-efficiency gains reduce or rebound total system resource consumption.
  4. What The Log: Compile-Time AST Template Mining and Two-Tier Statistical Temporal Compression for Agent-Queryable Parquet Log LakehousesVidigal · 2026 · vault draft · opens in the graph↳ Applies Hoefler & Belli's reproducible systems-benchmarking rules to compile-time AST log template mining and local-vs-cloud telemetry processing.

Study paths: /ai-operations/13-unit-economics-of-local-inference-loops/ · /ai-operations/16-lost-in-the-middle-and-the-log-pareto-frontier/ · /ai-operations/04-prompt-caching-on-vertex/

Open bets — where this graph goes next

Refuted, out loud

3 drawn relations dispute a specific finding. Keeping them is the point: a graph without contradictions is decoration.

  • refutes Guidelines for assessing the accuracy of log message template identification techniques → Tools and Benchmarks for Automated Log Parsing

    Proves that Grouping Accuracy (GA) in Zhu et al. (ICSE-SEIP 2019) masks severe template-identification errors and introduces message-level Parsing Accuracy (PA) and template-level FTA.

  • refutes Agile Effort Estimation: Have We Solved the Problem Yet? Insights From a Replication Study → A Deep Learning Model for Estimating Story Points

    Replication finds Deep-SE story-point estimates do not reliably beat naive baselines.

  • refutes Putting an Agent on the Incident Queue: What I Measured Before I Trusted It. → Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs

    Across n=368 incidents and 787 production runs, self-reported verbalized LLM confidence carries no usable discriminative signal (AUROC 0.460, p=0.315; and 0.518, p=0.661, 95% CI 0.447-0.580), refuting reliance on verbalized confidence for operational gating.

Research areas

Applied AI in Operations

Agents, retrieval, and verification gates on live production workloads. One program, deliberately kept whole: the study series publishes here, so the count is high and the area card says exactly how much of it is drawn. 1129 papers in the vault, 165 drawn here.

468 curated relations touch this area · 4 authored papers drawn

Open Applied AI in Ops →

Calibration & Conformal Prediction

When a model's confidence is allowed to mean something: prior shift, selective prediction, risk-controlled decisions. 60 papers in the vault, 10 drawn here.

25 curated relations touch this area · 0 authored papers drawn

Open Calibration →

AI Governance & Security

ISO/IEC 42001 and 5259 controls, EU AI Act obligations, privacy firewalls, sandboxed agent tools. 85 papers in the vault, 32 drawn here.

64 curated relations touch this area · 2 authored papers drawn

Open Governance →

Distributed Systems

The database floor: replication, LSM engines, grey failures, fleet upgrades under measurement. 76 papers in the vault, 23 drawn here.

78 curated relations touch this area · 4 authored papers drawn

Open Distributed Systems →

Econometrics & Causal Inference

Deconfounding and causal discovery — separating what correlates from what an intervention changes. 26 papers in the vault, 13 drawn here.

43 curated relations touch this area · 0 authored papers drawn

Open Causal Inference →

Sports Analytics

Public-data football studies: the proving ground where the site's method was formed. 15 papers in the vault, 9 drawn here.

6 curated relations touch this area · 0 authored papers drawn

Open Sports Analytics →

Maritime Emissions

AIS telemetry and EU maritime regulation (ETS, FuelEU Maritime) behind the live dashboard.

5 curated relations touch this area · 0 authored papers drawn

Open Maritime →

Energy & Grids

Grid economics and the energy cost of computation — the counterpart ledger.

2 curated relations touch this area · 0 authored papers drawn

Open Energy →

Graphics & Perception

Graphical perception, visual encoding, and honest data communication: the evidence this page's own design choices follow. 89 papers in the vault, 66 drawn here.

187 curated relations touch this area · 0 authored papers drawn

Open Graphics & Perception →

Statistics & Experimentation

Inference, calibration, survival, experimentation: the statistical standing orders behind every study. 193 papers in the vault, 97 drawn here.

222 curated relations touch this area · 0 authored papers drawn

Open Statistics →

Log & Failure Analysis Classics

Console-log mining, anomaly detection, debugging at scale — the direct ancestry of the AI-operations work. 159 papers in the vault, 77 drawn here.

273 curated relations touch this area · 0 authored papers drawn

Open Log Classics →

Systems & Infrastructure

Storage engines, networking, cloud operations: where distributed-systems research begins. 161 papers in the vault, 86 drawn here.

210 curated relations touch this area · 0 authored papers drawn

Open Systems →

Software Engineering & Human Factors

Modularity, bug reports, performance evaluation, reproducibility, usability: the engineering discipline layer. 140 papers in the vault, 80 drawn here.

159 curated relations touch this area · 0 authored papers drawn

Open Software Eng. →

Foundations — Other Fields

Classics from adjacent fields the keyword rules did not place — medicine, forecasting, domain science. Kept honest by the count. 287 papers in the vault, 67 drawn here.

129 curated relations touch this area · 0 authored papers drawn

Open Other Classics →

Retracted — kept on the page, struck through

Putting an Agent on the Incident Queue: What I Measured Before I Trusted It.

Self-reported confidence was reported as inverted at n=30.

Did not replicate. At n=368 (and across 787 runs) high-confidence answers were slightly more often right, not less: the defensible claim is that self-reported confidence carries no usable signal (AUROC 0.460, p=0.315; and 0.518, p=0.661, 95% CI 0.447–0.580).

What replaced it — the study with the replication →

Authored records — all 23, drawn or not

Status labels are the builder’s honest classification: peer-reviewed is only claimed where a resolvable publisher record exists; future-dated or identifier-less records are drafts, and that is what the badge says.

RecordYearStatusDrawn relations
Putting an Agent on the Incident Queue: What I Measured Before I Trusted It.2026retracted claim7
Four Frontier Models on the Same Support Queue: What Separates Them, and What Does Not.2026preprint2
From TextBlob to LLM Agents: Sentiment Model Selection for B2B Technical Support with CSAT Ground Truth2026peer-reviewed (vault claim)5
AgentSearch-56K: Production Enterprise RAG Multi-Fidelity Evaluation and Langfuse Agent Search Telemetry Dataset2026vault draft0
Value-Budgeted Diagnostic Support Bundles and Source-Side Secret Redaction Across Distributed Database Fleets2026vault draft0
Breaking the Single-Expectation Bottleneck and Stage-Conditional Rank Discordance in Multi-Fidelity Enterprise RAG via Calibrated System-1 Relevance Judges2026vault draft0
FleetCanary-605 & UpgradePaths-68: De-Identified Multi-Tenant Cloud Database Release Canary and Upgrade Regression Benchmark2026vault draft0
Systems Language Trade-Offs in Production Telemetry Collectors: Controlled Evaluation of Go and Rust Across Regex Redaction, Long-Distance Zstandard, and Parquet Generation2026vault draft1
The Prompted-Prior Inversion Law: Deterministic Floor and Calibrated Logit Fusion for Hybrid Diagnostic Engines2026vault draft0
Defense-in-Depth for Forensic MCP Agents: DuckDB AST Sandboxing and Context-Aware Runtime Egress Redaction2026vault draft1
Detecting Behaviour Changes in New YugabyteDB Builds from Aeon Fleet Telemetry: Stage-1 Pre-Registration Protocol for Observational Release Canarying2026vault draft0
Confidence Inversion and Cyclic Ensemble Runoff in Multi-Turn LLM Forensic Root-Cause Analysis2026vault draft0
Most Health Rules Do Not Know Their Fleet: Population-Calibrated Thresholds, Three-Layer Effective Defaults, and Stratified Epidemiology of 212 Database Diagnostic Rules2026vault draft2
Machine-Checked Log Compression: Zero-Axiom Lean 4 Verification and C-ABI Differential Testing for Streaming Log Parsers and Temporal RLE2026vault draft4
What The Log: Compile-Time AST Template Mining and Two-Tier Statistical Temporal Compression for Agent-Queryable Parquet Log Lakehouses2026vault draft10
Pre-Registering the Speedup: Cost-Bounded, Multiset-Equivalent Performance Engineering of a Terabyte-Scale Log-to-Parquet Engine2026vault draft1
YugabyteDB at Scale: Operating Distributed SQL in Production2026vault draft0
Beyond the Alignment Tax and the Invisibility Floor: Calibrated System-1 Decision Models for Rare-Event CSAT and Silent Churn Detection in B2B Technical Supportn.d.vault draft0
Neuro-Symbolic Memory Gatekeeping for Multi-Session Agentic Codingn.d.vault draft0
De-confounding multi-tenant database telemetry: Coarsened-stratum studentized permutation and calibrated System-1 text de-confoundingn.d.vault draft0
Sovereign Fleet Ledgers and Preventable Support Economics: Density-Ratio Conformal vCPU Imputation, Multi-Plane EOL Reconciliation, and Pre-Upgrade Gate Attributionn.d.vault draft0
When Account Concentration Masquerades as a Software Regression: De-Confounding Observational Cloud Fleet Telemetry and Prior-Weighted e-BH Release Canaryingn.d.vault draft1
Sovereign Enterprise Knowledge Graphs in DuckDB: Cross-Family System-1 Adjudication, Statutory Hierarchy Resolution, and Domain-Deduplicated Noisy-OR Pathfindingn.d.vault draft0

What this page does not claim

  • Relation notes are agent-drafted (drafting model on every edge) and reviewed in the study pipelines. Nothing on this page is labelled human-curated. Peer-reviewed status is only shown where a resolvable external identifier exists.
  • The drawn dataset also holds 643 mechanically extracted citation edges (from 1387 publisher reference-list links in the vault); they are hidden by default and can be switched on in paper views.
  • 13 of 23 canonical authored records have no curated relation yet; the authored index lists all of them, drawn or not.
  • 81 curated edges in the vault point at a key the citation table does not contain and were dropped by the builder; the manifest lists the count and every other gate the build applied.
  • Foundations display areas are assigned by published title-keyword rules (they ship in the manifest); misclassification is possible and is not hidden.

Take the data

The graph you are looking at is a static JSON extract of the vault, served from this origin with no runtime requests. Every count on this page is reproducible from these files and scripts/build-research-graph.mjs.

research-graph-full.jsonresearch-graph.manifest.json