Foundations — Other Fields

Dostoevsky shows that standard leveled LSM-trees perform superfluous merging at the largest level (where almost no lookups terminate early and obsolete entries are already bounded by the capacity ratio T), introducing Fluid LSM-Tree (Lazy Leveling) that uses tiering at L1..L_{L-1} and leveling at L_L.

Monkey proves that assigning uniform bits-per-key to Bloom filters across all LSM levels is asymptotically suboptimal because the largest level dominates filter memory while smaller levels dominate false-positive SST probes.

Large preregistered experiment: first-person uncertainty phrases from an LLM ('I'm not sure, but...') reduced over-reliance and raised accuracy on hard questions, with a small trust cost; general-perspective phrasing had weaker effects.

Systematic review of automation bias in clinical decision support: users accept wrong advice (commission) and miss unflagged problems (omission); mitigators include accountability, training and displaying confidence. For human-in-the-loop AI notes: engineer review is a control only if overrides and corrections are recorded.

Adapts Weill & Ross's IT governance framework to information assets, distinguishing data governance (what decisions must be made and who holds the locus of accountability/decision rights) from data management (making and implementing day-to-day decisions).

Consumer-centred data-quality framework (intrinsic, contextual, representational, accessibility). In the lake-chaos tests the failures were contextual (timeliness: stale _derived; completeness: unpinned frames, dropped CE cases) more than intrinsic accuracy, which mostly held.

Normal approximation to Erlang C yielding square-root staffing rules and insight that delay probability depends on safety capacity relative to sqrt(load). Useful for quick what-if staffing tables on a support dashboard, but approximation degrades at very low offered load (small on-call pools).

Objective task-complexity attributes: multiple paths, multiple outcomes, conflicting interdependence, uncertainty about path-outcome links. Supports grading support tickets on objective characteristics rather than perceived difficulty.

Readiness rubric over data, model, infra, monitoring.

Derives a weighted comorbidity index (integer weights 1, 2, 3, 6 from relative risks in a Cox model) on one cohort and validates it on another for 10-year mortality. The method, regression-derived integer points validated out of sample, is the classical precursor of points scorecards such as RiskSLIM and of our ex-ante points card for expected ticket effort.

Google data validation: schema-based checks per batch, alerting on drift, data treated like code. Motivates per-table contracts for silver/_derived (key uniqueness, timestamps inside the bundle window, schema version) run server-side before the gallery export.

Clinical NER gold-standard study: dictionary pre-annotation shortened annotation time without a measurable bias in their setting, but authors caution that pre-annotation can bias annotators toward the pre-labels; motivates keeping AI labels hidden from human raters when humans are the reference for measuring AI agreement.

Hidden debt in ML systems: entanglement (CACE), undeclared consumers, unstable data dependencies, glue code, configuration debt.

KCS Evolve Loop: aggregate reuse and new-vs-known analysis feeds root cause and product improvement back to engineering. Bundle signatures + KB coverage + Jira links are raw material for a ranked known-issue-by-org-reach feed to PM.

Metrics need validated directionality and sensitivity before they steer decisions.

Comparative survey of 13 DQ assessment/improvement methodologies; consolidates dimensions (accuracy, completeness, consistency, and time-related currency/timeliness/volatility) with their metric definitions. Useful as the canonical reference for distinguishing currency (age vs real world) from timeliness (fit for the task) and volatility.

Model-robustness: report the share of models where the estimate is significant and the sign holds, plus the influence of each modelling choice.

Reward systems pay for the measurable proxy and get it, not the goal. In support ops: a bundle-coverage target (activity) rewards asking every customer for a bundle, raising customer effort and time-to-relief, while the goal is faster correct diagnosis. KCS v6 states the same rule: no goals on activity metrics (link rate, coverage);

Framework of task-complexity dimensions (size, variety, ambiguity, relationship, variability, unreliability, novelty, incongruity, action complexity, temporal demand), separating objective from subjective complexity; vocabulary for ex-ante (presented) vs ex-post (realised) ticket complexity.

Extends Wood/Campbell: complexity as the number of pathways through the task's event network, so it grows non-additively with components and links. Reviews over 700 citing studies; nearly all use the Wood/Campbell typology without a network indicator.

fsync T7: sessions cut at the valley of a mixture over log inter-activity times. On 4 cached POC cases the stall-line gap histogram is unimodal (peak ~30 s, chronic cases) and episode counts fall steadily from gap 0 to 60 s with no plateau, while acute EA is one 43.4 s episode at any gap >= 1 s.

Source of the popular formulation of Goodhart's law ("when a measure becomes a target, it ceases to be a good measure") in the context of audit culture. Cited for the risk that dashboard targets (backlog 100, CSAT 90, FRT 30 min) reshape behavior rather than reflect service quality.

Longitudinal test finds no evidence that NPS is a superior predictor of revenue growth versus other satisfaction measures, contradicting Reichheld's 'one number' claim. Lesson for support leadership: do not replace a weak CSAT signal with a single NPS figure;

Mines event logs to show resource service times depend on workload (Yerkes-Dodson-like effects: people work faster under higher load). Implication for ticket effort: a fixed minutes-per-comment prior ignores load; effort per event should be allowed to vary with concurrent queue load, or at least load should be a sensitivity covariate.

NIST AI RMF 1.0 organizes AI risk management into GOVERN, MAP, MEASURE, MANAGE functions and defines trustworthiness characteristics (valid/reliable, safe, secure/resilient, accountable/transparent, explainable, privacy-enhanced, fair). Voluntary, not certifiable. Crosswalks cleanly to ISO 42001: GOVERN ~ clauses 5/7 and A.2-A.3; MAP ~ 6.1/8.2 and A.5;

The multitrait-multimethod matrix. Convergent validity needs agreement across different methods measuring the same trait; method variance shared within a source inflates within-source correlations. For ticket effort, Zendesk, Slack, Jira and PagerDuty are methods.

Customer-base model linking complaints to subsequent defection and spend, and quantifying recovery effectiveness: complaints signal elevated churn risk, and effective recovery offsets much of it. Justifies an account-risk tab that tracks complaint/escalation episodes per account and whether recovery happened, rather than only aggregate satisfaction.

Articulates the RUM Conjecture: every storage access method faces a fundamental 3-way trade-off among Read Amplification (RA), Update/Write Amplification (WA), and Memory/Space Amplification (SA), such that optimizing any two necessarily degrades the third.

Argues diagnostic errors stem mostly from knowledge deficits rather than from fast-thinking biases, and debiasing training has limited effect. Implication for LLM triage: retrieval of the right domain knowledge (known issues, mechanisms) matters more than adding generic 'consider alternatives' instructions.

Independent eval of a hosted decision model on the Ziems et al. CSS annotation tasks vs 19 LLMs. Decision model trails per-task best LLM on 14/15 tasks (median -11.6 macro-F1) at ~44x lower cost. Its confidence is better calibrated than verbalized confidence of 16/19 LLMs, but three frontier models have lower median calibration error (0.066 vs 0.157).

Field experiment with ~1,000 high-school students: unguarded GPT-4 access raised practice performance (+48%) but lowered later unassisted exam scores (-17%); a tutor version with guardrails avoided the harm. For support onboarding: AI notes may raise junior engineers' immediate output while slowing skill acquisition.

Review of retention management: prediction, targeting, and measuring the incremental effect of retention actions.

Formalizes Jevons' paradox: technological increases in resource efficiency lower the effective price of a work unit, causing elastic demand to expand total consumption rather than reducing it.

Demonstrates that LLM accuracy in multiple-choice selection can swing by 13 to 75 percentage points simply from reordering candidate options, driven by positional and token-prior biases. Motivates position-debiased candidate scoring and cyclic/swap-order averaging in single-pass 4B decision gates.

The Nursing Activities Score (NAS) replaced TISS-28 with 23 activity items whose weights were fitted to nursing time measured in a multinational time-and-motion study; it explains about 81% of nursing time versus 43% for TISS-28.

Review linking on-call work to sleep disruption, stress and health effects even when not called. Supports measuring on-call burden (night pages, pages per shift, interruptions per person-week at team level) as an agent-experience metric, especially for regions taking most out-of-hours pages.

Standard reference for process mining. Its organizational perspective (who did what, handover-of-work and working-together social networks) and the distinction between waiting time and service time are the formal basis for reconstructing case-level resource involvement from an event log.

Targeting the highest churn-risk customers can be futile; target those most responsive to intervention (uplift). A support-derived risk list (stalls, repeat P1s) is decision-useful only with an intervention whose effect is measured, e.g. proactive outreach to accounts with recurrent disk stalls.

Return on Quality: quality spending must be justified through satisfaction and retention to revenue; many programmes fail that test. The gallery measures process quality (coverage, latency) but not the retention or revenue link ROQ requires.

Documents how Revolut reduced AI coding and verification inference costs by 90% in a regulated European banking environment by migrating high-volume loops from per-token frontier APIs to self-hosted open-weight models.

Defines task complexity by a priori determinability: how far the worker can know, before starting, the outcome, the process and the information needed. Five task categories from automatic processing to genuine decision tasks. Higher complexity makes people consult more sources and more general ones.

Classic estimate of poor-data-quality cost; practitioner-level evidence (CACM viewpoint), not a controlled study

Card argues the 5 Whys forces a single linear causal chain, depends on the investigator's prior knowledge, is not reproducible across teams, and stops at a convenient 'root'.

LLM optimised for differential diagnosis evaluated on 302 NEJM cases with top-k accuracy and specialist ratings of list quality/appropriateness/comprehensiveness, plus an assisted-clinician arm. Relevant to the triage audit: separates standalone list accuracy from whether the list helps the human, which mirrors helpfulness vs directional correctness.

Taxonomy of freshness: currency (time since source change not reflected) vs timeliness (age of data relative to update frequency), and propagation of freshness through integration pipelines.

Dual-process model of diagnostic reasoning: fast pattern recognition (System 1) for familiar presentations, analytic reasoning (System 2) for unfamiliar ones, with calibration and override between them. Maps to known-issue matching (System 1) versus hypothesis-driven investigation plans (System 2) in DB support triage.

GPT-4 on 70 NEJM CPC cases: correct diagnosis anywhere in the differential 64%, top-ranked 39%, mean Bond quality score 4.2; inter-rater kappa 0.57. Operational lesson for LLM triage audits: report top-1 and any-of-N separately, use an ordinal partial-credit scale, and expect only moderate rater agreement on partial credit, so adjudicate.

Evaluates LLM synthesis of QuickCheck/Hypothesis property-based tests (custom input generators + invariant properties), showing that compiler/execution feedback loops dramatically increase generator validity and edge-case bug discovery.

Field and lab experiments (cafeteria cooks/diners) show transparency works in both directions: customers who see the work value it more, and workers who see the customer exert more effort and produce higher quality. For a support dashboard, an automated 'acknowledged by engineer X' relay is one-way transparency;

B2B evidence that customer-success health scores from usage and service signals predict retention. Supports adding support-derived signals (repeat P1, stalls, CE escalations, EOL exposure) to account health once joined to ARR and renewal dates, which the gallery lacks.

D16as_v5 uncached cap is 384 MBps vs 768 for D32as_v5. With an 8 TiB-class Premium v1 disk (P70, 750 MB/s), NewBlack's D16 VM cap binds before the disk; on D32 the disk binds. NewBlack had S4 stalls on both shapes (42.5 s on D16, 8.9 s on D32), so the VM cap alone does not explain them.

Synthesizes 145 data governance publications (2001-2019) into a 6-dimensional conceptual framework: Governance Mechanisms (structural roles/accountabilities, procedural policies/standards/compliance monitoring, and relational training/shared language), Organizational Scope (intra- vs. inter-organizational), Data Scope (traditional structured vs.

100 diagnostic-error cases: system-related and cognitive factors co-occur; premature closure (stopping at the first plausible diagnosis) is the dominant cognitive failure. For support triage: the key failure to score is a confident single cause that stops the investigation, not merely a wrong list.

Daily-diary study of on-call workers: being on call (even without being called out) is associated with worse well-being and impaired recovery the same day compared with off-call days.

Decomposes long-form generations into atomic facts and scores the share supported by a knowledge source.

The headline result — a small specialized PRM guiding best-of-N search outperforms GPT-4's own critic — is the strongest literature support for System-1 gating of a System-2 generator without the small model ever producing the answer itself.

SERVPERF: performance-only measures predict overall service quality better than SERVQUAL gap scores. For support dashboards: measure delivered performance (resolution, reopen, effort proxies) rather than expectation-gap surveys.

MLOps practice centres on velocity, validation and versioning; ad-hoc deploys are common.

Vector Model of Complexity: patient complexity spans several axes (biologic, socioeconomic, cultural, behavioural, health-system) and is more than comorbidity count. Each axis has magnitude and direction. A later critique says the maths is only metaphorical and a single score may not help.

DECIDE-AI: reporting guideline for early-stage live clinical evaluation of AI decision support; AI-specific items on user characteristics, AI system version, how output is presented, user exposure/modification of AI output, safety/errors and human factors.

SLIM learns sparse linear models with small integer coefficients by integer programming, directly optimising accuracy and sparsity rather than rounding a fitted logistic model. Shows that such scoring systems can match black-box accuracy on many medical tasks while staying usable by hand.

Common-good lens: in technical support, repository use trades learning for speed. A triage note that surfaces the known fix maximises speed for the current ticket; whether engineers learn from it is a separate outcome. Pair known-issue reuse metrics with a learning metric (later unassisted handling of the same signature).

Evaluates Typst (written in Rust by Laurenz Mädje and Martin Haug) as a programmable, incremental typesetting engine. Unlike LaTeX (Turing-complete macro expansion intertwined with single-pass layout) or AsciiDoctor-PDF (Ruby Prawn interpreter), Typst separates pure functional scripting from constraint-based layout memoization (`comemo`), achieving…

Closest methodological template for converting a public activity trace into labor-hours. Edits are grouped into sessions with a 1-hour inter-edit cutoff (vetted with EM-fitted within/between-session inter-edit distributions), session duration is last-minus-first edit plus a pad of 430 s (the mean within-session inter-edit time) to account for work…

Field experiment (telemarketing): AI taking the routine first stage raised creativity and sales, mostly for higher-skilled agents. Operational translation: the bundle note is first-stage triage for skilled engineers; benefit, if any, should show as engineer time-to-hypothesis, not volume, and may concentrate in senior staff.