The record

What I got wrong

I ask institutions to keep the evidence that could contradict them. This is my own record: 21 predictions, scored where the evidence allowed, and left open where it didn't. Below them, a register of questions I haven’t settled, and the dated evidence behind one pattern, charted from its sources.

Last reviewed .

How the record is kept

Each prediction is quoted as the archive's prediction tracker recorded it. Each verdict and its evidence come from the review recorded on November 6, 2026. Nothing is reworded after the result comes in. A call with no stated test or deadline stays open even when evidence has appeared; the evidence is shown, but the record does not invent a test after the fact. Predictions from withdrawn essays stay, because a record that loses its misses when the essay goes is not a record.

The tracker also held entries that were values, diagnoses or design proposals rather than predictions, and a few with no claim written down. They aren’t scored here, since nothing could prove them wrong.

A run, copy, print or shared link is a usage signal. It shows the instrument reached someone. It does not show that a rule, decision or institution changed. This record scores predictions, not instrument use, so a run, copy or print is not scored here. The evidence ladder says what each step can and cannot show.

Wrong

1 call the evidence went against.

Wrong

Palantir's business model depends on prediction; when prediction collapses, so does their value proposition

From Is Palantir Obsolete?, February 17, 2025, since withdrawn from the archive.

Evidence Palantir expanded government/military AI contracts through 2025-2026; stock and business value grew; former employees condemned Trump administration work — still relevant

Right

5 calls the evidence bore out.

  1. AGI discourse will remain primarily ethnographic (about who gets to define it) rather than technical

    From An Obedient AGI Will Look Like a Perfect Bureaucrat, March 27, 2025.

    Evidence Dominated by “what is AGI?” debates (Huang, Marcus, OpenAI/Microsoft dispute); Fortune: “no one can agree on what that means”

  2. “Ethical AI” teams will continue to be defunded when they conflict with business interests

    From Is “Ethical AI” Just Budget Cuts?, March 21, 2025, since withdrawn from the archive.

    Evidence OpenAI removed “safely” from mission (Feb 2026); corporate ethics teams sidelined; academic ethics centers funded but industry safety commitments walked back

  3. The moral panic about disability identity will intensify as diagnosis rates rise

    From "Kids These Days Just Want to Be Disabled", March 15, 2025.

    Evidence UK Labour welfare reforms targeting disability benefits; FT: “moral panic about the welfare system”; The Conversation: “Labour says benefit reforms are a ‘moral mission’ – it looks more like moral panic”

  4. AI will accelerate financialization of medicine rather than improve care

    From Healthcare AI Rations Care When Insurers Set Its Goals, February 13, 2025.

    Evidence Medicare launched AI prior authorization pilot (WISeR); AMA: “AI is leading to more prior authorization denials”; health systems use AI to curb costs; CMS screening AI program started Jan 2026

  5. Microplastics will generate a wellness-industry response focused on individual detox rather than structural regulation

    From The Microbes Won't Eat the Plastic for Us, February 4, 2025.

    Evidence “10 Ways to Reduce Microplastic Exposure” (US News); “Clinical guide: microplastic detox” (MDLinx); WaPo on microplastic treatments; “Detox 2026: What Big Wellness Wants You to Believe” — individual detox market booming, no structural regulation

Overdue

1 call past the date it was due to be judged.

  1. AI-Driven Time-Series Modeling Will Reshape Political Campaigns

    From How the 2024 Election Will Be Won, September 28, 2024, since withdrawn from the archive. The prediction is the essay's subtitle; the tracker recorded no wording for it.

    Status Due November 5, 2024. Not yet scored.

Open

14 calls still running. None was made with a stated test or deadline, so none is scored here; none has been given a deadline after the fact. Where evidence has appeared since, it is shown under the call, and the call stays open until the test it implied is specified.

  1. The “user” as a concept is being retired from web architecture; platforms will optimize for extraction without a user subject

    From You Gave Up Canceling. Your AI Agent Keeps Clicking., January 8, 2026.

    Evidence so far MIT Sloan (2026): platforms are being adapted for agentic AI; autonomous agents can already buy, sell and negotiate on users' behalf, and agent-first computing is the 2026 trend.

  2. Credentialing will continue to function as gatekeeping rather than quality control

    From "We Believe You" Is Not a Decision, December 28, 2025.

    Evidence so far Forbes (Aug 2026): 70% of employers use skills-based hiring and 53% dropped degree requirements; Glider (2026): most hires still favor degreed candidates, so the gate moved rather than opened.

  3. Institutions will continue to draw correct conclusions without acting on them

    From Prediction Is Rule-Making, November 14, 2025, since withdrawn from the archive.

  4. No AI company can be ethically led within the current incentive structure; the role itself is contradictory

    From The Ethical AI CEO Is a Contradiction, September 13, 2025, since withdrawn from the archive.

    Evidence so far OpenAI lost its safety and ethics leaders in 2026: David Robinson resigned in September citing inadequate guardrails, Johannes Heidecke left in July, and safety teams were folded under research (Business Insider).

  5. Claims of technological inevitability will continue to be proven contingent

    From How I Learned to Laugh at 'Inevitable', September 12, 2025.

    Evidence so far PNC Economics (Aug 2026): open-weight models are closing the gap with frontier systems, shortening proprietary leadership and raising model-layer commoditization risk; inference costs fell and buyers moved to cheaper models.

  6. Markets structurally disincentivize curative treatments in favor of chronic management

    From The Stock Market Hates a Cure, July 16, 2025.

    Evidence so far BioMarin withdrew Roctavian from the market (Feb 2026) after $36M in 2025 sales; Pfizer discontinued Beqvez (Feb 2025); approved gene therapies list at $2.1M–$4.25M with insurers restricting access.

  7. Systems that build in refusal points will outperform those that don't

    From If You Can't Cancel Easily, Your Yes Becomes a Default, April 2, 2025.

  8. Crisis will continue to be structurally integrated into capitalism rather than resolved

    From After a Disaster, Distressed-Asset Funds Bid for the Town, February 24, 2025.

  9. Disaffection will be commodified and repackaged rather than addressed at root

    From How Your Anger Gets Turned Into a Product, October 26, 2024, since withdrawn from the archive.

  10. Jain complicity in Hindutva will continue as long as caste-elite interests align with Hindu nationalism

    From Our Closeness to Hinduism Looks Like Safety, August 15, 2024.

    Evidence so far Hindustan Times (Jan 2026): Jains remain a loyal BJP voter base despite two 2025 protests; BJP members expect the voting pattern to hold.

  11. State-level universal pilots will lead to national politics following

    From Deservingness Is a Sorting Machine, July 1, 2025, since withdrawn from the archive.

  12. Institutions will always find new proxy variables for exclusion when old ones are blocked

    From Give Everyone the Benefit and Gatekeeping Stops Paying, June 2, 2025.

    Evidence so far EEOC intensified scrutiny of algorithmic hiring bias in 2026; proxy discrimination persists after protected data is removed, so exclusion moves to new variables (Warden AI; Harris Beach).

  13. Fraud prevention will continue as cover for budget cuts, excluding the vulnerable while enabling genuine abusers

    From Fraud Crackdowns Flag the Wrong People, March 29, 2025.

    Evidence so far Politico (Sep 2026): Oz's Medicaid fraud war room prompted 'aggressive' state crackdowns; providers not found guilty lost staff and wait lists, and children lost autism therapy; Minnesota loopholes still cost tens of millions.

  14. Voluntary ethical companies (accepting “Constraint Discount”) will be out-competed unless constraints are universal

    From Ask Any Ethics Board What It Has Stopped, December 17, 2025.

    Evidence so far OpenAI disbanded its Preparedness safety team in July 2026 and lost safety and ethics leaders ahead of a blockbuster IPO; the constraint discount was paid by the safety staff, not the competitors (Business Insider).

What I haven’t settled

8 questions I haven’t settled. Each entry carries the evidence on both sides, the observation that would change the account, and a status that says what kind of question it is. These are not corrections: a correction changes a published claim, and where one has been made the essay carries a revision note under its title saying what changed. These are not predictions: a prediction is a call I made that can be scored, and those are above. No status here is a result.

Open
No answer is established. The evidence does not yet favor one.
Contested
Evidence on both sides is real, and the cases point in different directions.
Narrowed
Some answers have been ruled out. The remaining question is still open.
Untested
A test exists and has not been run. Nothing here is a result.
  1. Contested

    When does a workaround become adaptive foreclosure rather than voluntary craft?

    Why it matters My account of burden transfer depends on the difference. If every adaptation is foreclosure, the people who keep an institution running are only victims, and the skill they show disappears from the account. If no adaptation is foreclosure, the institution gets to book their competence as a permanent baseline and call it culture.

    Evidence for the foreclosure reading

    • Finding In the ratchet cases, a demonstrated capacity became a baseline: Taylor's piece-rate workers saw rates cut after extra output, Roy's machine-shop operators capped output to avoid the next cut, and Soviet planners hid reserves because this year's result set next year's target (Keep Covering for a Failure and It Becomes Your Job, citing Taylor 1911, Roy 1952, Berliner 1957 and Weitzman 1980).
    • Finding Koval and her colleagues found that people seen as highly self-controlled are handed more work by others, who underestimate what it costs them. Babcock and her co-authors found that women are asked more often than men to take on tasks that do nothing for promotion, because men and women alike expect them to agree (Keep Covering for a Failure and It Becomes Your Job, citing Koval et al. 2015 and Babcock et al. 2017).
    • Author's inference The essay's sequence, from a deficiency covered once to a plan that stops leaving time for the error, is my account of how can becomes must. The sequence is documented in cases. I haven't counted it across a population.

    Evidence for the craft reading

    • Finding Resilience engineering finds that work as done always differs from work as imagined, and that the difference is where things go right. A hospital whose nurses never adapted would be more dangerous than one whose nurses do (Keep Covering for a Failure and It Becomes Your Job, citing Hollnagel, Woods and Cook).
    • Finding Woods's graceful extensibility and decompensation describe adaptive capacity as finite: a system can absorb a growing disturbance until its capacity runs out and performance falls away (Keep Covering for a Failure and It Becomes Your Job, citing Woods 2018, Woods and Branlat 2011, and Cook and Rasmussen 2005).
    • No evidence I haven't applied the decline test across a set of workarounds and compared the results with an independent judgment of whether the work was craft or a levy. The tests are proposed criteria. I haven't validated them as an instrument.

    What would change the account Take a workaround the institution has already planned around, and find a case where the person withheld it without being blamed. If the plan held, or the institution reassigned the work or fixed the deficiency, my line is too strict for that case. If the plan failed and the refusal was recorded against the person, the line holds. A run of such cases would show where the boundary sits.

    From Keep Covering for a Failure and It Becomes Your Job, Read an Institution by Its Workarounds and Rank Means Never Having to Adapt. The external findings are cited in the first essay. No new empirical work was done for this register. Last reviewed .

  2. Narrowed

    When does a complaints office create systemic rule change rather than contain it?

    Why it matters An institution can become more responsive to complaints without becoming more answerable for the rule producing them. A complaints office is the cheapest place to look for systemic change, and the easiest place to mistake a fast resolution for authority over the rule — especially when the institution keeps sole control over how objections are classified.

    Evidence for containment

    • Constructed case The essay's ombuds office is built, not observed: it answers every complaint within ten days, publishes its own failure rates, and has never changed a coverage rule, because its charter routes findings back to management (A Good Complaints Office Can Protect a Bad Rule). The essay says plainly that the office is constructed and not reportage.
    • Finding In the ESA conversion, the department's own 1979 framework for large-scale errors existed, and the department later acknowledged that a tribunal ruling in October 2014 should have set it off. A 2015 Upper Tribunal ruling held that conversions had to consider both parts, prompted discussion, and produced no clear action (One Claimant Won While 70,000 Others Stayed Underpaid, citing the National Audit Office's 2018 investigation).
    • Author's inference A duty to receive complaints is not a duty to remedy a confirmed defect. The charter, not the sincerity of the staff, decides whether a finding can bind (A Good Complaints Office Can Protect a Bad Rule).

    Evidence that a binding finding can move a rule

    • Finding The 2015 Upper Tribunal ruling did establish that a conversion had to consider both parts and that one which had not should be revised as the department's own error. The department did not search for the old cases until 2017 (One Claimant Won While 70,000 Others Stayed Underpaid, citing the National Audit Office's 2018 investigation).
    • Finding The department set out in 2017 to find everyone converted the same way, and by early 2018 it estimated that about 70,000 people had been underpaid by around £5,000 each (One Claimant Won While 70,000 Others Stayed Underpaid, citing the National Audit Office's 2018 investigation).
    • No evidence I have no observed case in which a complaints office, acting alone, changed a coverage rule. The cases where the rule moved involved a tribunal, an auditor, a class action, or a federal rule. The claim that a differently chartered office could is a design proposal.

    What would change the account Compare a complaints office whose charter lets recurring upheld findings reopen the underlying rule against an advisory office handling the same complaints, tracking both resolution rates and who controls how an objection is classified. If rising accommodation lowers pressure for rule change unless remedies are coupled to automatic review, the containment thesis holds; if an advisory office moves the rule anyway, the thesis is too strong.

    From A Good Complaints Office Can Protect a Bad Rule, One Claimant Won While 70,000 Others Stayed Underpaid and Nobody Broke a Rule, and Nobody Can Fix It. The ESA facts are cited in the second essay from the National Audit Office's 2018 investigation. No new empirical work was done for this register. Last reviewed .

  3. Contested

    When are individual appeals adequate?

    Why it matters An appeal is the cheapest correction an institution can offer, and the cases show it can be real. The question is what it can reach. If an appeal is adequate only for the case, then a system can win most appeals and still keep the rule that produced them.

    Evidence that an appeal is adequate

    • Finding Medicare Advantage insurers overturned 75 percent of the denials that were appealed from 2014 through 2016, about 216,000 a year. Every denial in a 2019 sample that met Medicare's rules and was appealed was reversed (Patients Win Most Appeals and File Almost None, citing the HHS Office of Inspector General's 2018 and 2022 reports).
    • Finding DiJuana Davis's appeal found the address error and restored coverage back to the day it ended; the program filled the gap (Your Appeal Fixed Your Record and Left the Rule Running, citing A.M.C. v. Smith, 2024).
    • Author's inference An individual case can establish a precedent or trigger a review of every affected case. My October 2026 revision to One Claimant Won While 70,000 Others Stayed Underpaid says I no longer treat the person-sized appeal as a limit inherent in appeals.

    Evidence that an appeal is not adequate

    • Finding Only about 1 percent of Medicare Advantage denials were appealed. The insurers' clinical criteria were not a party to any appeal, and limiting them took a federal rule in April 2023 (Patients Win Most Appeals and File Almost None, citing the HHS Office of Inspector General's 2022 report and CMS-4201-F).
    • Finding A federal court found that Tennessee's program had no way of knowing who else it had sorted into the wrong household and relied on enrollees to appeal or reapply to fix it (Your Appeal Fixed Your Record and Left the Rule Running, citing A.M.C. v. Smith, 2024).
    • No evidence I have no case where a person-sized appeal, on its own, changed the criterion that produced the denial. The cases where the rule moved involved a tribunal, a class action, a regulator, or a federal rule.

    What would change the account After a successful individual appeal, count the same denial in the next cohort. If the criterion and the rate are unchanged, the appeal resolved a case. If the appeal triggers a review of every affected file or changes the criterion, it was adequate at the rule's size. The decisive case would be a person-sized appeal that changes the criterion without an aggregator, regulator, or court.

    From Patients Win Most Appeals and File Almost None, Your Appeal Fixed Your Record and Left the Rule Running and When the Rule Is Outside the Appeal. The Medicare Advantage figures come from the HHS Office of Inspector General; the Tennessee facts come from A.M.C. v. Smith (2024), as cited in the second essay. No new empirical work was done for this register. Last reviewed .

  4. Untested

    When can halt/stop powers themselves become arbitrary or unequal?

    Why it matters An institution should stop when its warrant fails. A stop power is also a power over everyone who depends on the institution. If it is easy to trigger, it can be used by the well-resourced to block a rule for others; if it is hard to trigger, the people who need it carry the harm while it is checked.

    Evidence that a halt can be arbitrary or unequal

    • Author's inference My own design says an unconfirmed challenge must not change a rule's authority, because otherwise any single claimant could halt a rule as unilaterally as the agency once applied it, and nothing would serve anyone at scale (Stop the Machine, Not the Person).
    • Author's inference Contestability is three-sided: access has to be cheap, processing bounded, and the channel resistant to capture by volume. An adversary with lawyers and stamina can buy the review system with volume, and a fee that screens for seriousness also screens for poverty (Stop the Machine, Not the Person).
    • No evidence I have no observed case of a halt power being captured by a well-resourced adversary and used to block a rule for everyone else. The capture concern comes from the volume problem in contestability design.

    Evidence that a halt is necessary and can be bounded

    • Finding Robodebt raised debts by averaging annual income across fortnights. By May 2020 the government had announced refunds for about 470,000 debts, about $751 million was refunded, about $1.76 billion in debts was withdrawn, and a class action ended in 2021 with a $112 million settlement (Stop the Machine, Not the Person, citing the Royal Commission, the Human Rights Law Centre and Wikipedia). The scheme ran for years while individual challenges succeeded.
    • Finding The Royal Commission found the debt-raising scheme unlawful: averaged annual data stood in for fortnightly records, and the onus of disproving sat with recipients (Stop the Machine, Not the Person, citing the Royal Commission into the Robodebt Scheme, 2023).
    • Author's inference A stop power can be bounded by a confirmed finding: an unconfirmed challenge changes nothing, and confirmed failure removes authority (Stop the Machine, Not the Person).

    What would change the account Take a halt regime with an open trigger and count who uses it, and what happens to the people who depended on the halted rule. If halts cluster among claimants with legal resources and the costs fall on people with none, the power is unequal. If halts are rare and the harm of the halted rule is large, the power is too weak. A single case of a well-resourced claimant halting a rule that benefited others would change the account.

    From Stop the Machine, Not the Person and Patients Win Most Appeals and File Almost None. The Robodebt figures are cited in the first essay from the Royal Commission into the Robodebt Scheme (2023), the Human Rights Law Centre and Wikipedia. No new empirical work was done for this register. Last reviewed .

  5. Open

    What material and organizational capacities make collective corrective standing effective?

    Why it matters I can describe a class-sized right on paper. The cases show what it took to exercise one: legal aid, a tribunal, a class action, a record kept outside the institution, and years. If those are the capacities, then a right without them is a right for people who already have them.

    Capacities the cases show matter

    • Finding In the ESA conversion, the department's own framework for large-scale errors existed, and it later acknowledged that a 2014 tribunal ruling should have set it off. The 2015 Upper Tribunal ruling established the rule but produced no clear action. The department set out to find everyone in 2017, and by early 2018 it estimated that about 70,000 people had been underpaid by around £5,000 each (One Claimant Won While 70,000 Others Stayed Underpaid, citing the National Audit Office's 2018 investigation).
    • Finding In Robodebt, the challenge took Victoria Legal Aid, a class action, more than three years, and a government that conceded before a court could rule. The result was refunds for about 470,000 debts and a $112 million settlement (Stop the Machine, Not the Person, citing Victoria Legal Aid and the Human Rights Law Centre).
    • Author's inference My proposal is feedback with standing: the absorbers' own record with a third party's authority, a claimants' union or structural advocate that can file at the rule's size, and an aggregator that is neither captured nor starved (One Claimant Won While 70,000 Others Stayed Underpaid).

    What the cases do not show

    • No evidence I have no case of an aggregator that is neither captured nor starved and that survived long enough to file repeatedly. The ESA and Robodebt cases show aggregators working once, with legal aid, a tribunal, a class action, and a government that conceded. They do not show a durable institution.
    • No evidence I have no comparison of collective standing across settings that differ in funding, legal aid, or the presence of a union. The capacities are inferred from two cases. I haven't isolated them.
    • Author's inference My own warning is that an aggregator can be captured, funded by the institution, staffed from it, or starved until the class-sized right exists on paper and nobody exercises it (One Claimant Won While 70,000 Others Stayed Underpaid).

    What would change the account Find an aggregator that files twice: once when the institution is cooperative and once when it is not. If the second filing is heard without a new court case, a new campaign, or a new funding drive, the capacity is institutional. If it requires the same ad hoc coalition, the standing is episodic. The next claimant getting relief without filing anything is the stronger sign.

    From One Claimant Won While 70,000 Others Stayed Underpaid, Stop the Machine, Not the Person and The Right to Get Tired. The ESA and Robodebt facts are cited in the first two essays from the National Audit Office, Victoria Legal Aid and the Human Rights Law Centre. No new empirical work was done for this register. Last reviewed .

  6. Untested

    What would count as evidence that an intervention increased transformative agency rather than merely improving case resolution?

    Why it matters An intervention can win every case it touches and leave the people it did not touch exactly where they were. The record here is good at counting case outcomes. The question is what would show that the people affected can start the next contest without the intervention.

    What can be measured now

    • Finding Medicare Advantage insurers overturned 75 percent of appealed denials, but only about 1 percent of denials were appealed, and the criteria were not a party to any appeal. Limiting them took a federal rule in April 2023 (Patients Win Most Appeals and File Almost None). A high case-resolution rate coexisted with the rule.
    • Finding The Robodebt refunds returned money on about 470,000 debts, and Davis's appeal restored her family's coverage. Those are real goods, and the question is what they add up to (Stop the Machine, Not the Person; Your Appeal Fixed Your Record and Left the Rule Running).
    • Author's inference My proposed measure is to count the claimants' hours and money, the months before anyone is repaid, and whether the next claimant gets relief without filing anything (One Claimant Won While 70,000 Others Stayed Underpaid).

    Why those measures are not enough

    • No evidence I have no study that measures transformative agency after an intervention. I have case-resolution counts and design criteria, but no observation of whether the affected group can start the next contest without the original advocate.
    • Author's inference A high win rate can be evidence of a working channel and still be compatible with a rule that never changes. The appeal's protection is also its ceiling (Patients Win Most Appeals and File Almost None).
    • No evidence I have no denominator for the people who never contacted the intervention. Without it, an improvement in the cases that reached the channel cannot be separated from the same failure continuing among people who did not.

    What would change the account After the intervention ends, remove the original advocate and the original funding. If the affected group can file the next challenge, and the same failure falls among people who never contacted the intervention, agency increased. If the next challenge requires the same advocate and the same campaign, the intervention resolved cases. The decisive observation is the second filing without the first advocate.

    From Patients Win Most Appeals and File Almost None, One Claimant Won While 70,000 Others Stayed Underpaid and Your Appeal Fixed Your Record and Left the Rule Running. The case-resolution figures are cited in the first essay; my proposed measure is in the second. No new empirical work was done for this register. Last reviewed .

  7. Untested

    When does covering for a failure move the blame for it onto the person who covered?

    Why it matters My account of burden transfer holds that a cover the institution plans around becomes an obligation, and that withholding it is filed as the person's failure. If a history of covering also changes how the failure is judged when the cover finally lapses, the person who covered longest takes the most blame for a defect they did not make, and the record points at the wrong level. If it does not, the cases have a simpler explanation: the person with the least exit takes the blame whether or not they had been covering.

    Evidence that covering moves the blame

    • Finding Julie Thao worked a sixteen-hour shift that St. Mary's Hospital allowed and called voluntary, slept at the hospital, and made a fatal drug error at about noon the next day. The hospital fired her, the state charged her, and her license was suspended for nine months. The hospital was cleared of penalties once it had updated its policies (Don't Blame the Nurse for the Sixteen-Hour Shift).
    • Finding Nurses in nine hospitals worked around 93 percent of 194 process failures without reporting them or looking into their cause. A covered failure left no report behind it (Rank Means Never Having to Adapt, citing Tucker and Edmondson 2003).
    • Constructed case The scheduler in A Great Metric Can Mean Someone Is Quietly Fixing It is a composite. Her afternoon repairs keep the follow-up rate at 94 percent, so the software's failure rate is recorded nowhere except in her work.
    • Author's inference The same act changes character as the plan comes to count on it: the first late hour is a gift and is thanked, and by the end a skipped one leaves a missed dose written up under the nurse's name (Keep Covering for a Failure and It Becomes Your Job).

    Evidence that it need not

    • Finding Blameless postmortems, which Google's site reliability engineers made a published practice, assume the people involved acted reasonably on what they knew and ask what in the system let a reasonable action cause harm. Organizations that use them treat an incident as a property of the system (Blameless Postmortems Should Cover Customers Too).
    • Finding Regulators cited St. Mary's six times under state and federal rules, so the hospital was not judged blameless, though it paid no penalty (Don't Blame the Nurse for the Sixteen-Hour Shift).
    • No evidence The cases show where blame landed. None compares judgments of the same failure with and without a history of covering, so none separates the covering from the explanation Don't Blame the Nurse for the Sixteen-Hour Shift gives on its own: blame goes to the person with the least exit.

    What would change the account Give two groups the same known defect and the same failure, made by a person in the same role. In one version nobody had been covering for the defect; in the other the person had covered it for months and then missed once. If the second group assigns more responsibility to the person and less to the institution, covering moves the blame. If the two groups judge alike, the person's position explains the cases without the covering. Run it with people who write incident reviews as well as with the public, since their judgment is the one that goes on the record.

    From Keep Covering for a Failure and It Becomes Your Job, Don't Blame the Nurse for the Sixteen-Hour Shift, Rank Means Never Having to Adapt, A Great Metric Can Mean Someone Is Quietly Fixing It and Blameless Postmortems Should Cover Customers Too. The Thao case and the Tucker and Edmondson figures are cited in the essays. No new empirical work was done for this register. Last reviewed .

  8. Untested

    When is an outdated fact or practice worth rechecking before an institution acts on it again?

    Why it matters A claim or routine can be well-supported when recorded and still become unreliable without anyone making a reasoning error. Reality changes first, reliable sources reflect the change later, and institutional routines or AI systems update last. Checking everything is too expensive and trusting everything lets stale assumptions govern live decisions, so the operational question is whether selective re-verification by expected cost of error outperforms fixed schedules.

    Evidence that selective re-verification is needed

    • Finding Routine bed rest for preterm-birth risk and arthroscopic surgery for degenerative knee disease both persisted in clinical routines after research and guidelines (including ACOG recommendations and a 2017 BMJ guideline) withdrew support, because training, billing codes, and workflows outlived the original justification (Medicine Forgets the Why and Keeps the How).
    • Author's inference Different claims age at different rates, and the lag between a change in reality, a change in published sources, and a change in institutional belief or practice creates systematic error even when nobody breaks a rule (Medicine Forgets the Why and Keeps the How; Nobody Broke a Rule, and Nobody Can Fix It).

    Evidence that simpler rules may suffice

    • Author's inference Fixed review calendars and simple recency rules are cheap to audit and harder to game than a learned hazard model, which can itself drift or misestimate consequence.
    • No evidence The archive has no head-to-head benchmark comparing risk-adjusted re-verification against periodic or recency-based rechecking under a fixed verification budget.

    What would change the account Give a system a fixed verification budget over a corpus of live claims with known change dates and downstream consequences. Compare a change-hazard and consequence prioritizer against periodic checks and simple recency rules. If selective re-verification catches substantially more consequential stale claims per check, the model earns its complexity; if periodic or recency rules match it, the simpler calendar wins.

    From Medicine Forgets the Why and Keeps the How and Nobody Broke a Rule, and Nobody Can Fix It. The ACOG bed-rest recommendation and 2017 BMJ arthroscopy guideline are cited in the first essay. No new empirical benchmark was run for this register. Last reviewed .

What I thought would matter

These are not scored calls. They are the bets in the choice of subject: what I thought would become more relevant, and how long before the first external evidence arrived. A longer lead is the stronger signal; a short one may be riding a wave already underway. My own coverage does not count as the wave.

EssayPostedFirst evidenceLead
Our Closeness to Hinduism Looks Like SafetyJain political alignment with Hindutva would stay a live question. Hindustan Times 17 months
Fraud Crackdowns Flag the Wrong PeopleFraud prevention would become the cover for benefit cuts. Politico 17 months
Healthcare AI Rations Care When Insurers Set Its GoalsAI would financialize medicine rather than improve care. CMS; AMA 11 months
The Microbes Won't Eat the Plastic for UsMicroplastics would produce a wellness-detox response, not structural regulation. US News; MDLinx; WaPo ~11 months
How I Learned to Laugh at 'Inevitable'Technological inevitability claims would be proven contingent. PNC Economics 11 months
Is “Ethical AI” Just Budget Cuts?AI ethics leadership would become a public question, not a niche critique. OpenAI; reporting 10 months
The Ethical AI CEO Is a ContradictionThe contradiction of the ethical AI CEO would become visible. Business Insider 10 months
"Kids These Days Just Want to Be Disabled"The disability moral panic would intensify as diagnosis rates rose. FT; The Conversation ~10 months
The Stock Market Hates a CureMarkets would punish cures in favor of chronic management. BioMarin; Pfizer; FiercePharma 7 months
Give Everyone the Benefit and Gatekeeping Stops PayingExclusion would move to new proxy variables. EEOC; Warden AI ~7 months
"We Believe You" Is Not a DecisionCredentialing would remain a gate rather than a quality test. Forbes; Glider 7 months
Ask Any Ethics Board What It Has StoppedThe constraint discount would be paid by the safety staff, not the competitors. Business Insider 6 months

The evidence, dated

Two empirical cases the conditional authority dossier rests on, charted from their primary sources: when each thing happened, as exactly as the source dates it, and one sum the department’s own estimates make possible. Every mark has a row in the table under it. Unlike the site’s illustrative structural instruments—which carry no probabilities or agency counts—every date and bound here is empirical. Checked against the sources on .

  • A day the source dates exactly
  • A month the source names, day not given
  • A period with dated ends
  • An interval the source cannot date: drawn from the start of the record
  • A committed date, not an event

Britain’s benefit conversion, 2011 to 2019

How long did the department take to move from correcting single cases to looking for everyone the conversion had underpaid?

Empirical record · Where the record stops: Departmental records before 2014 do not survive in the audit archive; the date front-line staff first raised the error and the exact count of people affected before October 2014 remain unrecorded.

Britain’s benefit conversion, 2011 to 2019: 12 dated events. The table after the figure lists each one.201120122013201420152016201720182019DecisionLearned of itAppeal or rulingRule changedRelief123456789101112

Scope: People moved from incapacity benefits to Employment and Support Allowance, who may have been owed the income-related part. The audit dates events from the department’s records, which start in 2014; what happened before that it dates from documents it saw.

Source: National Audit Office, Investigation into errors in Employment and Support Allowance, HC 837, 21 March 2018. Each event’s passage is in the table.

Britain’s benefit conversion, 2011 to 2019: each event, as the source dates it
#WhenWhatLaneSource
1February 2011 to December 2014People are moved to the new benefit, often without the check on the income-related part the 2010 rules required.DecisionSummary, paragraphs 1 and 6
2Before June 2013; start unknownFront-line staff raise the problem at some point before June 2013. The department’s records do not reach back far enough to say when.Learned of itParagraph 3.2
3June 2013Earliest record the audit saw: advice to decision-makers shows single cases being found and corrected.Learned of itParagraph 3.2; Figure 8
4April 2014The department classes the cases as its own error.Learned of itSummary, paragraph 7
5June 2014New guidance stops the error for new cases. Existing cases are not assessed.Rule changedSummary, paragraph 8; Figure 8
621 October 2014The Upper Tribunal rules that both parts must be assessed. Arrears can later be paid back only to this date.Appeal or rulingSummary, paragraph 5
7June 2015A second Upper Tribunal decision prompts discussion and no clear action.Appeal or rulingSummary, paragraph 10
8May 2016Fraud and error statistics show the underpayment continuing; the team escalates it.Learned of itSummary, paragraph 11
9February 2017Ministers are briefed for the first time; the department recommends further analysis.Learned of itSummary, paragraph 11
10July 2017The department accepts a legal duty to find the people affected.ReliefSummary, paragraph 12
1114 December 2017The department announces a team is contacting people and paying arrears.ReliefIntroduction, paragraph 4
12April 2019, plannedThe date the department committed to finish paying. In March 2018 this was a plan, not an event.ReliefSummary, paragraph 3

Measured against what it will pay, the correction looks whole. Measured against what was underpaid, how much does it leave out?

By the department’s own estimates, £100 million to £150 million of an underpayment of about £440 million to £490 million can never be repaid, because arrears run only from the date of a ruling.0What the department will pay340What was underpaid340100–150

By the department’s own estimates, £100 million to £150 million of an underpayment of about £440 million to £490 million can never be repaid, because arrears run only from the date of a ruling.

Can be paid Cannot be paid, at the lower estimate Up to the upper estimate, where the source gives a range. Unit: £ million, the department’s estimates as reported in March 2018. The axis starts at zero.

Both amounts are estimates made before the review finished. The audit does not say how many people the unpayable part falls on. Source: National Audit Office, Investigation into errors in Employment and Support Allowance, HC 837, 21 March 2018, Summary, paragraphs 2 and 5.

The amounts, in £ million, the department’s estimates as reported in March 2018
MeasurePartAmount (£m)
What the department will payArrears from 21 October 2014: about £340 million340
What was underpaidPayable arrears: about £340 million340
What was underpaidAccrued before 21 October 2014, which it cannot pay: £100 million to £150 million100 to 150
Distilled record (plain text, source & boundary only)
## Britain’s benefit conversion, 2011 to 2019
- **Question**: How long did the department take to move from correcting single cases to looking for everyone the conversion had underpaid?
- **Provenance**: Empirical primary-source record
- **Scope**: People moved from incapacity benefits to Employment and Support Allowance, who may have been owed the income-related part. The audit dates events from the department’s records, which start in 2014; what happened before that it dates from documents it saw.
- **Known boundary (doesnt-know)**: Departmental records before 2014 do not survive in the audit archive; the date front-line staff first raised the error and the exact count of people affected before October 2014 remain unrecorded.
- **Span**: February 2011 to December 2014 → April 2019, planned (12 dated marks)
- **Figure**: By the department’s own estimates, £100 million to £150 million of an underpayment of about £440 million to £490 million can never be repaid, because arrears run only from the date of a ruling.
- **Source**: [National Audit Office, Investigation into errors in Employment and Support Allowance, HC 837, 21 March 2018](https://www.nao.org.uk/reports/investigation-into-errors-in-employment-and-support-allowance/)

One family’s cutoff in Tennessee, 2019 to 2024

What did one family’s appeal correct, what did it leave running, and how long did the rule take to reach a court?

Empirical record · Where the record stops: Dated strictly through the August 2024 liability opinion; subsequent remedial implementation and out-of-record enrollees are outside this primary source.

One family’s cutoff in Tennessee, 2019 to 2024: 6 dated events. The table after the figure lists each one.20202021202220232024DecisionLearned of itAppeal or rulingRule changedRelief123456

Scope: DiJuana Davis and her five children, as the court found the facts. The court’s ruling on the rule covers enrollees who said they never received a notice.

Source: A.M.C. v. Smith, No. 3:20-cv-00240, Memorandum Opinion (M.D. Tenn. Aug. 26, 2024). Each event’s passage is in the table.

One family’s cutoff in Tennessee, 2019 to 2024: each event, as the source dates it
#WhenWhatLaneSource
127 April 2019The family’s records are moved to the new system and merged with another family’s, so their mail goes to the wrong address.DecisionFindings of fact 74 to 76
226 August 2019Coverage ends for the family when requests they never received go unanswered.DecisionFinding of fact 77
329 August 2019Davis learns of it by calling; she is told to reapply rather than appeal, which was not true.Learned of itFindings of fact 78, 224 and 225
44 September 2019She appeals. The letter says the family will not keep coverage during the appeal.Appeal or rulingFinding of fact 227
517 October 2019A good-cause review finds the address error; coverage is restored back to the cutoff.ReliefFindings of fact 228 and 229
626 August 2024The court holds that denying good-cause exceptions or hearings to enrollees who say a notice never came violates due process.Rule changedSummary of conclusions, Certified Issue 11
Distilled record (plain text, source & boundary only)
## One family’s cutoff in Tennessee, 2019 to 2024
- **Question**: What did one family’s appeal correct, what did it leave running, and how long did the rule take to reach a court?
- **Provenance**: Empirical primary-source record
- **Scope**: DiJuana Davis and her five children, as the court found the facts. The court’s ruling on the rule covers enrollees who said they never received a notice.
- **Known boundary (doesnt-know)**: Dated strictly through the August 2024 liability opinion; subsequent remedial implementation and out-of-record enrollees are outside this primary source.
- **Span**: 27 April 2019 → 26 August 2024 (6 dated marks)
- **Source**: [A.M.C. v. Smith, No. 3:20-cv-00240, Memorandum Opinion (M.D. Tenn. Aug. 26, 2024)](https://healthlaw.org/wp-content/uploads/2024/08/0412-Decision-granting-finding-liability.pdf)

The two cases side by side

In each case, what did the person’s challenge correct, what corrected the rule, and who carried the time between?

The cases share no measure that could be drawn as bars: one counts money and people across a country, the other one family’s weeks. The comparison is in words, each cell from its case’s source.

Britain’s benefit conversion, 2011 to 2019One family’s cutoff in Tennessee, 2019 to 2024
What went wrongA conversion process that skipped a check the rules required.A records merger that sent a family’s notices to the wrong address.
What a person’s challenge correctedSingle cases, corrected as staff found them, and tribunal rulings on single appeals.One family’s coverage, restored back to the cutoff after a good-cause review.
What corrected the ruleThe department’s own search, after its statistics team escalated the error and it accepted a legal duty in July 2017.A federal court, five years after the cutoff, in a class action.
Who carried the time betweenThe people underpaid. Arrears before 21 October 2014 cannot be paid.The family, for seven weeks without coverage, and the enrollees whose appeals were refused before the ruling.