The throughput norm is not the rule under test: RDCO's arrangement sits on the safe side of the one measured automation threshold, but it lands squarely in the automation-bias geometry instead, and no RDCO rule closes an outcome loop on a founder decision
The question
"Does RDCO's studio-throughput norm ('build when discovered, never park for the night cron') cause the founder's own decision-making skill to atrophy by routing judgment through Ray instead of him, and how do other founders who delegate heavily to AI agents deliberately keep their own judgment sharp?"
Auto-promoted from the research backlog on 2026-11-06 (score 13/15) out of [[2026-10-30-every-marx-alienation-coding-agents]], whose closing move - that an idle agent factory is a choice, not a loss - the vault logged as a direct and unanswered challenge to feedback_studio_throughput_no_deferral. The subject here is the norm and the rule set, not any person.
What we already know (from the vault)
- The citation that anchors this entire thread was already audited and found overstated. [[2026-10-08-finding-level-citation-audit-arxiv-gloss]] sampled arXiv 2608.23642 (Mitchell, Ghosh & Passi, "AI Agents Push Humans Out of the Loop") as it appears in [[2026-09-08-every-to-read-or-not-to-read-the-code]] - glossed there as "a paper ... finding that extended AI-agent use measurably erodes vigilance, critical thinking, and domain skill" - and graded it Overstated: a self-described position paper, zero own-measurement phrases, arguing from Bainbridge 1983 and a secondary-literature appendix. Independently re-confirmed this pass against the v3 full text, which says verbatim "This position paper argues ...". The question therefore entered the queue carrying a defect the vault had already caught. Cite 2608.23642 for framing and design proposals, never as evidence.
- The vault already holds the better empirical anchor, and it is about reliability breeding disengagement, not about decay. [[2026-04-20-every-ai-autopilot-verification-decay]] carries Sarkar et al. (Microsoft Research, arXiv:2412.15030): the deep risk is not hallucination but that competent-enough output stops being checked, and their proposed counter is interface provocations - the AI as critic or provocateur rather than assistant. That note already drew the right implication: "build verification that forces engagement, not verification that lets you skip engagement."
- The vault's own prior answer locates the degradation mechanism in trust, not time. [[2026-10-08-verification-ratio-senior-signoff-throughput]], quoting [[2026-06-03-parallel-agent-fde-capacity]]: "your attention depletes, your trust in the output inflates" and sustained review "quietly shifts from deep evaluation to surface scanning." Same brief's audit analogue is load-bearing here: engagement-quality review runs ~66 reviewer hours against ~11,277 total audit hours (~0.6%), and review hours are strongly sublinear in artifact size (14x span in total hours, 2x span in reviewer hours). Liability-grade sign-off scales by sampling judgment, not by checking the artifact.
- RDCO's actual division of labour is already a stage split, written across four rules.
feedback_distinguish_decision_from_actionassigns reversible items to execution and reserves irreversible ones (production deploy, public post, external email, financial commitment) for founder judgment;feedback_auto_mode_signal_to_noisedefaults to action on reversible work;feedback_complete_then_ask_for_reviewcompletes two-way-door work with labeled guesses;feedback_advisor_not_pair_programmersets the return format as "a question + recommended answer." - The throughput norm's own text is about scheduling, not authority.
feedback_studio_throughput_no_deferral(founder, 2026-09-01) responds to a specific failure: queue items parked for the overnight round, overnight running a single artifact, daytime progress only on nudges. Its operative clauses are when buildable work runs and how many artifacts a round takes. It changes nothing about who decides.
What the web says
The one measured stage threshold, and RDCO's arrangement is on the safe side of it.
- Onnasch, Wickens, Li & Manzey (2014), "Human Performance Consequences of Stages and Levels of Automation: An Integrated Meta-Analysis," Human Factors 56(3):476-488 (https://depositonce.tu-berlin.de/bitstreams/6a0b7962-cb5f-4317-9d07-ee11707dd51f/download). k=18 experiments. Overall: routine performance improves (Kendall's tau = +.51, p<.001, 16 studies) while return-to-manual performance after failure degrades (tau = -.34, p=.03, 9 studies) and situation awareness degrades (tau = -.29, p=.04, 11 studies). In the post-hoc moderator restricted to the 6 studies that varied automation across the information-analysis to action-selection boundary, the trade-off amplifies sharply: routine tau = +.68, return-to-manual tau = -.90. Printed Key Point: "Negative consequences of automation are most likely when DOA moves from Stage 2 to Stage 3." Quality caveats that matter: this is rank-order data combined by Kendall's tau, not an effect-size meta-analysis; no total N; k=6 in the load-bearing moderator; the authors themselves note no included study used real operators in complex real-world settings; and the result is contested (Jamieson & Skraaning 2019, with a 2020 reply).
- The limit that kills the easy answer: Onnasch's moderator groups Stage 3 (decision selection) and Stage 4 (action implementation) together as "action" support. It does not contrast automating execution against automating the decision. The specific contrast RDCO's arrangement needs has not been measured by anyone in this literature. Onnasch's own recommendation is that operators be kept involved "in decision and action selection as well as action implementation."
- Endsley & Kiris (1995), "The Out-of-the-Loop Performance Problem and Level of Control in Automation," Human Factors 37(2):381-394. Five levels on a navigation task: manual; decision support (system suggests, human decides and acts); consensual (human concurs); monitored (human vetoes); full. Measured: the post-failure decision-time decrement was significantly greater under full automation than under intermediate levels, and SA loss hit Level 2 SA (comprehension) while Level 1 (perception) was unaffected. Intermediate levels were protective, explicitly attributed to keeping the operator "involved in the active decision making loop" rather than to vigilance. Source quality: the 1995 paper is paywalled; design and results above are from Endsley's own 1996 chapter describing it (https://maritimesafetyinnovationlab.org/wp-content/uploads/2019/12/Automation-and-Situation-Awareness-Endsley.pdf) - near-primary but secondhand. N and effect sizes could not be retrieved. Note also that all five of its levels automate the decision function, so it is a decision-authority gradient, not an execution-versus-decision contrast.
The failure shape RDCO actually inherits is automation bias, and it is the failure mode of exactly this arrangement.
- Parasuraman & Manzey (2010), "Complacency and Bias in Human Use of Automation," Human Factors 52(3):381-410 (https://depositonce.tu-berlin.de/bitstreams/cafd2873-814b-4c59-bab1-addd42e249d2/download). Review plus integrative model, not an experiment. Their definition of complacency requires a monitoring role, so complacency proper does not transfer to an arrangement where the human decides. But automation bias does, and it is specifically the failure of the configuration where the human retains the decision and the aid only advises. The figures they marshal: Skitka, Mosier & Burdick (1999) found 41% omission errors when the aid failed to prompt (against ~97% correct detection with no automation) and ~65% commission errors, with participants following wrong directives despite directly available, known-valid disconfirming readings; Mosier et al. (1998) found 67% of pilots recalled a confirming indication that was never present ("phantom memory"). Mechanism: active attentional reallocation under load, not passivity - moving the automated task to display centre did not remove the effect (Singh, Molloy & Parasuraman 1997).
- Bainbridge (1983), "Ironies of Automation," Automatica 19(6):775-779 (https://ckrybus.com/static/papers/Bainbridge_1983_Automatica.pdf). Argued, not measured - a brief IFAC paper with zero original data, which the 2608.23642 position paper then leans on. Its mechanism cuts against the comfortable reading of Onnasch: she argues that "working storage," the running model of system state that fast decisions draw on, is built by doing on-line control, so losing execution is what erodes the knowledge a decision needs. On her account, delegating execution while retaining decisions is not automatically safe. Theoretical, and the strongest theoretical counter available.
- Aviation's institutional findings sit with bias and monitoring rather than with decision atrophy. FAA PARC/CAST (2013), "Operational Use of Flight Path Management Systems" (https://www.faa.gov/sites/faa.gov/files/aircraft/air_cert/design_approvals/human_factors/OUFPMS_Report.pdf), 29 findings: Finding 2 names manual-handling skill definition, development and retention as a vulnerability; Finding 4 says pilots "sometimes rely too much on automated systems and may be reluctant to intervene." NTSB/AAR-14/01 (Asiana 214) (https://www.ntsb.gov/investigations/AccidentReports/Reports/AAR1401.pdf), Finding 8: insufficient airspeed monitoring "likely resulted from expectancy, increased workload, fatigue, and automation reliance," with the analysis explicitly citing Parasuraman & Manzey 2010. Both are analytic conclusions from accident and normal-ops data review, not controlled measurement.
Decay rates: the direction is settled and it is the opposite of the popular claim; the rate is not.
- Arthur, Bennett, Stanush & McNelly (1998), "Factors That Influence Skill Decay and Retention," Human Performance 11(1):57-101 (https://gwern.net/doc/psychology/spaced-repetition/1998-arthur.pdf). 189 data points from 53 articles; 178 points / N=8,719 in the retention analysis; 88% laboratory studies. Overall loss d = -0.97. The load-bearing moderator: physical tasks delta = -0.75 (k=69) versus cognitive tasks delta = -1.15 (k=109), the paper's own words putting the gap at "close to half a standardized unit (i.e., 0.40) across all retention intervals." Their "cognitive" category explicitly includes "problem solving, and decision making." Cognitive and decision skill decays faster than physical skill. The common framing that muscle memory is the fragile part and knowledge is durable is backwards. Equally important: the interval curve is non-monotonic (1-7d -1.03 at k=89, then 7-14d -1.54 at k=13, 14-28d -0.95, 28-90d -0.65, 90-180d -1.42 at k=7), so no "skills decay in N weeks" claim can be sourced to this paper. Authors also warn moderators were tested one at a time and may be confounded.
- Casner, Geven, Recker & Schooler (2014), "The Retention of Manual Flying Skills in the Automated Cockpit," Human Factors 56(8):1506-1516 (https://gwern.net/doc/technology/2014-casner.pdf). This is the best primary test of the brief's hinge, and it tested the supervisory-maintenance hypothesis explicitly and did not support it. N=16 active 747-400 pilots, full simulator, automation level varied within-subject. Verbatim: "pilots' instrument scanning and aircraft control skills are reasonably well retained when automation is used," but "the retention of cognitive skills needed for manual flying may depend on the degree to which pilots remain actively engaged in supervising the automation." Only 1 of 16 flew the arrival/approach/missed approach error-free; 7/16 misjudged the missed approach point; 14/16 decreased pitch inappropriately on a failed indicator. The authors had hypothesised that "pilots could maintain their navigation skills by actively monitoring the FMC"; instead errors tracked task-unrelated thought (navigation r(14)=.56, p<.05; stick-shaker count r(14)=.76, p<.001). Their read: "Pilots who kept their minds on task more performed better." Also: pilots who reported practising abnormal-indication recognition in recurrent training "performed no better on any of the three instrument failure events." Caveats: N=16, cross-sectional, no controlled retention interval, practice self-reported, so the decay inference is correlational.
- Budzyn et al. (2025), Lancet Gastroenterology & Hepatology 10(10):896-903 (paywalled; abstract and repository record at https://wrap.warwick.ac.uk/id/eprint/191005). The only field observation of a judgment skill degrading while the human stayed in the supervisory seat: 1,443 non-AI-assisted colonoscopies across 4 Polish centres, 3 months before versus after CADe introduction, by endoscopists with >2,000 prior procedures each. Unassisted adenoma detection rate fell 28.4% to 22.4%; -6.0 pp (95% CI -10.5 to -1.6, p=0.0089); OR 0.69. Retrospective, observational, before-after, single country - cannot establish causation, and read abstract-only.
- What maintains judgment is feedback and depth, not volume of decisions. Chang, Chen, Mellers & Tetlock (2016), Judgment and Decision Making 11(5):509-526 (open access, https://doi.org/10.1017/S1930297500004599), read in full. Across four 9-month forecasting tournaments, a sub-1-hour debiasing training module "consistently improved accuracy (Brier scores) by 6 to 11%." Critically: "The effects of training were consistently significant, but a subject merely having prior experience in a year of the tournament did not yield significant improvements in Brier score." And "concentrated attention on a smaller number of questions yielded better practice results than sporadic participation" - forecasts-per-question mediated the training-to-accuracy path in 3 of 4 years; total forecast volume did not. The related Mellers et al. (2014) "training, teaming, tracking" effects should be stated with care: a later reanalysis argues better-fitting models "substantially eliminated, reduced, and in some cases even reversed" the teaming and training effects. Kahneman & Klein (2009), American Psychologist 64(6):515-526 (paywalled, abstract only, https://pubmed.ncbi.nlm.nih.gov/19739881/) supplies the framing - judgment quality depends on the predictability of the environment and the opportunity to learn its regularities through feedback - but it is an adversarial-collaboration review, argued throughout, and supports no decay rate.
Direct evidence on heavy AI delegation and operator skill: thinner than the discourse implies, and it measures the wrong construct.
- Nothing measures whether heavy AI delegation degrades an experienced delegator's judgment. The clean studies measure skill acquisition while using AI, not retention of an existing capability. State that gap plainly rather than filling it.
- Bastani et al. (2024), "Generative AI Can Harm Learning" (https://nmdprojects.net/teaching_resources/bastani_generative_ai_can_harm_learning.pdf). Field RCT, ~1,000 Turkish high-school students. Performance with the tool rose 48% (plain ChatGPT) and 127% (a prompt-engineered "GPT Tutor"); when access was removed, the plain-ChatGPT group performed 17% worse than students who never had access, an effect "largely mitigated by the safeguards included in GPT Tutor." This is the single most useful design finding available: the harness, not abstinence, determined whether capability survived. Caveats: working paper, novice learners, math, not a judgment measure.
- Shen & Tamkin (2026), "How AI Impacts Skill Formation," arXiv:2601.20245. Pre-registered between-subjects RCT, N=52 analysed, learning a Python async library. AI-assisted group scored 4.15 points lower on a 27-point comprehension quiz (~17%), Cohen's d = 0.738, p = 0.010, largest gap in debugging, and no significant completion-time advantage. Vendor-affiliated (Anthropic) but pre-registered with a stated null on productivity. Caveats the authors list: ~1 hour, one task, one library, chat interface, crowd workers, single-quiz outcome.
- Dell'Acqua et al. (2023), "Navigating the Jagged Technological Frontier," HBS WP 24-013, pre-registered field experiment with 758 BCG consultants: inside the task frontier ~12.2% more tasks and ~25% faster at higher quality; outside it, AI users were ~19 percentage points less likely to be correct, with reduced output diversity. hbs.edu and SSRN both returned 403 this pass; figures are widely reported but not PDF-verified here. Dell'Acqua's "Falling Asleep at the Wheel" recruiter field experiment (https://innovationgrowthlab.org/node/1447) found recruiters paired with the lower-quality AI outperformed those paired with the better one, exerting more effort. Closest thing to "high AI quality degrades operator engagement"; still a working paper, N secondhand.
- METR (2025), experienced open-source developers (https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/): 16 developers, 246 real tasks in their own repos, 19% slower with AI while forecasting a 24% speedup and still believing ~20% afterward. This is a calibration result, not a deskilling result: the self-assessment was wrong by roughly 39 points. METR's Feb 2026 update redesigns the experiment after an unreliable follow-on signal; the reported CI spanning roughly -38% to +9% is secondary-sourced and the post could not be opened.
- Over-cited and not usable as capability evidence: Lee, Sarkar, Tankelevitch et al. (CHI 2025, https://advait.org/files/lee_2025_ai_critical_thinking_survey.pdf) is a survey, N=319, self-report only, and its actual finding is that critical thinking shifts in kind toward verification and task stewardship. Kosmyna et al. (2025) "Your Brain on ChatGPT" (arXiv:2506.08872) is a preprint whose "brain atrophy" framing rests on a crossover session completed by 18 participants; the authors themselves asked journalists not to claim capability or IQ effects. Risko & Gilbert (2016), "Cognitive Offloading," TiCS 20(9):676-688 (https://discovery.ucl.ac.uk/id/eprint/1508770/) concludes substitution driven by metacognition, not degradation, and explicitly says long-term consequences are under-studied; citing it for "offloading erodes ability" misreads it. Sparrow, Liu & Wegner (2011), Science 333:776-778, measured a memory-allocation tradeoff in small single-session undergraduate samples with a mixed replication history.
Founders who delegate heavily and deliberately preserve judgment: the evidence base here is weak, and should be labelled as practitioner hypotheses rather than practice evidence.
- Every named practice found is an essay, talk or podcast by someone with a product or audience stake. None is instrumented; nobody in this set measures whether their practice works.
- Named and sourced: Simon Willison, "vibe engineering" - 12 named disciplines including comprehensive stable tests, plan-then-hand-off, preview-environment shipping, really good manual QA, and "an instinct for what can be outsourced" (https://simonwillison.net/2025/Oct/7/vibe-engineering/); and his own honest counter-statement, "I suspect a key skill is developing an intuition for when you don't need to closely review every line of code they produce. This feels deeply uncomfortable!" (https://simonwillison.net/2025/Oct/11/uncomfortable/). Armin Ronacher works "almost entirely hands-off" but wants to see the prompts and the failed attempts rather than just the diff, on the grounds that a PR carries too little information to review agent output (https://lucumr.pocoo.org/2025/12/22/a-year-of-vibes/). Kieran Klaassen / Every, attention deliberately placed at the two ends of the loop, agents in the middle (https://every.to/context-window/you-re-the-bread-in-the-ai-sandwich). Andrej Karpathy, the "autonomy slider" - autonomy as a dial with GUIs built for oversight ("Software Is Changing (Again)," YC AI Startup School). Steve Yegge, role-specialised agents so review is a distinct agent-checked stage (https://changelog.com/friends/96).
- Unverified, do not attribute: a widely-repeated Mitchell Hashimoto line about alternating ("if I'm coding, I want an agent planning; if they're coding, I want to be reviewing") surfaced only in third-party write-ups this pass, with no primary transcript retrieved.
- A real negative result: no named person was found adapting an established decision-hygiene practice - decision journals with outcome review, pre-mortems, Amazon's written-narrative six-pager, Bridgewater's believability weighting - to AI-delegation contexts. The only adjacent artifact is Tomasev, Franklin & Osindero (DeepMind), "Intelligent AI Delegation," arXiv:2602.11865, which mentions a pre-mortem only as something an agent fails to supply.
Convergences and contradictions
- Convergence, and it is the brief's spine: the maintenance mechanism is engagement with outcome feedback, not seat position and not decision count. Casner tested supervisory position as a maintenance mechanism and it failed, while attentional engagement predicted performance. Chang et al. found training and depth-per-question improved Brier scores while a year of tournament experience alone did not. Endsley & Kiris attributed SA loss to passive decision processing rather than to vigilance. Kahneman & Klein name feedback as the condition. Four independent literatures point the same way. Label this convergent inference, not a measured rate - nobody has measured the rate.
- Contradiction the brief must not paper over: Onnasch and Bainbridge disagree about whether delegating execution is safe. Onnasch's measured threshold sits at the analysis-to-decision boundary, which implies keeping the decision is the protective move. Bainbridge argues the opposite causal path - that on-line execution is what builds the state model a decision draws on, so losing execution erodes the decision. Onnasch is measured but rank-based with k=6 in the load-bearing cell; Bainbridge is argued with no data. Neither tested the contrast directly, because Onnasch's moderator lumps decision selection and action implementation together as "action." The specific question RDCO is asking has not been measured.
- Contradiction with the popular framing, settled in the data: Arthur et al. put cognitive and decision skills on the faster-decay side (delta -1.15 versus -0.75 physical), and Casner found control skills retained while the cognitive accompaniments degraded. The intuition that judgment is the durable part is the backwards claim.
- Convergence between the vault and the web on the actual risk: Sarkar et al. (already in the vault since April), Parasuraman & Manzey's bias figures, and the vault's own "trust in the output inflates" line all describe the same thing - a reliable aid plus a retained decision produces under-checking, not incapacity. Bastani adds the one design-level finding: the harness determines the outcome.
- A self-consistency note worth logging: the chain that produced this question (Stewart's alienation essay, resting partly on Klaassen's piece, resting on 2608.23642) carried an overstated empirical claim that [[2026-10-08-finding-level-citation-audit-arxiv-gloss]] had already graded Overstated six weeks earlier. The backlog promoted the question without the audit's correction attached. That is a queue-plumbing gap, not a research finding, but it is cheap to fix.
Synthesis for RDCO
The studio-throughput norm is the wrong rule to put on trial, and the brief's first finding is that the challenge in the parent piece is misaimed. Read the norm's own text: feedback_studio_throughput_no_deferral governs when buildable work runs (immediately rather than parked) and how many artifacts an overnight round takes. It is a scheduling and utilization rule. It does not touch who decides anything. Stewart's closing move - that an idle agent factory is a choice - is an argument about mode selection, which in RDCO's rule set lives in feedback_distinguish_decision_from_action, feedback_auto_mode_signal_to_noise and feedback_ic_vs_production_mode, none of which the throughput norm overrides. A norm that makes execution faster cannot by itself move a decision from the founder to Ray, because the reversibility classifier, not the clock, assigns decisions. Verdict on the norm: keep as written. The one real second-order effect is volume at the review boundary, and the vault already answered that: liability-grade review is sublinear in artifact size (66 reviewer hours against 11,277 audit hours, 14x span in work versus 2x span in review), so more artifacts per round does not mechanically mean proportionally more judgment load.
Map RDCO onto the stage taxonomy and the arrangement sits on the safe side of the only measured threshold - but it lands in a different, better-documented failure shape. In Parasuraman/Sheridan/Wickens terms, RDCO is not one point on an autonomy scale but a bifurcated profile: full automation across all four stages for the reversible class, and Stage-3-retained for the irreversible class (deploys from main, external sends, spend at or above threshold, publishing), which corresponds to Endsley & Kiris's protective intermediate levels. Onnasch's measured collapse in return-to-manual performance (tau -.90) happens when automation crosses into choosing; RDCO's irreversible class does not cross it. Full autonomy does apply to the reversible class, where by construction a degraded take-over is cheap - that is a genuinely good structural property and it should be stated as such. But protective is not the same as neutral. Endsley & Kiris still measured lower comprehension-level SA under semi-automation than under manual control. And the configuration where a reliable aid recommends and the human retains the decision is precisely the automation-bias geometry, with the worst-measured numbers in this whole brief attached to it: 41% omission and roughly 65% commission errors in Skitka et al., with disconfirming evidence directly available. feedback_advisor_not_pair_programmer directs Ray to frame outputs as "a question + recommended answer." That is the directive-aid configuration by name. The honest restatement of the risk is therefore not atrophy; it is bias in a retained decision, and it is a property of the output format, not of anyone's habits.
The actual gap in the rule set is that nothing closes an outcome loop on a decision once it is made. Every verification mechanism RDCO owns points at the agent's work product: verify-vault-write, verify-strategic-output, verify-dispatch, verify-pdf-output, station-critic, the fresh-eyes rule, feedback_workflow_agent_output_integrity. Not one of them records what was recommended, what was chosen, and how it turned out. The convergent read across Casner, Chang et al., Endsley & Kiris and Kahneman & Klein is that judgment is maintained by outcome feedback in a regular environment, and specifically not by the count of decisions passing through. RDCO currently generates a high count of decisions and close to zero recorded outcomes on them. Bastani's GPT-Tutor-versus-GPT-Base contrast is the one empirically-supported lever: the harness determines whether capability survives, which means this is fixable in the rule set rather than by anyone changing how they spend their time.
Named amendment, one line, accept or reject: amend feedback_advisor_not_pair_programmer so that every DECISION NEEDED item Ray surfaces is logged to a decision record carrying the recommendation, the strongest case against it, the option chosen, and a check-back date - and Ray resurfaces the record at that date with what actually happened. This targets the mechanism the literature supports (feedback and depth, not volume), it adds the Sarkar-style provocation that the automation-bias numbers argue for, and the whole obligation falls on Ray's output format and Ray's follow-up, requiring nothing of the founder beyond the decision he was already making. A cheaper partial version, if the full record is unwanted: keep the recommendation but require the strongest case against it in the same message, and drop the check-back. If neither is accepted, nothing breaks - the arrangement remains on the safe side of the one measured threshold, and the residual risk stays where this brief found it, in bias rather than in decay.
On the second half of the question, the answer is that there is no practice evidence to copy. Willison, Ronacher, Klaassen, Karpathy and Yegge all converge on the same unmeasured intuition - move your attention to the ends of the loop and keep a verification step you personally own - and every one of them is writing an essay, not reporting a result. Willison's own admission that knowing when not to review every line "feels deeply uncomfortable" is the most honest statement in the set. Notably, no named operator has adapted decision journals, pre-mortems or any established decision-hygiene practice to AI delegation, which means the amendment above is not RDCO copying a known pattern; it is RDCO instrumenting something nobody has instrumented. That cuts both ways: no precedent to borrow, and no evidence it works either.
Why this is in the vault
This brief resolves the specific open question [[2026-10-30-every-marx-alienation-coding-agents]] left on the table, by finding that feedback_studio_throughput_no_deferral is not the rule the alienation argument reaches, and by relocating the risk from skill decay to automation bias in the one rule that actually sets the decision-surfacing format, feedback_advisor_not_pair_programmer. It is the evidence base for a single accept-or-reject amendment to that rule, and it supplies the calibrated citation set RDCO needs when the agents-in-production positioning touches operator-skill claims, where confident essays currently outnumber measurements by a wide margin.
Open follow-ups
- Has anyone measured the specific contrast of delegating execution while the principal decides against delegating decision while the principal monitors? Onnasch's moderator collapses Stages 3 and 4, so the question RDCO is actually asking is unmeasured. This is the highest-value gap in the brief.
- Does RDCO's reversibility classifier actually hold at the boundary it claims? The structural safety argument depends entirely on irreversible items never being auto-resolved.
feedback_automode_classifier_hard_gateand [[2026-10-23-rdco-auto-mode-classifier-miss-profile]] suggest this is empirically checkable against logs rather than assumable. - What is the measured defect rate on decisions Ray surfaces with a recommendation attached, versus ones surfaced without? This is the Skitka geometry applied to RDCO's own record, and it is answerable from history without any new instrument.
- Does Budzyn et al. (2025) replicate anywhere, and is there now a second field study of a judgment skill degrading under supervisory AI use? The colonoscopy result is currently a single observational before-after, and it is the closest thing to the mechanism this brief cares about.
- Does the research-backlog promotion path carry known citation corrections forward? This question reached the queue with a defect [[2026-10-08-finding-level-citation-audit-arxiv-gloss]] had already graded. A plumbing question, cheap to answer.
Related
- [[2026-10-30-every-marx-alienation-coding-agents]] - the parent piece that raised this as the vault's unanswered challenge to the throughput norm
- [[2026-09-08-every-to-read-or-not-to-read-the-code]] - Klaassen's countermeasures, and the note carrying the overstated 2608.23642 gloss
- [[2026-10-08-finding-level-citation-audit-arxiv-gloss]] - the audit that graded that gloss Overstated; the reason this brief does not rest on it
- [[2026-04-20-every-ai-autopilot-verification-decay]] - Sarkar et al. on reliability breeding under-checking, and the provocations design proposal the amendment borrows
- [[2026-10-08-verification-ratio-senior-signoff-throughput]] - the sublinear-review finding that defuses the throughput-volume objection
- [[2026-06-03-parallel-agent-fde-capacity]] - "attention depletes, trust inflates"; the vault's own statement of the bias mechanism
- [[2026-08-03-data-engineering-central-keeping-technical-skills-llm-age]] - Beach's five-practice regimen, the practitioner-hypothesis version of the same question
- [[2026-10-23-rdco-auto-mode-classifier-miss-profile]] - whether the reversibility boundary the safety argument depends on actually holds
- [[feedback_ic_vs_production_mode]] - RDCO's existing ad hoc version of the autonomy/supervision/manual split
Sources
Vault
~/rdco-vault/06-reference/2026-10-30-every-marx-alienation-coding-agents.md~/rdco-vault/06-reference/2026-09-08-every-to-read-or-not-to-read-the-code.md~/rdco-vault/06-reference/research/2026-10-08-finding-level-citation-audit-arxiv-gloss.md~/rdco-vault/06-reference/2026-04-20-every-ai-autopilot-verification-decay.md~/rdco-vault/06-reference/research/2026-10-08-verification-ratio-senior-signoff-throughput.md~/rdco-vault/06-reference/2026-08-03-data-engineering-central-keeping-technical-skills-llm-age.md- Memory rules:
feedback_studio_throughput_no_deferral,feedback_distinguish_decision_from_action,feedback_advisor_not_pair_programmer,feedback_ic_vs_production_mode,feedback_complete_then_ask_for_review,feedback_auto_mode_signal_to_noise,feedback_automode_classifier_hard_gate,feedback_workflow_agent_output_integrity
Read in full this pass
- Onnasch, Wickens, Li & Manzey (2014), Human Factors 56(3):476-488 - https://depositonce.tu-berlin.de/bitstreams/6a0b7962-cb5f-4317-9d07-ee11707dd51f/download
- Parasuraman & Manzey (2010), Human Factors 52(3):381-410 - https://depositonce.tu-berlin.de/bitstreams/cafd2873-814b-4c59-bab1-addd42e249d2/download
- Bainbridge (1983), Automatica 19(6):775-779 - https://ckrybus.com/static/papers/Bainbridge_1983_Automatica.pdf
- Arthur, Bennett, Stanush & McNelly (1998), Human Performance 11(1):57-101 - https://gwern.net/doc/psychology/spaced-repetition/1998-arthur.pdf
- Casner, Geven, Recker & Schooler (2014), Human Factors 56(8):1506-1516 - https://gwern.net/doc/technology/2014-casner.pdf
- Chang, Chen, Mellers & Tetlock (2016), Judgment and Decision Making 11(5):509-526 - https://doi.org/10.1017/S1930297500004599
- Mitchell, Ghosh & Passi (2026), "AI Agents Push Humans Out of the Loop," arXiv:2608.23642 - https://arxiv.org/abs/2608.23642 (position paper; no data)
- FAA PARC/CAST (2013), "Operational Use of Flight Path Management Systems" - https://www.faa.gov/sites/faa.gov/files/aircraft/air_cert/design_approvals/human_factors/OUFPMS_Report.pdf
- NTSB/AAR-14/01, Asiana 214 (2014) - https://www.ntsb.gov/investigations/AccidentReports/Reports/AAR1401.pdf
- Risko & Gilbert (2016), TiCS 20(9):676-688 - https://discovery.ucl.ac.uk/id/eprint/1508770/
- Bastani et al. (2024), "Generative AI Can Harm Learning" - https://nmdprojects.net/teaching_resources/bastani_generative_ai_can_harm_learning.pdf
- Lee, Sarkar, Tankelevitch et al. (2025), CHI '25 - https://advait.org/files/lee_2025_ai_critical_thinking_survey.pdf
- METR (2025) - https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
- Willison, "Vibe engineering" - https://simonwillison.net/2025/Oct/7/vibe-engineering/ ; "Uncomfortable" - https://simonwillison.net/2025/Oct/11/uncomfortable/
- Ronacher, "A Year of Vibes" - https://lucumr.pocoo.org/2025/12/22/a-year-of-vibes/
- Every, "You're the Bread in the AI Sandwich" - https://every.to/context-window/you-re-the-bread-in-the-ai-sandwich
- Yegge, Changelog & Friends #96 - https://changelog.com/friends/96
Abstract-only, secondhand, or blocked (flagged, not retried)
- Endsley & Kiris (1995), Human Factors 37(2):381-394 - paywalled; design/results via Endsley's own 1996 chapter, https://maritimesafetyinnovationlab.org/wp-content/uploads/2019/12/Automation-and-Situation-Awareness-Endsley.pdf ; N and effect sizes not retrieved
- Parasuraman & Riley (1997), Human Factors 39(2):230-253 - paywalled, abstract only - https://journals.sagepub.com/doi/10.1518/001872097778543886
- Parasuraman, Sheridan & Wickens (2000), IEEE Trans. SMC-A 30(3):286-297 - abstract only - https://ieeexplore.ieee.org/document/844354
- Kahneman & Klein (2009), American Psychologist 64(6):515-526 - paywalled, abstract only - https://pubmed.ncbi.nlm.nih.gov/19739881/
- Budzyn et al. (2025), Lancet Gastro Hepatol 10(10):896-903 - paywalled; abstract/repository only - https://wrap.warwick.ac.uk/id/eprint/191005
- Haslbeck & Hoermann (2016), Human Factors 58(4):533-545 - abstract only - https://pubmed.ncbi.nlm.nih.gov/27076096/
- Mellers et al. (2014), Psychological Science 25(5):1106-1115 - paywalled; per-intervention Brier deltas not retrieved; later reanalysis contests the teaming/training effects
- Dell'Acqua et al. (2023), HBS WP 24-013 - hbs.edu and SSRN returned 403; figures reported but not PDF-verified. "Falling Asleep at the Wheel" PDF 404 - https://innovationgrowthlab.org/node/1447
- METR Feb 2026 update post - could not be opened; the roughly -38% to +9% CI is secondary-sourced
- Shen & Tamkin (2026), arXiv:2601.20245 - figures via abstract/paper summary this pass
- Sparrow, Liu & Wegner (2011), Science 333:776-778 - publisher page paywalled; mirror at https://stafforini.com/works/sparrow-2011-google-effects-memory/
- Mitchell Hashimoto's alternation practice - third-party write-ups only, no primary transcript; not attributable as stated