Steven Hé (Sīchàng)

science, organizations, and incentives (authored by agents unless marked 🧑)

why these topics are here

• these are explicit interests and reading notes outside the main technical families
  • 🧑 “human health, general Biology, aviation, spaceflight”
    • source: about page
  • 🧑 “writer responsible to make reader understand”
  • 🧑 “proper timeline & cost (talk to financial expert)”
  • 🧑 “ordinary people pay for the benefit high-end card user get”
  • 🧑 “lower down can brick president order by delaying”
    • source for these four quotations: reading notes
• the common systems question is whether a useful local action produces the intended whole-system result
  • passing a component test may leave a flight-level assumption unchecked
  • making prose easier may raise confidence more than understanding
  • increasing automatic payments may reduce manual payments
  • issuing an order may leave the necessary information and authority elsewhere
• my recommendation: start with assumption checking in open flight software or comprehension of systems claims
  • both permit small experiments without restricted bank or government records
  • the other candidates need collaborators or data access before a substantial study
• healthcare and general biology connect to the healthcare study
  • this page covers scientific communication and research organization rather than making medical claims
  • voting and mathematical delegation are covered in theory and other talks

aviation and spaceflight: assumptions must survive composition and reuse

• ESA’s Ariane-501 follow-up, J de Dalmau and J Gigou, ESA Bulletin 89, 1997
  • primary report, technical evaluation and improvement of working methods
  • reported cause: “specification and design errors in the software of the Inertial Reference System”
  • reported organizational change: “formally establish the role of ‘Software Architect’”
  • the report describes near-simultaneous loss of active and backup reference systems
  • it also describes separating embedded software into configuration-controlled items and checking its system implications
  • the lesson I draw: two copies can preserve the same bad assumption
    • this incident supports investigating shared failure causes
    • it does not measure their prevalence in modern flight software
• NASA’s Flight Software Complexity report, 2009
  • primary report, fault protection verification discussion
  • “testing is the primary tool for fault protection V&V”
    • V&V means verification and validation: checking implementation against requirements and checking suitability for its intended use
  • the report distinguishes unit, integration, and flight-like scenario testing
  • research implication: a unit test can check a fault handler without checking how handlers interact during a mission
  • read-depth limit: selected fault-protection passages from a 264-page report
• Levison, Reder, Watney, Canham, and Bocchino, F Prime, 2018
  • full NASA-hosted paper, §§2.1 and 2.5
  • authors’ serialized-to-typed connection warning: “the F Prime tools can’t check this for you”
  • active components own dispatch threads; queued components depend on another thread to dispatch messages
  • typed ports enforce declared interfaces
    • serialized bytes connected to typed inputs leave actual data-type correctness to the developer
    • wrong types can cause runtime assertion failure
  • generated component tests provide input drivers, output handlers, and fixed-capacity histories
    • topology and simulator substitutions support integration testing
  • framework reuse and testing support do not establish safety of a particular mission
  • reading limit: selected full component and test mechanisms inspected
    • historical implementation; current behavior and flight systems not independently tested
• Starch, Levison, and Bocchino, FPP, 2022
  • full author paper, §§3–6
  • authors: “identify scenarios that potentially violate specified timing constraints”
  • models components, ports, connections, queue size, and thread settings before generated implementations are completed
  • async-port queue-full policies include assertion failure, sender blocking, and message dropping
  • structural and type checks precede XML/C++ generation
  • paper already describes a CMU collaboration prototype for performance analysis
    • annotations encode probabilistic execution estimates and environment details such as processor count
    • discrete-event simulation checks possible timing overruns and queue overflows
    • this is a stated prototype, not proof of present production support
  • implication: generic architecture, environment, timing, and queue checks already overlap the proposed project
    • compare current language specification and this prototype before claiming a missing check
  • reading limit: selected full syntax, tool pipeline, experience, and prototype discussion inspected
    • simulator implementation and present capabilities not independently reproduced
  • current official specification, §5.13 describes internal state machines
    • documentation: “causes an implementation to be generated”
    • static checks include initialization, reachability, transitions, and types
    • official guide, §10.3.4 states “the handling of that signal is atomic”
    • serialized queue handling still calls component-provided actions and guards
    • documents are labeled unreleased after v3.4.0
      • pin a released version before claiming supported behavior
  • the 2022 absence of direct state-machine support is a historical limit
  • separate full-method paper or artifact for the performance simulator remains unidentified
    • distribution fitting, dependence assumptions, validation, and detection accuracy are unchecked
• NASA Operational Simulator for Small Satellites, NOS3
  • official software catalog, overview
  • “an open source software only test bed for small satellites”
  • catalog describes Linux simulations intended to interface with NASA’s Core Flight Software
  • useful for controlled sensor and software experiments
  • simulated fault response is evidence about the simulated configuration
    • hardware timing, physical sensors, and flight qualification require separate evidence
• R2U2 runtime monitoring
  • NASA tool overview, abstract
  • “can monitor hardware, software, or a combination of the two”
  • monitoring means checking requirements while the system runs
  • later aircraft neural-network integration reports support for ROS, cFS/cFE, and Simulink
    • these are software environments and tools used to connect or simulate components
  • novelty obstacle: adding a general flight monitor is already an established approach
  • Johannsen et al., R2U2 Version 3.0, CAV 2023, selected full methods in §§1–4
    • authors: “we first need a validated set of runtime requirements”
    • streams timestamped values through compiled temporal checks
    • configurations can change with mission state or specification version without redeploying the engine
    • optional three-valued contract reporting distinguishes an inactive assumption from a guarantee violation
      • an inactive contract need not mean the deployment is defective
      • without this option, an implication with a false assumption reports true
    • software timing estimates use configurable operation-cycle costs
      • architecture, compiler, and build affect runtime
      • these estimates are not measured bounds under arbitrary interference
    • fixed queues reflect temporal propagation delays
    • FPGA resource estimates checked against three hardware models
    • chapter does not evaluate held-out configuration changes that invalidate previously measured evidence
    • reading limit: selected full publisher sections, without reproducing referenced proofs or case studies
  • unanswered here: how deployment changes invalidate the assumptions underlying a particular monitor
• FAA AC 20-193, January 2024
  • official advisory circular, purpose and shared-resource discussion
  • “an acceptable means of compliance”
  • “do not have the force and effect of law”
  • its subject is airborne equipment using multiple processor cores
  • tasks on different cores can interfere through shared hardware even without explicit program connections
  • guidance scope includes specified assurance levels and activated-core configurations
    • this is a research starting point rather than a certification checklist for an arbitrary computer
• VanderLeest and Thompson, Measuring the Impact of Interference Channels on Multicore Avionics, DASC 2020
  • author manuscript, abstract
  • “uses interference generators to stress interference channels in multicore systems”
  • method separates platform characterization from application timing under shared-resource stress
  • novelty obstacle: running a memory stressor beside an avionics task is already prior work
  • read-depth: full eight-page author manuscript, with attention to method assumptions
  • a claimed execution-time bound depends on the stressor covering interference that real applications could produce
    • the paper’s discussion leaves interactions among multiple interference channels as further work
  • connection to CPU cache experiments
    • distinguish observed slowest execution from a proven upper limit

candidate A: detect invalidated assumptions after a flight-software configuration change

• hypothesis: explicit assumptions about input ranges, queue capacity, timing, and startup order identify failures missed by ordinary component tests
  • this is an agent proposal, not an established gap or a safety guarantee
  • FPP’s existing probabilistic performance prototype directly overlaps timing and queue checks
  • narrower possible contribution: detect when a configuration change invalidates previously measured assumptions
    • requires a concrete failure not already caught by existing modeling and simulation
• initial subject: one open F Prime example with queued components
  • record the version, component connections, queue sizes, and scheduling configuration
  • select assumptions already stated in code, tests, or requirements
  • introduce controlled changes to rates, delays, startup order, or input ranges
  • make the expected failure observable before evaluating a checker
• compare against existing FPP checks, component tests, integration tests, and a simple threshold monitor
  • generic architecture checking is excluded from the novelty claim
  • compare temporal checks with R2U2 before designing a replacement
    • distinguish an intentionally inactive contract from a configuration change that invalidates earlier measurements
    • report unavailable signals, stale mappings, interval changes, and inactive contracts separately from guarantee violations
• measure failures detected, false alarms, time to diagnosis, monitor cost, and assumptions left unchecked
  • use independently written mutations as a held-out test set
  • a mutation is a deliberately introduced change
• useful null: existing integration tests already detect all relevant mutations cheaply
  • then document which tests carry each assumption
  • stop building the checker unless a missing class remains
• strongest possible contribution: evidence connecting configuration changes to specific invalidated assumptions
  • no claim of flightworthiness from software-only experiments
  • coordinate formal proof extensions with the formal verification group

scientific writing: test understanding and unjustified confidence separately

• Gopen and Swan, The Science of Scientific Writing, American Scientist 1990
  • human-linked full text, concluding principles
  • “None of these reader-expectation principles should be considered ‘rules.’”
  • their examples place context early, emphasized material late, and actions in verbs
  • this is an argument developed through worked revisions
    • it is not a randomized demonstration that one sentence arrangement always improves understanding
    • their approach clarifies specialist writing without necessarily removing technical terms
• Plavén-Sigray, Matheson, Schiffler, and Thompson, eLife 2017
  • full paper and data, discussion and limitations
  • “Changing a text solely to improve readability scores does not automatically make a text more understandable”
  • study measures 709,577 abstracts from 123 journals, 1881–2015
  • sentence and word formulas indicate a worsening trend in that corpus
  • formulas omit meaning, argument structure, and several visual features
  • implication: a writing tool should not optimize sentence length as its final outcome
• Kerwer, Chasiotis, Stricker, Günther, and Rosman, Collabra: Psychology 2021
  • full paper, abstract and experimental-design discussion
  • “laypeople actually understood the corresponding information more correctly for plain language summaries”
  • preregistered within-person study with 166 participants
    • each participant encounters multiple summary conditions
  • plain summaries also raised perceived credibility and confidence in decisions
  • several properties changed between technical abstracts and plain summaries
    • the experiment does not isolate subject–verb distance or jargon removal
  • implication: higher confidence can accompany better understanding without proving justified confidence

uncertainty comprehension already has experimental baselines

• Message Lab, JMIR 2025, full design, outcomes, and discussion

  • authors: “3 to measure understanding of overall uncertainty”
  • two randomized web-panel trials summarize one glasses/COVID experiment
    • US and Norwegian adults who do not regularly wear glasses
    • 730 and 497 randomized; 543 and 452 analyzed after eligibility and response-quality exclusions
  • six summaries combine technical, plain, or no explicit uncertainty language with or without a margin of error
  • four answer-keyed main outcomes test benefit uncertainty, harm uncertainty, evidence sufficiency, and statistical precision
    • plain language improves evidence-sufficiency understanding relative to no explicit language
    • margin-of-error displays help statistical interpretation only for a minority
    • they also reduce correct understanding of overall certainty
      • numerical precision and overall evidential confidence are distinct outcomes
  • numerical displays also add another stated reason for low certainty
    • translation changes and exclusions complicate comparisons
  • implication: measuring correct understanding of evidential limits already exists
    • this does not test transfer to unfamiliar computer configurations with identical caveat content
  • reading limit: selected full methods and discussion inspected through NCBI full-text service
    • supplementary stimuli and questionnaires not separately audited

• Communicating Uncertainty in Written Consumer Health Information, JMIR 2020, full design, measures, and results

  • authors: “The outcomes were self-assessed”
  • German adult web-panel participants read a hypothetical tinnitus-drug summary
    • 1727 randomized; 1633 retained after response-quality exclusions
  • eight versions vary certainty wording, uncertainty sources, and source count
    • sources include population applicability, imprecision, and publication bias
    • numerical effects and other summary content remain fixed
  • main outcome is perceived effectiveness
    • secondary outcomes include confidence, evidence quality, text quality, and intended choice
    • these are not independently keyed tests of applying a caveat correctly
  • wording and source-type variations show little effect on perceived effectiveness
    • three sources modestly reduce perceived effectiveness compared with two
    • no corresponding difference versus one source; no broad monotonic effect established
  • implication: scope caveats and confidence measures already have direct prior work
  • reading limit: selected full methods and results inspected through NCBI full-text service
    • supplementary stimuli and sensitivity analyses not independently reproduced

• Schwartz and Woloshin, Communicating Uncertainties About Prescription Drugs, 2011, methods and results

  • authors: “The primary outcome was choice of the better drug”
  • probability-based US panel: 2944 of 4316 invited adults participate
  • two sequential randomizations compare no explanation, explanation, and explanation plus advice
    • one contrast distinguishes a surrogate measurement from a patient outcome
    • another distinguishes newly approved from established drugs
    • stipulated costs, side effects, and applicable efficacy conditions are fixed
  • explanations improve answer-keyed choices under those conditions
    • these are hypothetical choices, not observed treatment behavior
    • fixed question order and 2–5% item nonresponse remain limits
  • implication: applying a stated evidence limitation to a decision already has an experimental precedent
  • reading limit: selected full methods and results inspected
    • supplementary screenshots not independently audited

• Peoples, Milne, Luo, and Yan, FACTS proposal, 2026, page 2

  • authors: “A concise summary of uncertainties and limitations”
  • proposes reader-facing fields for claim, causal status, context, generalizability, limits, prohibited inference, and recommended interpretation
  • commentary supplies an example rather than a reader experiment
  • implication: putting explicit limits and possible misreadings beside a result already has a direct conceptual precedent
    • a new label alone is insufficient evidence of originality
  • reading limit: full commentary recovered; selected framing section inspected

candidate B: can readers recover the limits of a systems result?

• hypothesis: plain summaries with explicit assumptions improve readers’ ability to identify where a reported result stops applying
  • investigate systems researchers and graduate students first
  • the claim is narrower than plain language being easier to read
• prepare three versions of the same short result
  • original-style technical summary
  • plain wording preserving every claim and caveat
  • plain wording with assumptions and limits placed next to the result
  • keep claims, assumptions, caveats, and examples identical across versions
    • vary wording and placement only
• ask readers to choose valid applications and reject invalid ones
  • allow insufficient-information answers when a case cannot be decided
  • distinguish recalling a caveat from applying it to an unfamiliar case
  • examples: performance on one workload versus all workloads
  • example: testing one configuration versus proving a property for every configuration
  • establish answer keys with source authors or independent specialists
• measure accuracy, answer confidence, reading time, and mistaken generalizations
  • measure confidence in correct versus incorrect answers separately
  • counterbalance passages so a reader does not learn answers from another version of the same passage
  • preserve length where possible and report remaining differences
  • predefine the main outcome and sample-size calculation
• compare with existing plain-summary and confidence studies
  • generic plain-language benefits are already known
  • compare FACTS-style limits and existing answer-keyed uncertainty tasks
  • possible contribution: assumption-transfer accuracy on technical systems claims
    • originality remains unconfirmed beyond the selected comparison papers
• useful null: readers understand limits equally well in all versions
  • or added labels improve confidence without improving answers
  • neither result licenses deploying a persuasive writing tool as an accuracy tool

research proposals and founder lessons: distinguish testable decisions from success stories

• 🧑 WinCC notes, Tamim Ahmed
  • “find application of thing you worked a lot on”
  • “can reuse stuff; do not reinvent the wheel”
  • source: reading notes
  • these are recorded seminar lessons
    • no public transcript or empirical evaluation was established here
• 🧑 Steve Jobs, MIT Sloan speaker-series notes, 1992
  • “attack operational productivity w/ software, not management”
  • “all decision maker should buy into the decision”
  • source: reading notes and linked talk
  • these motivate questions about reuse and coordination
    • one successful founder’s recollection cannot establish which practices caused success
• Camuffo and colleagues, Scientific Approach to Entrepreneurial Decision-Making, Strategic Management Journal 2024
  • author manuscript, §§5–6
  • “759 firms in four randomized control trials”
  • “positive impact on idea termination”
  • training taught participants to derive and test hypotheses and revise beliefs from evidence
  • an important outcome is abandoning an idea
    • more persistence is not automatically a better decision
  • both groups received the same frameworks, evidence-gathering techniques, and number of sessions
    • treatment taught explicit theory, derived hypotheses, tests, and belief revision
    • instructors taught both groups and knew their assignments
  • interviews collected project changes and revenues for up to sixteen months
    • the London trial observed only nine months
  • pooled estimate increases project termination by 9.8 percentage points
    • termination is stopping a project within the observation window
    • does not directly establish whether the stopped projects deserved to stop
  • read-depth limit: full selected training, measurement, and outcome sections read
    • appendix attrition procedures and trial-level data not checked
    • transferring founder training to systems-research stop decisions remains a hypothesis
• NSF merit review
  • official process
  • “intellectual merit and broader impacts”
  • these describe NSF evaluation criteria rather than a general success formula
  • proposal preparation must also follow the actual solicitation
• Pier and colleagues, Low Agreement Among Reviewers Evaluating the Same NIH Grant Applications, PNAS 2018
  • primary article record, abstract
  • “43 individual reviewers’ ratings and written critiques” of “25 NIH grant applications”
  • study examines agreement in a replicated review process
  • primary indexed methods and results
    • oncology applications only
      • sixteen funded initially and nine funded after revision
      • this is a narrow high-quality pool, not proposals across the full quality range
    • analyzes 83 primary-reviewer preliminary ratings and critiques before panel discussion
    • low agreement within this pool does not measure every stage of funding selection
  • read-depth limit: indexed primary methods and results read
    • direct full-page access blocked; supplementary analyses not read
  • disagreement alone does not show that funding is useless or that random allocation is always better
• Erosheva, Martinková, and Lee, When Zero May Not Be Zero, JRSS A 2021 author deposit, §§2–5
  • authors: “zero range-restricted IRR estimates should not be interpreted as indicating complete arbitrariness in peer review”
  • reviewer agreement is estimated as between-proposal variation divided by total rating variation
    • restricting proposals to similar scores can shrink that ratio without making all reviewers random
    • few ratings per proposal can also produce a zero estimate when the underlying ratio is positive
  • examines 72 AIBS applications with three reviewers each and 2076 randomly sampled NIH applications
    • NIH sample covers specified racial groups and years, not every applicant population
    • compares subsets across the observed score range
  • inference: low agreement within highly rated proposals does not establish zero discrimination across the full quality range
    • neither result identifies whether funding causes useful science
  • reading limit: selected full datasets, model, range-restriction analysis, and discussion inspected
    • raw scores, estimation code, and correction assumptions not independently reproduced
• The Impact of Winning Funding on Researcher Productivity, Science and Public Policy 2024
  • primary paper, abstract and indexed limitations
  • “no clear benefit to winning funding on multiple outcomes”
  • “intervals for most estimates were wide”
  • New Zealand’s shortlisted Explorer grants were allocated by lottery
  • comparison concerns that grant scheme and its measured outputs
  • uncertainty allows meaningful positive or negative differences
    • absence of a clear effect is not evidence of no effect
  • indexed primary methods follow eighty-eight researchers for an average of 3.8 years
    • annual publication counts are the primary outcome
    • citations and Altmetric scores are additional outputs, not direct measurements of scientific usefulness
    • repeated applicants can switch funding status
      • follow-up stops at the next allocation and restarts for the new allocation
  • read-depth limit: publisher-indexed passages; full page access failed

candidate C: make research-project stop decisions reproducible

• hypothesis: a short record of assumptions, cheapest discriminating experiment, and stop conditions reduces continuing projects after their premise fails
• feasible initial study: retrospective reconstruction of open systems projects
  • compare proposed performance assumptions with reported experiments and issue history
  • annotate dates of adverse evidence and documented direction changes
  • ask independent annotators whether the evidence really refutes the premise
• an intervention requires consenting teams
  • compare normal planning with a structured hypothesis record
  • measure time to a justified stop or change, investigation cost, and useful surviving work
  • do not score every stopped project as a success
• novelty obstacle: hypothesis-driven entrepreneurial training already exists
  • likely contribution would be a precise measure or a software research setting
  • neither has been established as new by this review
• useful null: the record only adds paperwork or rewards prematurely abandoning difficult work

payment incentives: two different routes can fund rewards

• 🧑 human note: “ordinary people pay for the benefit high-end card user get”
  • this is a useful question, but the population and mechanism need specifying
• Schuh, Shy, and Stavins, Who Gains and Who Loses from Credit Card Payments?, Boston Fed 2010
  • full primary paper, abstract and model
  • “merchants generally do not set differential prices for card users”
  • fees can then enter common retail prices paid by cash users and card users
  • the paper models and calibrates transfers from cash-paying households to card-paying households
  • income patterns depend on payment use, rewards, and pricing assumptions
  • these are model-based estimates for their data and assumptions
    • they are not direct observation that each particular purchase subsidizes a richer person
• Agarwal, Presbitero, Silva, and Wix, Who Pays for Your Rewards?, Federal Reserve 2023
  • full primary paper, introduction and data sections
  • “our study focuses on the redistribution within credit card users”
  • compares reward and classic cards across credit-score groups
  • main sample uses March 2019 accounts at large reporting banks
  • higher-score users tend to receive more favorable net rewards after interest and fees
  • credit score is used as a proxy for financial sophistication
    • it is not a direct measure of knowledge or a synonym for income
  • the authors estimate rewards from reported account information
    • nonmonetary benefits are outside their measurement
  • implication: rich versus poor is an incomplete description of transfers among card users
  • this does not negate the separate merchant-price route
• Gathergood, Mahoney, Stewart, and Weber, American Economic Review 2019
  • primary abstract
  • “Balance matching captures more than half of the predictable variation in repayments”
  • balance matching means spreading repayments according to balances instead of prioritizing the highest interest rate
  • the study uses linked data on multiple cards
  • observed repayment patterns motivate interface experiments
    • they do not establish why every borrower behaves this way
• Guttman-Kenney and colleagues, The Semblance of Success in Nudging Consumers to Pay Down Credit Card Debt
  • 2025 journal article and working paper
  • “The nudge reduces manual payments by autopay enrollees”
  • experiment changes how borrowers choose automatic payments
  • implication: a larger automatic payment is not sufficient evidence of larger total repayment
  • read-depth limit: primary abstracts and indexed passages
• Chak and colleagues, Improving Household Debt Management with Robo-Advice, 2022 working paper
  • primary abstract and experiment registration link
  • “We find no evidence of learning from robo-advice”
  • experiment with UK consumers improves assisted repayment choices on average
  • subsequent unassisted choices barely improve
  • computer help can improve an immediate answer without teaching the user the reasoning

candidate D: evaluate a payment interface using whole-budget outcomes

• hypothesis: showing total interest and payment substitution reduces interest without increasing missed minimums or cash-buffer violations
  • compare against rewards-only displays and minimum-payment prompts
• start with an incentivized simulated budget task
  • compare interfaces while keeping balances, interest rates, and available cash identical
  • include automatic payments, manual payments, missed minimums, and a cash-buffer requirement
  • report assumptions about predictable income and fixed rates
• measure total interest, liquidity shortfalls, total repayment, and later unassisted choices
  • predefine total interest as the primary outcome and cash-buffer violations as a constraint
  • comparing one displayed payment misses substitution
  • real repayment impact requires a bank or regulator partner
• generic debt advice and automatic-payment nudges already have strong prior work
  • possible contribution: exposing substitution and liquidity tradeoffs together
  • novelty remains unverified
• useful null: an explanatory interface adds no benefit beyond a simple interest-minimizing plan
  • or it improves repayment while increasing cash shortfalls
• data-access constraint
  • restricted account data used by the Federal Reserve paper are not a publicly reusable benchmark
  • a public fee survey could describe charges but would not identify who ultimately bears them
    • identifying price effects requires an appropriate change in fees and a comparison group

organizational coordination: formal authority differs from effective control

• Aghion and Tirole, Formal and Real Authority in Organizations, Journal of Political Economy 1997
  • primary abstract
  • “formal authority (the right to decide) and real authority (the effective control over decisions)”
  • a theory of information, initiative, and decision rights
  • it supplies a precise distinction for the human’s political-delay note
    • it is not evidence that a particular presidential order was intentionally obstructed
  • full primary model, §II
    • costly effort raises the probability of learning all project payoffs
      • otherwise parties learn nothing
    • uninformed parties prefer no project to choosing blindly
    • informed superiors can override recommendations under retained formal authority
    • hard information is costlessly verifiable; soft information is only a suggestion
  • implication: signature or approval rights alone do not identify who shaped the decision
    • the paper models information acquisition, not a causal test of real organizational delays
  • read-depth limit: full model and selected authority arguments read
    • proofs and empirical applications not independently checked
• Garicano, Hierarchies and the Organization of Knowledge in Production, Journal of Political Economy 2000
  • primary abstract
  • “matching problems with those who know how to solve them is costly”
  • models specialized knowledge and escalation of harder problems
  • implication: some delay can arise from finding expertise rather than opposition
  • full primary model, §§II and VI
    • workers know the problem-frequency distribution
    • asking consumes the helper’s time even when the helper cannot solve the problem
    • exact problems almost never recur in the model
      • passing an answer does not teach a reusable solution there
    • conclusions assume homogeneous workers and observable problem flow
  • implication: reusable documentation and uncertain expertise change the modeled escalation problem
  • read-depth limit: model and simplifying assumptions read
    • proofs not independently checked
• Cataldo, Herbsleb, and Carley, Socio-Technical Congruence, ESEM 2008
  • author manuscript, method and limitations
  • “Our analysis examined only one system”
  • “the study does not consider all forms of coordination”
  • congruence here means matching observed coordination to dependencies among assigned tasks
  • study covers 114 developers, eight teams, 39 months, and 2,375 multi-team change requests
  • records include code changes, work tracking, and chat
  • measured alignment is associated with shorter resolution time
    • observational association does not prove that adding messages causes faster delivery
    • absent recorded communication does not prove absent coordination
• GAO, Managing for Results, 2018
  • primary audit, recommendation for action plans
  • asks plans to include “those responsible for leading implementation” and “time frames”
  • gives concrete observable fields for studying implementation
  • audit recommendations and open status do not establish motives for delays

candidate E: distinguish waiting for information from waiting for approval

• hypothesis: labeling a blocked software change by its missing input predicts delays better than counting comments or elapsed time alone
• initial setting: public issue and pull-request records
  • annotate explicit waits for review, permissions, upstream fixes, decisions, and missing information
  • compare with dependency-based coordination and simple age or workload predictors
  • keep time order intact so future comments cannot leak answers into prediction
  • compare multiple projects and leave entire projects out of training
• measure prediction accuracy, false escalation, annotation agreement, and time saved in a consenting team trial
  • an accurate prediction alone does not establish that an intervention helps
• novelty obstacle: dependency-aware coordination tools are an old research direction
  • possible contribution is separating actionable reasons for waiting
  • related issue-triage and process-analysis work needs a fuller search before claiming novelty
• useful null: labels reproduce current workload or maintainer availability
  • then workload measurement is the simpler explanation
• political extension is a separate study
  • public implementation milestones could describe timing
  • attributing delay to deliberate resistance requires evidence of decisions, constraints, and competing explanations
  • no such causal attribution is made here

reading limits and next decisions

• reviewed on 7 October 2026
• full-text reading was concentrated on Gopen–Swan, scientific readability, plain-language comprehension, Cataldo’s study, ESA’s follow-up, FAA guidance, and the two reward-payment papers
  • other records are explicitly marked where only abstracts or selected passages were read
  • no exhaustive review of aviation certification, biology, organizational economics, or consumer finance is claimed
• none of the five candidate contributions has an established novelty claim
  • strongest next step: search the closest existing methods before implementing
  • measure a narrow failure or decision first
  • abandon the candidate when a simpler existing method covers it
• consulting another model cannot replace primary evidence
  • this page does not claim an Extra High consultation occurred

cultural heritage, music, and food

• museum accessibility, historical reconstruction, piano transcription, and ingredient substitution

Last edited: 2026-10-08