Software Engineering & Human Factors

Reviews stationary-independent-period-by-period (SIPP) and lagged/pointwise-stationary staffing for time-varying demand, showing when stationary Erlang approximations fail. Relevance: shift hand-offs and event windows (e.g., Black Friday) need time-varying staffing analysis rather than annual-average λ.

Parnas (1972) establishes information hiding as the foundational criterion for decomposing systems into modules: modules must encapsulate design decisions that are likely to change (such as representation formats, byte-scanning algorithms, or storage schemas) rather than steps in a processing flowchart.

Brodie et al. (IBM, ICAC'05) match symptoms from logs/stack traces against a database of known problems to find known issues quickly; precision depends on curated, versioned symptom definitions.

Defines appropriate reliance as trust that is calibrated to the automation's actual capability, with good resolution (it differentiates situations) and specificity (it applies to particular functions). Trust rests on performance, process and purpose information.

Customer-perceived quality (defects/incidents per installed system) is driven as much by deployment time, platform and install size as by code; comparing releases needs exposure normalisation and a release-age / time-since-install axis, otherwise early-adopter and calendar effects masquerade as release quality.

Static-analysis framework that detects ORM performance anti-patterns (one-by-one processing i.e. N+1, excessive data retrieval) in Java/Hibernate code and ranks them by measured performance impact, evaluated on open-source and an industrial system.

Explanations raised acceptance of AI advice whether the AI was right or wrong, so they did not produce complementary human-AI performance beyond the better of the two alone.

Task complexity = component (distinct acts and information cues), coordinative (relations, timing, sequencing between acts) and dynamic (change in those relations over time). Triage-audit mapping: planes and artifacts (component); engineers, hand-offs, CE escalation (coordinative); novelty and new bugs (dynamic).

Requirements should be driven by continuously mined usage and feedback data (implicit telemetry plus explicit feedback), combined, with attention to representativeness. Supports using bundle/ticket evidence as a discovery input, but alongside usage telemetry (call-home), not instead of it.

App-store feedback is frequent but skewed to few users and topics, mixes bug reports with feature requests, and needs classification before it can drive requirements. Analogy for tickets/bundles: dominated by a few heavy orgs and must be classified (bug vs gap vs how-to) before it feeds a roadmap.

Comprehensive survey of mutation testing (IEEE TSE 2011), establishing fault-injection mutation kill rates as the rigorous empirical measure of whether a test or conformance suite actually detects semantic defects injected into production code.

Review integrating automation complacency and automation bias: humans over-rely on reliable automated aids, under-monitor them, and miss the aid's errors; expertise and training do not eliminate it. For an AI 'safe to close' list: a human click-through is not a sufficient guardrail;

PM seat (council 5): Feature-level sentiment mining of app reviews ties feedback to named features. The support analogue maps tickets and IDEA links to named product features so pain can be ranked per feature, next to featadopt usage shares.

Combined computer-interaction monitoring with self-reports for professional developers at several companies: work is highly fragmented (frequent activity switches within minutes), and a large share of the day goes to email, meetings and collaboration rather than coding.

Cognitive forcing functions (asking people to decide before seeing the AI, or adding a wait) reduce over-reliance on wrong AI advice more than explanations do, but people like them least; benefits are largest for people high in need for cognition.

Hard-threshold targets (NHS 4-hour, ambulance 8-minute) induced gaming: reclassification, clock-boundary manipulation, hitting the target and missing the point. A support SLA defined as a one-tap PagerDuty ack is exposed to the same mechanism; counter-measure is a paired substantive-engagement metric and audits of the ack-to-reply gap.

Shadowing study of information workers (analysts, developers, managers): about 3 minutes on an event before switching and roughly 10.5-12 minutes in a 'working sphere' (a thematically connected unit of work) before switching sphere.

Data cascades are delayed and opaque.

Survey of metamorphic testing: metamorphic relations replace a missing oracle by checking relations between outputs of related inputs.

PM seat (council 5): Systematic review of app-review analysis for SE (elicitation, bug/feature classification, prioritisation); flags weak evaluation and representativeness. The same caution applies to support tickets as a roadmap signal.

ITiCSE 2019 working-group landscape review of text-based programming error-message research. Consolidates design guidelines (readability, specificity, actionable fix hints, context, no blame) and shows evidence for enhanced messages is mixed and measurement-dependent (time-to-fix, repeated-error rates, population).

Models over-reliance as a strategic cost-benefit choice. People verify AI output only when verification is cheap relative to the task cost and the stakes. Explanations that make verification easier, along with higher task difficulty or rewards, reduce over-reliance. 'Evidence-link clicked before accept' is a useful verification proxy.

Introduces executable specifications and composable random generators for property-based testing.

Demonstrates empirically that ~95% of regular expressions used in production software in Python, Java, and JavaScript do not require non-regular features and can be evaluated in guaranteed O(N) linear time using Thompson NFA/DFA simulation.

Concurrent-session (chat) staffing and routing; closest queueing analogue to engineers working several async tickets at once

Systematizes 93 papers on Regular Expression Denial of Service (ReDoS), contrasting Spencer-style backtracking NFA engines (Python `re`, PCRE, Java, JavaScript) that exhibit worst-case O(2^N) or polynomial O(N^k) super-linear complexity with Thompson NFA / lazy-DFA engines (Rust `regex`, Google RE2) that guarantee O(N) time complexity by tracking…

Mylyn captured per-task interaction context in the IDE, the ancestor of task-attributed activity timing (the approach WakaTime-style tools later popularised with heartbeat-plus-timeout heuristics). Relevant as precedent for attributing interaction time to a task id (here, a ticket) only while the task is active; it does not itself validate time estimates.

Common-good lens: in follow-the-sun support the agent note is a handoff artefact present at minute 0 regardless of shift. Measure whether tickets opened in thin-staffing hours get a usable working diagnosis before the next shift's first touch, and whether the note reduces handoff re-work.

Source of Campbell's law: the more a quantitative indicator is used for social decision-making, the more it is subject to corruption pressures and the more it distorts the processes it monitors. Underpins pairing every SLA attainment metric with a counter-metric and an unadjusted series, and against tying individual reward to ack speed.

Nine-hospital prospective study of the I-PASS structured handoff bundle (mnemonic, training, verbal + written template): medical errors fell 23% and preventable adverse events 30% with no measured increase in handoff time.

Finds a positive association between ACSI customer satisfaction and shareholder value (Tobin's q). Evidence that satisfaction is a legitimate business-outcome dimension, which a support dashboard can link to account retention/expansion once CRM revenue data is joined.

Post-release defect counts scale with usage; normalising by usage gives a quality metric independent of adoption. Relevant to relrate and sigrel: release comparisons must use exposure (node-months or call-home usage).

Parnas (1979) shows that designing software for ease of extension and contraction requires a strict acyclic 'uses' hierarchy where lower-level virtual machine layers have zero dependency on upper-level consumers.

Formulates differential testing (Digital Technical Journal 1998): feeding identical randomized inputs to two independent implementations (such as a verified Lean 4 reference oracle and a SIMD-optimized Rust/C++ engine) so any divergence exposes a translation or implementation defect without manual test oracles.

Firefox rapid release: no more post-release bugs but bugs surface and are fixed faster, and adoption dynamics change what is observed in the field. Warns that release cadence and adoption timing confound field-defect comparisons between releases.

Logged foreground email-client time for 40 information workers over ~12 workdays: nearly 1.5 h/day on email, 77 checks/day, about 4.5 h/day total computer time. Email duration was sessionized with a 5-minute keyboard/mouse inactivity cutoff. Gives a daily email-time anchor and an example of a short inactivity cutoff for in-application activity;

Controlled experiment (EMSE 2025): SQL error messages rewritten per the 2023 design framework led to faster error fixing than stock DBMS messages. Strongest available evidence that rewriting messages (not only adding docs) reduces time-to-fix. Limits: novice population, single-query tasks;

PM seat (council 5): Practitioners capture feedback from many channels but rarely link it systematically to requirements or usage data. Motivates linking tickets to IDEA keys and to fleet feature usage, which the gallery does not do yet.

Interval quality: probability a customer hits a field failure within a window after installing a release, joining support data to deployment counts; defect density alone mis-ranks releases. Template for a per-release field-quality metric with an installed-base denominator, which the relrate/upgrade cards approximate with YBA bystander universes.

A taxonomy of eight leakage types in ML-based science, including temporal leakage (features measured after the outcome) and non-independence between train and test. It proposes model info sheets.

Defines dependency freshness as the time (libyears) or version distance between the used and the newest available release, and shows version-count distance is distorted by release cadence.

CACM 2024 (GitHub). Of many telemetry metrics, acceptance rate of shown suggestions best predicted developers' perceived productivity, better than persistence or retained-code measures.

Interaction-logged field study of 78 professional developers over 3,148 working hours: 57.6% comprehension, 24.0% navigation, 13.4% other, only 5.0% editing. Sessions split at 1-hour idle; within-session gaps over 5 minutes excluded.

METR RCT: experienced open-source developers were 19% slower with early-2025 AI tools while believing they were faster. Operational translation: self-reported time saved by the AI bundle note is not evidence; require measured touch time or a blinded comparison.

Brooks (IEEE Computer 1987) distinguishes essential complexity (inherent domain logic: YugabyteDB Raft consensus, DocDB compaction, YSQL query planning, Bayesian System-1 calibration, and LLM orchestration) from accidental complexity (version skew across 4 worktrees, 4 forked Finding schemas, 3 regex redactors, and OOM-prone multi-pass tarball unpacking).

Framework: organisational learning converts task experience into knowledge embedded in members, tools and routines; knowledge embedded in tools transfers and persists better than knowledge held by individuals.

LinkedIn industry paper (SIGIR 2024 industry track); self-reported A/B, vendor-adjacent

TiKV/PD slow-store detection uses a time-integrated slow score (1-100, slow at >=80) and evicts leaders from one slow store, recovering after 600-900 s. This is a persistence/volume-style trigger, the counterpart to CockroachDB's per-stall 20 s crash: it catches chronic slowness that never crosses a per-stall threshold.

SPACE framework: Satisfaction and well-being, Performance, Activity, Communication and collaboration, Efficiency and flow. Productivity cannot be captured by one metric. Use at least three dimensions, mix perceptual (survey) and system data, and avoid activity counts as performance proxies or for individual ranking.

Compares satisfaction and loyalty metrics (average satisfaction, top-2-box, repurchase intent, recommend intent, NPS, complaints) as predictors of firm business performance; average and top-2-box satisfaction predict best while recommendation-based metrics such as NPS have little or no predictive value.

Systematic review (127 studies) splitting trust in automation into dispositional, situational and learned layers; learned trust is updated by observed performance, and design features (transparency, feedback) shape it.

Firefox rapid vs traditional releases: post-release bug counts similar, but bugs surface and get fixed faster; release comparisons must align on time since release. Supports day-0 alignment in relrate.

PM seat (council 5): Users report quality aspects (reliability, performance, usability) in feedback channels, i.e. feedback reveals non-functional requirements. Support tickets are a richer non-functional channel; supports treating fsync and upgrade cards as roadmap input.

WCAG 2.2 AA audit of the gallery: text contrast passes in both themes (muted 4.8:1 on the light page); 1.4.11 graphic contrast passes for blue, vermilion and green, with deemph-on-sheet borderline at 3.3:1; 2.4.1 bypass is missing; 2.5.3 Label in Name fails (visible 'Theme: light' vs name 'Colour theme');

SZZ: link fix commits to issues via ids in messages, trace fix-inducing changes. Basis for joining field-linked Jira keys to fix commits (fixtest); linkage only finds commits that cite the key.

Formalises technical lag as a function of (used version, ideal version available at reference time t) with pluggable lag measures (version, time, number of releases).

Of ~1.4M GitHub notebooks only ~24% ran without error and ~4% reproduced stored results; out-of-order execution, hidden state and unpinned dependencies dominate.

DevEx proposes three core dimensions: feedback loops, cognitive load and flow state. Each is measured through perceptions (surveys) and workflows (system data), with KPIs at the top level.

Field logging of Microsoft information workers responding to email/IM alerts: after switching to answer an alert, people took on the order of 10 minutes (Mark et al. 2016 cite 'over nine minutes') to return to the suspended task, often drifting to other applications.

Property-based testing in industry: generators plus invariants instead of examples, with shrinking to minimal failures. Applied to a figure gallery: generate perturbed card JSON (deleted keys, nulls, empty arrays, out-of-range numbers) and assert page-level properties (all cards present, failing card fails closed with a message, no NaN/undefined in text).

Introduces PIE (Performance-Improving Code Edits), showing that conditioning LLMs on execution/profiling feedback and sampling multiple candidate rewrites yields substantial speedups over human reference code, whereas single-shot unguided prompting often degrades performance or breaks correctness.

Original metamorphic testing proposal: derive follow-up tests from successful ones through expected input/output relations.

Demonstrates empirically that selective transparency via progressive disclosure in high-stakes diagnostic decision-support interfaces significantly improves practitioner comprehension and calibration over flat explanation dumps, while affective cues and decorative emojis degrade expert trust.

Proposes five requirements for DQ metrics (defined min/max, interval-scaled values, quality of configuration parameters, sound aggregation, economic efficiency) and evaluates existing metrics incl. currency metrics against them. expected cadence) and aggregate only with a defensible rule (weakest-link rather than averaging).

Investigates SemVer adherence and binary/source breaking changes across >100,000 artifacts, motivating automated API/schema compatibility gates (cargo-semver-checks) and immutable version-tag enforcement in CI.

Soufflé (CAV 2016) demonstrates how declarative Datalog rule specifications can be both interpreted at runtime for rapid rule prototyping and ahead-of-time compiled/embedded into high-performance native C++ binaries via staged specialization.

Track provenance of every result, no manual data steps, archive tool versions, record seeds, store data behind plots, link claims to results. Bundle Insights meets seed and package-version rules (wavec RNG seed, manifest) but not input pinning (row counts, not content hashes) nor the no-manual-step rule (fsync JSON).

Establishes the Lean 4 dependently typed theorem prover and functional programming language (CADE-28). In Report 08, Lean 4 v4.13.0 serves as the zero-axiom proof kernel across five production subsystems (157 theorems, 0 sorry, 0 axiom) and emits C code with @[export] C-ABI entry points for direct in-process differential testing against Rust.

Oracle-problem survey: implicit, derived and specified oracles.

Experiments showing people distrust an automated aid after seeing it err, even when it outperforms them, unless told why it errs; explaining errors restored appropriate reliance. Supports a visible error taxonomy and changelog for the triage note.

Empirically studies SemVer compliance across Cargo, npm, Packagist, and RubyGems, highlighting the perils of 0.y.z ('zero-space') permissive ranges vs lockfile drift.

WatchDog IDE telemetry on 416 developers (2,443 in the TSE extension) versus their own estimates: developers believed they spent about half their time testing but telemetry showed about a quarter, a twofold self-report overestimate.