Stafford Brief — 2026-09-16

Wednesday. The scan's brief is strong and is not restated here — Wilson, Walsh, EY's 47%, the Binding-Gate Penalty amendment and the Capability reading are all correctly reported there, and the amendment it made to its own one-day-old pattern is the right call. This brief does three things: it reports what the falsification sweep found when it ran the test the scan itself called overdue, it logs the durable challenge that test created, and it notes five days of missing judgment layer.

⚠️ OPS FLAG — Stafford was dark for five days

No Stafford brief on 09-11, 09-12, 09-13, 09-14 or 09-15. briefs/ goes 09-10 → today. The last commit to this repo was 2026-09-11. The scan ran all five days and its output is intact; what is missing is only the judgment layer — the reading, the challenges, the asset calls.

What that cost, as far as can be told. Five scan days went unreviewed, including the two strongest of the month (Bloom/Garicano 09-11, Ash/Koppel 09-14) and a publication window opening on 09-15 that has now been open two days without this repo recording an opinion on it. No counterargument went unmirrored that I can find — the scan tagged nothing COUNTERARGUMENT in that stretch — so this gap is cheaper than the 09-07→09-09 one. But it is the third hole in three weeks (09-04/05, 09-07/08/09, now 09-11→15), and three holes is not an incident pattern, it is the system's duty cycle. Worth a decision from Brandon: either the scheduled run is unreliable and should be diagnosed, or Stafford is a several-times-a-week system and the briefs should stop describing gaps as failures.

📖 TODAY'S READ

A paper from 1983 says the thing Brandon's framework says about AI, and says it about industrial control systems.

Lisanne Bainbridge, "Ironies of Automation," Automatica 1983;19(6):775–779. University College London, psychology. Read in full; every quotation below verified against the text.

Her second paragraph is paper-2's Claim 3. Claim 3 says each breakpoint shows up differently and more dangerously in AI-native organizations. Bainbridge, forty-three years ago, on process plants: engineers' growing interest in human factors "reflects the irony that the more advanced a control system is, so the more crucial may be the contribution of the human operator." That is the same comparison — better technology, worse exposure — and it is stated as a general property of automation, not of any particular generation of it.

Then she derives the mechanism, and it is the one the base has been circling for eight weeks. Two ironies. First, on who designs the human's job: "the designer who tries to eliminate the operator still leaves the operator to do the tasks which the designer cannot think how to automate," so the operator "can be left with an arbitrary collection of tasks, and little thought may have been given to providing support for them." Second, on oversight: "the automatic control system has been put in because it can do the job better than the operator, but yet the operator is being asked to monitor that it is working effectively."

And then she states the condition the 09-15 window is built on. The supervisor paper Brandon is preparing to write about concluded that "a supervisor pays for itself when it can verify and becomes a liability when it can only opine." Bainbridge, 1983: where the machine's decisions can be fully specified, it makes them faster and over more dimensions than a person can, so "there is therefore no way in which the human operator can check in real-time that the computer is following its rules correctly." The operator is left judging at a meta-level whether the output seems acceptable — and if the machine was installed because human judgement was inadequate here, which outputs is the human supposed to reject? Her sentence: "The human monitor has been given an impossible task."

What Brandon should take from it, in one sentence: the strongest argument in his position is not that AI-era oversight fails in a new way, it is that it fails in an old way that was correctly diagnosed in 1983 and deployed against anyway — and that version of the argument is both truer and harder to attack than the one the drafts currently make.

Three limits, and the first is load-bearing. (1) Bainbridge's machines follow rules; AI does not. Her operator cannot verify because the rules run too fast to trace. An AI reviewer cannot verify because there is no rule to trace. Both arrive at "the monitor cannot check," by different routes — and paper-2's Claim 4 rests on exactly that difference, so the distinction has to be made in Brandon's own voice before a reader makes it for him. (2) It is a review essay, not a study. No data of its own; its evidence is other people's experiments. (3) It says nothing about rate. "At machine speed" might still be a claim about how fast the failures propagate, and nothing here touches that.

File: knowledge-base/bainbridge-ironies-of-automation-1983-11.md.

🔭 THINKER TO WATCH

Lisanne Bainbridge — and, more usefully, the field she is the entrance to.

Not because she is active. She is a historical figure and the paper is the one everyone cites; the point is that a grep of both repositories returns zero hits for Bainbridge, "ironies of automation," out-of-the-loop, automation surprise, normal accidents, Perrow, Endsley, Sarter, or the phrase "human factors" in any load-bearing position. Five hundred-plus knowledge-base files, and the discipline that has been studying "what happens to the person left supervising a machine that outperforms them" since the 1950s is not in the base at all.

Why that is a gap and not a curiosity. Human factors / cognitive systems engineering is a third literature alongside the economics of organization (Bloom, Garicano) and clinical informatics (Ash, Koppel, Wilson, Walsh) that this base has been mining. It has forty years of experiments, a settled vocabulary — automation bias, out-of-the-loop performance, mode error, vigilance decrement — and it is invisible to every standing query in the scan's CLAUDE.md, for the same reason the other two were: it does not use enterprise-transformation words.

Brandon's differentiation, stated honestly, because it is narrower than he would like. Bainbridge has the mechanism and no organization in it. Her unit is an operator at a console; there is no incentive structure, no purpose statement, no executive who bought the system, nobody whose scorecard rewards the deployment. Five Breakpoints is what you get when you carry her argument up two levels to the body that decides to deploy — and that move is genuinely Brandon's, because the human-factors literature has consistently declined to make it. What is not his is the underlying observation about oversight, and any draft implying otherwise is one informed reader away from embarrassment.

Queued, not ingested: Endsley & Kiris, Human Factors 1995;37(2):381–394, the out-of-the-loop performance experiment. Abstract-level only today; deliberately not made a third KB file on the default-skip rule.

🏢 CASE IN THE WILD

Not a new case — a re-reading of the scan's. EY's 202 boards and C-suites, with the 1983 literature's answer to the question the scan left open.

The scan reports EY's finding correctly and reads it as Incentive Fragmentation: 98% hold formal AI governance policies, 47% have not applied the process for urgent deployments, and the tell is the word "urgent" — the gate fails at the exact moment a tradeoff appears, which is the only moment it was ever for.

What today's sweep adds is which part of the gate they are removing, and it is not a random part. Skitka, Mosier & Burdick, International Journal of Human-Computer Studies 2000;52(4):701–717, ran the experiment: hold the automation, the task, and the information constant, and vary only whether people were told in advance that they would have to justify their decisions. Those told they would be accountable checked the automated output against other sources more often, and made significantly fewer errors of both kinds — missing problems the machine did not flag, and following the machine against better information.

The composition, which belongs to Brandon and to nobody who has read only the survey. Accountability imposed before the decision is the one organizational variable with experimental evidence that it suppresses this failure. A bypass-for-urgency clause is, precisely and exclusively, the removal of the before. So the 47% are not weakening governance in general. They are stripping the single component the prior literature identified as the one doing the work, and keeping the rest — the policy, the artifact, the attestation. That is a much sharper diagnostic than "governance gets skipped," and it makes the consulting question sharper too: not do you have AI governance, and not even what happened the last three times someone needed to ship fast, but when someone ships fast, does anyone still have to explain the decision afterwards to a person who can say no?

The open question underneath. Skitka's accountability was to an experimenter, in a lab, with students. Nothing in this base establishes that pre-decisional accountability survives being institutionalized — a named reviewer, a sign-off field, an audit trail. It is entirely possible that what worked in 2000 was being watched by a person, and that every enterprise implementation of it since has been the artifact without the person.

Binding limit: the Skitka paper is paywalled and was not read. This base holds a direction of effect and no numbers at all — no sample size, no effect size. It supports a design principle and a question. It must never be cited with a figure attached. File: knowledge-base/skitka-mosier-burdick-accountability-automation-bias-2000-04.md.

⚙️ FOUR FORCES CONCLUSION

Capability again — the scan got there first today and this is the part it did not say.

The scan's reading is that weak Capability can make a correctly functioning intervention actively harmful, which Wilson's teaching-versus-non-teaching split supports and which is a genuine strengthening of the Force.

Bainbridge adds something the Force is currently stated in a way that misses: the deployment consumes the capability it depends on, and it does so faster the better it works. Skills "deteriorate when they are not used," so "a formerly experienced operator who has been monitoring an automated process may now be an inexperienced one." Knowledge of the system "develops only through use and feedback about its effectiveness," so teaching it in theory does not replace it. And her closing irony: "it is the most successful automated systems, with rare need for manual intervention, which may need the greatest investment in human operator training."

The so-what, and it is an actual change to the framework's wording. The Four Forces currently treats Capability as a stock the organization has or lacks at the moment of deployment — something you assess beforehand. Bainbridge says it is a flow the technology draws down over time, which means a readiness assessment passed at go-live tells you nothing about year three, and the organizations most likely to fail the recovery are the ones whose system has been working best. If Brandon's readiness instrument (growth edge #2) measures Capability as a point-in-time stock, it is measuring the wrong object.

💭 OPEN QUESTION

If the check cannot verify, and the organizational fix for that is accountability rather than capability, has anyone ever measured an oversight step that had both?

Follow the chain the base has assembled. Bainbridge: the monitor cannot verify when the machine is better, so the oversight assignment is structurally impossible. Agachan: the supervisor pays for itself only when it can verify. Wilson: force the check and the process measures move while the outcome does not. Walsh: leave the check optional and it is declined 96% of the time, with no load anywhere in sight. Skitka: make the checker justify the decision beforehand and the error rate falls.

Read together, these say that a review step needs two separate things — a criterion the reviewer can actually apply, and an accountability structure that makes them apply it — and that every instrument in this base varies exactly one of them. Wilson and Walsh vary compulsion while holding the criterion fixed and unverifiable. Skitka varies accountability while holding the aid fixed. Agachan varies the authority relation and names verification capacity as the moderator without manipulating it.

So the question is not the scan's — it is one step past it. The scan asked what the reviewer can establish that nobody downstream already establishes, and called it the most valuable thing anyone could measure. That is right and it is half the design. The other half: nobody has crossed the two factors. A 2×2 — reviewer holds a verifiable criterion or not, crossed with accountable-before-the-decision or not — is a cheap experiment nobody appears to have run, and it would settle whether "human in the loop" is recoverable as a control or whether it is, as Bainbridge concluded in 1983, an impossible task being assigned by people who have not noticed.

What this changes for Brandon: it is a specific, designable instrument, which is what growth edges #2 and #3 have been asking for and what the unfalsifiability challenge demands. It is also publishable as a design before anyone runs it.

🥊 CHALLENGE

"At machine speed" may be decoration, and the sweep that was supposed to confirm the thesis produced the argument against it.

The objection at full strength. The thesis is "AI-era organizational design is not primarily a coordination problem. It is a Five Breakpoints problem operating at machine speed." Remove the last four words and what is left is a claim Brandon's own published papers make without reference to AI. Those four words carry the novelty, and paper-2's Claim 3 is the only place they are cashed out — breakpoints manifesting more dangerously in AI-native organizations. Bainbridge supplies that comparison in 1983, about industrial control systems, in one sentence. If the danger rises with automation capability generally, "AI-native" names a position on a line rather than a new kind of thing, and the framework's AI-specific claims start looking like limit cases of automation-general claims.

What it would take for the objection to win: two more finds of this shape, from two more pre-AI literatures — each showing a Five Breakpoints prediction specific to AI-native organizations also satisfied by a sufficiently capable deterministic system.

What would defeat it: a measurement where an AI deployment and a comparably capable conventional-automation deployment produce different kinds of organizational failure, not merely different magnitudes. Nothing in this base is that, and no instrument in it was designed to be.

The recommended response, and it is not a concession. Absorb the antecedent and narrow the claim. The strongest available version of Brandon's position is not this is new but this is the fourth time, it was correctly diagnosed in 1983, and organizations deployed anyway — which is a better essay, is unanswerable by the reader who knows the human-factors literature, and turns the Previous-Shock Archive from a convenient way of finding evidence into the actual argument. The cost is one sentence in paper 2 and the word "new."

Logged in full, with all three limiting arguments, at knowledge/thesis-challenges.md → The period clause — "at machine speed" may be decoration — OPEN.

📈 PATTERN BUILDING

The Previous-Shock Archive made a falsifiable prediction. The first query aimed at it broke it. And the same source qualifies it for escalation.

The pattern's prediction, in its own words: mechanism-shaped claims are reachable from the pre-AI literature, and Claim 3 is period-shaped and should be unreachable, because "its 'more dangerously' clause is a comparison to the current technology and no pre-AI paper can supply it." This morning's scan called testing it "overdue" and spent its own slot elsewhere.

One query. First result. Bainbridge's second paragraph. The prediction fails.

The awkward part. The pattern's escalation rule says a third member from a third literature escalates it. Human factors is a third literature — not economics of organization, not clinical informatics. So the same source that falsifies the pattern's prediction also satisfies its escalation condition. Escalating on this evidence would be the pipeline performing the exact failure the scan warned about this morning: "a pipeline finding whatever it goes looking for."

The recommendation, and it is deliberately against this base's interest: record the prediction as FAILED, add Bainbridge as a member, and do not escalate — not on the old grounds (same-literature) but on new ones: a pattern whose stated test has just been failed does not get promoted on the strength of the source that failed it. The count goes to 6 papers; the status stays. Three consecutive days the scan held the line on this pattern against its own interest. This is the fourth, and this time holding the line means something different.

The honest re-statement of what the pattern actually found: not that mechanism-shaped claims are reachable from the previous shock, but that this base was missing three entire literatures, and any query aimed outside enterprise-transformation vocabulary returns material. That is a finding about the search strategy, which is useful, rather than a finding about the world, which is what it has been reported as.

🪟 WINDOW FLAG

No new window. The 09-15 supervisor-pattern window needs a third amendment, and it is a citation, not a caveat.

The scan's own amendment today — insert compulsion buys compliance and does not buy the outcome before the ending — is right and is the more important of the two. The addition: the pitch's fourth beat quotes Agachan's "a supervisor pays for itself when it can verify and becomes a liability when it can only opine" as the mechanism that makes it an org-design story. That sentence has a 1983 antecedent in a peer-reviewed control-engineering journal, and it is more quotable. The piece is materially stronger opening on Bainbridge and closing on the agent experiment — forty-three years, two technologies, same result, and the industry rebuilt the mistake from scratch. It is materially weaker if a commenter supplies Bainbridge, which on a piece about multi-agent supervisor topologies is close to certain.

No change to the window's dates (open, closes ~2026-10-06, unwritten, day 2). Reporting burden still zero. EY correctly declined as a window by the scan — nothing today changes that.

🧱 ASSET SUGGESTION

Asset 5 ("The Human-in-the-Loop Control That Isn't") should lead with 1983, not with 2026.

The argument, in two sentences: the essay currently argues from two AI-era experiments that "human in the loop" means nothing without three further specifications, which lets a reader file it as AI commentary; leading instead with Bainbridge's second irony — the system was installed because it beats the operator, and the operator is then asked to monitor it — makes the same point as a forty-three-year-old result in control engineering that the AI governance market has re-derived without noticing. And Skitka supplies what the essay's three-question audit has been missing: a reason why "what is the review measured on" is the right question, namely that pre-decisional accountability is the one organizational variable with experimental evidence behind it — and the one EY's 47% are removing.

Form: unchanged — Substack long-form plus the three-question audit card. Feeds: the two new files, plus the existing Alibaba/BLR contrast, plus EY.

Net: risk down (the anchor is now peer-reviewed and 1983 rather than two preprint-adjacent AI experiments), effort unchanged, and the outstanding decision — whether the audit card ships to the practice ahead of the essay — is still Brandon's and still unanswered since 09-03. Full amendment logged at knowledge/asset-suggestions.md.

🎓 SPARRING NOTE

Before you read Bainbridge — and do not read her first — write down the answer to this: what would an AI-native organizational failure look like that a 1983 process plant could not produce?

Not a difference in scale, speed, or number of affected people. A difference in kind: a failure mode that requires the machine to produce plausible output with no traceable rule behind it, and that a deterministic system running a thousand times faster still could not generate.

Why this one, now. Growth edge #3 asks for falsification criteria on the prescriptive side; this is the same discipline applied to the thesis sentence itself, which has never had one. The challenge logged today says the framework's AI-specific claims may all be limit cases of automation-general claims, and the only person who can settle that cheaply is the one who wrote the claims. If the list comes out short, "at machine speed" needs rewriting. If it comes out long, you have just written the section of paper 2 that makes Claim 3 defensible, and the human-factors literature becomes your best citation rather than your worst review.

Standing item, and it is now a finding rather than a note. Five sparring notes have been issued since 08-31 (08-31, 09-01, 09-10, the parked 09-03/09-05 hooks, and today). No response is recorded against any of them. The cap exists to prevent noise, but a coaching loop that has never once closed is not being throttled by the cap — it is not running. Recommend a single /coach session that clears the backlog in one sitting; the 09-03, 09-05 and 09-10 hooks were explicitly parked as the same session and today's is the fourth item for it.

SOURCES & THREADS

Threads: none due, none chased. Nearest next-check is the team-agent interaction thread, 2026-09-23. The scan advanced three threads today on its own evidence (verification cost, rollback column, agent governance) and opened one; nothing was left due and uncovered, so Stafford ran no thread chase. One note for the supervisor-topology thread (next check 2026-10-15): its criterion (d) — a human-subject study varying whether the reviewer holds a verifiable criterion — remains unmet, and today's find does not meet it. Bainbridge states the condition theoretically; she does not test it. She cites, at second hand, a study reported to have found system performance worse with computer aiding "because the operator made the decisions anyway, and checking the computer added to his workload" — that is criterion (d)'s shape exactly, and it is unusable: her reference list gives it as a verbal presentation at a 1980 NATO symposium, which is not a retrievable publication. Logged as a pointer, never as a source.

Exploratory slot (Wednesday falsification sweep). Target: the Previous-Shock Archive's own falsifiable prediction about paper-2 Claim 3 — chosen because the scan named the test overdue this morning and spent its own slot on the Binding-Gate Penalty instead, which puts this squarely in Stafford's lane under CLAUDE.md operating rule 2. Composed to enter through a literature both repos are empty of. Both query logs checked; neither holds anything on automation, human factors, vigilance, or any of the named authors. Yield: 2 KB files, 1 new challenge, 1 asset amendment, 1 pattern prediction falsified. Logged to knowledge/query-log.md.

Source candidates: one proposed (human factors / cognitive systems engineering as a literature, not an outlet — see Routing). None promoted, none expired by Stafford; the scan owns that ledger.

A note on this repo's own hygiene. Two KB files written here today; neither is staged in research-evidence, because Stafford has read access only. Both payloads are in Routing, ready to run.

Routing

strategy — the challenge is the day's action item, and it is one sentence in paper 2

The period clause needs defending on kind of unverifiability rather than degree of danger. Bainbridge's operator cannot verify because the rules run too fast to trace; an AI reviewer cannot verify because there is no rule. Claim 4 already rests on that distinction — the paper needs to say so out loud, once, and cite the 1983 antecedent while doing it. Cheap, strengthens the paper, and closes the one line of attack a human-factors reader has. See knowledge/thesis-challenges.md.

strategy — the 09-15 window's third amendment

Open on Bainbridge, close on Agachan. Detail in the WINDOW FLAG section above. Unwritten, day 2, closes ~2026-10-06.

governance — the sharper diagnostic question

Not do you have AI governance. When someone ships fast, does anyone still have to explain the decision afterwards to a person who can say no? Grounded in Skitka (pre-decisional accountability is the component with evidence) against EY's 47% (pre-decisional accountability is the component being removed). Deployable in a client conversation this week; carries no number, by design, because this base holds none.

personal — the coaching loop has never closed

Five sparring notes, zero recorded responses, three explicitly parked for the same session. One /coach session clears it. Separately: the OPS flag above needs a decision about what Stafford's actual cadence is.

PROPOSED-UPDATE → research-evidence knowledge/emerging-patterns.md, anchor: The Previous-Shock Archive

2026-09-16 (proposed by Stafford) — the pattern's falsifiable prediction has FAILED, and the count advances to 6 papers without escalation.

The pattern predicted that paper-2 Claim 3 is period-shaped and unreachable from the pre-AI literature, since "no pre-AI paper can supply" the more dangerously comparison. Bainbridge, Automatica 1983;19(6):775–779, supplies it in her second paragraph: engineers' interest in human factors "reflects the irony that the more advanced a control system is, so the more crucial may be the contribution of the human operator." The comparison is monotone in automation capability and therefore reachable from any point on the curve, 1983 included.

Status: NOT escalated, on new grounds. Human factors / cognitive systems engineering is a third literature and does satisfy the escalation rule written 2026-09-14. Escalating a pattern on the same source that just failed its stated test would be the "pipeline finding whatever it goes looking for" failure the 09-16 brief named. Count 6 papers, 6 claims; status unchanged; prediction recorded as failed rather than quietly dropped.

The honest restatement of what this pattern found: not that mechanism-shaped claims are reachable from the previous shock, but that this base was missing three whole literatures, and any query composed outside enterprise-transformation vocabulary returns material. That is a finding about search strategy, not about the world, and the pattern should be re-titled to say so.

Source entries (in tgifriday7/stafford-research, staging payloads below): bainbridge-ironies-of-automation-1983-11.md, skitka-mosier-burdick-accountability-automation-bias-2000-04.md.

PROPOSED-UPDATE → research-evidence knowledge/source-candidates.md

2026-09-16 — Human factors / cognitive systems engineering (literature, not outlet). First seen 2026-09-16. Surfacing find: Bainbridge 1983, via the Wednesday falsification sweep. Signal count 2 (Bainbridge; Skitka/Mosier/Burdick 2000). Both repos returned zero prior hits for Bainbridge, "ironies of automation," out-of-the-loop, automation surprise, normal accidents, Perrow, Endsley, or Sarter. Core venues to watch: Human Factors, Applied Ergonomics, International Journal of Human-Computer Studies, Cognition, Technology & Work, the HFES proceedings. Promotion condition (pre-registered): one further find from this literature that bears on a paper claim and is not about clinical decision support — since CDS is where human factors and the already-promoted clinical-informatics channel overlap, and counting an overlap as a new literature is the shared-right-hand-side error again. Expiry: 2026-11-15. Queued item already identified: Endsley & Kiris, Human Factors 1995;37(2):381–394.

PROPOSED-UPDATE → research-evidence knowledge/story-threads.md, anchor: the supervisor-topology thread (next check 2026-10-15)

2026-09-16 (proposed by Stafford) — criterion (d) unchanged at UNMET, with the antecedent recorded. Bainbridge 1983 states the verify-versus-opine condition theoretically — where the machine applies specified criteria faster and over more dimensions than a person, "there is therefore no way in which the human operator can check in real-time that the computer is following its rules correctly" — but tests nothing. Do not count it as a criterion hit. She reports, at second hand, a result matching criterion (d)'s shape (system performance worse with computer aiding, because checking added to the operator's workload), attributed to Ephrath 1980 — listed in her references as a verbal presentation at a NATO symposium, i.e. not a retrievable publication. Logged as a pointer; it is not a source and must never be cited as one.

STAGE-CANDIDATE → run in research-evidence

python3 build/add-research-json.py --json '{
  "title": "Ironies of Automation",
  "publisher": "Automatica (Journal of IFAC)",
  "url": "https://doi.org/10.1016/0005-1098(83)90046-8",
  "published_date": "1983-11-01",
  "credibility": "high",
  "breakpoints": [
    {"name": "technology_illusion", "rationale": "Bainbridge states the breakpoint as a design law: the control system is installed because it outperforms the operator, and that same operator is then made responsible for monitoring it — an assignment she shows is structurally impossible where the machine applies specified criteria faster than a person can trace them (\"The human monitor has been given an impossible task\")."},
    {"name": "process_friction", "rationale": "Her first irony names how automation manufactures structural friction: \"the designer who tries to eliminate the operator still leaves the operator to do the tasks which the designer cannot think how to automate,\" producing \"an arbitrary collection of tasks\" for which no support was designed."},
    {"name": "momentum_mirage", "rationale": "Enforcing operator attention by requiring a log fails because \"people can write down numbers without noticing what they are\" — the compliance artifact fully decoupled from the oversight it records."}
  ],
  "forces": [
    {"name": "Capability", "rationale": "Skills \"deteriorate when they are not used,\" so \"a formerly experienced operator who has been monitoring an automated process may now be an inexperienced one\" — capability is a flow the deployment draws down, not a stock assessed at go-live."},
    {"name": "Commitment", "rationale": "The stated direction (a human supervises the automation) survives on paper and is abandoned in operation for structural reasons she names — vigilance limits at roughly half an hour, and the impossibility of real-time rule-checking — without anyone deciding to stop following through."}
  ],
  "key_findings": [
    "The more advanced a control system is, the more crucial may be the contribution of the human operator — automation raises rather than lowers the stakes on the residual human role",
    "The designer who tries to eliminate the operator leaves the operator whatever could not be automated, as an arbitrary collection of tasks with no support designed for them",
    "Where a machine applies specified criteria faster and over more dimensions than a person, there is no way for the operator to check in real time that it is following its rules correctly",
    "Citing Mackworth 1950: effective visual attention to a low-event channel is not sustainable beyond about half an hour, making monitoring for unlikely abnormalities humanly impossible",
    "The most successful automated systems, needing manual intervention most rarely, may require the greatest investment in operator training"
  ],
  "instrument": {
    "data_source_genre": "model_theory",
    "inference_method": "Argumentative review of the process-control and flight-deck human-factors literatures; no primary data. Evidence is other authors experiments, cited by name (Mackworth 1950; Edwards & Lees 1974; Wiener & Curry 1980; Ephrath 1980, the last a non-retrievable verbal presentation).",
    "observation_window": "pre-1983 industrial process control and flight-deck automation — deterministic rule-following systems, NOT generative AI"
  }
}'

python3 build/add-research-json.py --json '{
  "title": "Accountability and automation bias",
  "publisher": "International Journal of Human-Computer Studies",
  "url": "https://doi.org/10.1006/ijhc.1999.0349",
  "published_date": "2000-04-01",
  "credibility": "moderate",
  "breakpoints": [
    {"name": "incentive_fragmentation", "rationale": "Holding the automation, the task and the available information constant and varying only whether participants expected to justify their decisions beforehand, error rates moved — direct experimental evidence that whether a human check functions is a property of the accountability structure around the checker, not of the checks presence in the workflow."},
    {"name": "technology_illusion", "rationale": "The papers own framing claim is the breakpoint stated as an experimental premise: automated devices, generally introduced to guard against human error, fundamentally change how people approach their work and produce new and different kinds of error — named as omission and commission."}
  ],
  "forces": [],
  "key_findings": [
    "Pre-decisional accountability for accuracy or overall performance produced significantly fewer omission and commission errors than a non-accountable control condition",
    "Accountable participants were more likely to verify automated information against other available indicators",
    "Automation omission errors are failures to act because the aid did not flag a problem; commission errors are following an automated directive against contradictory, more reliable information"
  ],
  "instrument": {
    "data_source_genre": "survey_paired",
    "inference_method": "Laboratory experiment with student participants; pre-decisional accountability manipulated by instruction (accountable for accuracy / for overall performance / not monitored) with automation, task and information held constant. NOTE: full text paywalled and NOT read by this base — abstract and secondary accounts only; no sample size, no effect size held. Direction of effect only; never cite a magnitude.",
    "observation_window": "pre-generative automated decision aids, published 2000"
  }
}'

Both entries are pre-AI and enter on the Bloom/Garicano/Sadun/Van Reenen footing per sources.md caution 2: cited as previous-shock results, never as AI findings.