About the framework
An engineering discipline for keeping automated decisions answerable
Institutions absorb the evidence that their automated systems are wrong. A worker quietly fixes the error. An appeal settles one case. The rule never changes. Ethotechnics names that mechanism, scores it in public cases, and specifies the records that carry the evidence back to the rule in any institution that adopts them.
Where a dashboard reports accuracy or efficiency as a property of the system, the casebook and the standards show it as a distribution among people: who carries the errors, who holds the authority, and who can make the institution respond.
One question tests most of it: would this institution still look successful if the people around it stopped saving it from itself? Where the answer is no, difficulty is living in the wrong place, and where difficulty lives is a design decision.
Put formally, the discipline keeps six things tied together as a system grows: what it can do, what it is permitted to do, the evidence behind the permission, how hard it would be to do without, who can challenge it, and who can halt it. When they come apart, a system that gains ability becomes power nobody can review. Those six are the safeguards the casebook scores. The framework is written by Kanav Jain as part of the institutional error-accounting research program, drawing upon the qualitative diagnostics developed across The Crumple Zone, and published openly as standards, a glossary, and diagnostics.
What it tests, and what it leaves open
From a complaint to a record an institution can breach
Complaints about an automated system start with the people it decides about and the staff who clean up after it, and an institution is free to treat a complaint as an anecdote. The standards restate each one as a record with an owner and a clock. A nurse's "we're constantly rescuing this system" becomes a measured count of the correction staff absorb. A claimant's "nobody listens when it gets me wrong" becomes standing: the right to a challenge that must be received, answered by a date, and able to change the decision.
The same records test what an institution says about itself:
| If you call the system… | …show the record |
|---|---|
| democratic or legitimate | Show who is exposed to its decisions, who can challenge them, whose challenge must be answered by a date, and who can force a reconsideration or a halt. Asked for by Law VII STD-02 §8.4 STD-07 §4.2 |
| efficient | Show that the figure counts the work the system pushes onto people with less power: the nurse fixing a scheduler's mistakes, the claimant proving a denial wrong. Asked for by STD-01 §7.1 Burden concealment evals |
| accurate | Show accurate for whom: who bears its false positives, its delays, and its denials. Asked for by Burden distribution evals |
| responsive | Show the challenges it upheld that changed a rule, not only a case. Asked for by Exception learning Corrective learning evals |
| authorized | Show the authority grant: who signed it, on what evidence, and when it ends. Asked for by STD-08 |
| chosen, not imposed | Show what it costs the person who depends on it to leave. Asked for by Law V Dependence runs both ways |
The restatement has a limit. It can turn "I am bearing this error" into standing against a decision. It cannot settle "why should this institution hold this power over me at all?" An election can authorize a program; it does not give the person the program decides about a way to challenge that decision. The standards force a named decision. They do not make it, and the choices built into them, such as who receives standing and how short a clock is, carry values of their own.
The method leaves the question of who should hold power out on purpose, so that people who disagree about the answer can demand the same records. The theory takes it up in deliberate non-use, democratic vs. coercive governability, and the compulsion problem.
What it is for
Relying on a system without being ruled by it
Institutions already rely on things they cannot fully inspect: colleagues, suppliers, instruments, the last shift's notes. That reliance is what makes an institution more than one person, and nothing here argues against it. What the framework is for is keeping reliance on a system from turning into subordination to it, in either direction. The institution should be able to depend on a system without losing the capacity to question, narrow, or replace it. The people the system's decisions land on should be able to depend on the institution without losing the standing to object and be answered.
Every law on the site is a constraint that keeps one of those two dependences from converting into command. Capability does not become authority; a quiet year does not become renewal; a captive population does not become a population that may be directed. The laws state the constraints, the axioms state what they are owed to and why, and the theory essays develop the argument. A reader can adopt the constraints without the ideal. The ideal is what they add up to.
The underlying project
Power that stays correctable, because every institution gets things wrong
Without the technology, the claim is about who bears a system's failures and what keeps its power from becoming final. The chain is short. An institution has to model the world to act on it. Every model leaves something out. Someone carries what it leaves out, and who that is gets decided by power. Power dominates when it pushes those costs onto others and cannot be corrected by them.
Domination, here, is power shielded from consequential contradiction. The institution decides what counts as the facts, acts on them, pushes the leftover costs down, decides what counts as evidence against itself, and cannot be made to revise by the people it gets wrong. The case a model leaves out is evidence about the model, not noise. And a system is not open to correction if the people who know it is wrong must ruin themselves to prove it.
No arrangement is innocent by design. Markets, states, democratic majorities, experts, and participatory processes all face the same question: when it gets someone wrong, what happens? The five concerns of the design discipline follow from that, and each is a working part of this site: jurisdiction (what the institution may decide at all), representation (what it models and how that can be challenged), residual custody (who carries what it cannot represent), corrigibility (whether error travels back into the rule), and non-finality (whether its decisions stay reopenable).
The question, in one line: how can an institution be strong enough to act without becoming strong enough to define away the evidence that it is wrong?
Institute and commissioned work
How the Institute and commissioned work fit together
Ethotechnics Institute is the public, open-access commons: standards, glossary, diagnostics, and field notes are free to use, cite, and extend.
Kanav Jain, who wrote this framework, does commissioned safety evaluation for healthcare AI: a one-hour safeguards review, a two-week readiness sprint on one workflow, an investor diligence memo in five business days, or a two-to-three-week evaluation against FDA and EU AI Act expectations. The Institute publishes the open framework; commissioned work applies it to specific systems, for health plans, healthcare AI companies, and investors.
For scholars
The values exist. The failure mechanics do not.
Contestability, meaningful control, oversight, and resilience are settled commitments in the neighboring fields. None of them comes with failure mechanics — a way to state when each value decays, what to measure, and who would have to act.
The scholarly crossings note sets this out three ways. Nine pairings match one scholar's concept to the instrument here that measures it. Eight completions take a mechanism the framework does not replace, such as markets or exit, and add the question its own theory does not ask: does it still correct after ten years of success? Five counterpositions state the strongest opposing views so their holders would accept them, each with the empirical question that would settle it.
Intellectual lineage and distinctive contribution
Where the argument comes from, and where it departs
Ethotechnics does not claim originality for the proposition that human struggle should be reduced, or that institutions exploit unpriced labor to absorb their own failures. Those insights belong to established traditions:
- Post-work socialism and Marx: The distinction between the realm of necessity and the realm of freedom, establishing that technical productivity should reduce compulsory labor rather than intensify it.
- The social model of disability: Shifting attention away from demanding extraordinary individual coping toward altering disabling environments and institutional barriers.
- Relational egalitarianism (Elizabeth Anderson): The principle that individuals are entitled to the social conditions of equal standing and ordinary dignity without having to prove exceptional merit or endurance.
- Feminist care ethics and social reproduction theory (Joan Tronto, Nancy Fraser, Eva Kittay, Alison Reiheld): Analyzing how formal economic success depends on unpriced care labor and identifying "privileged irresponsibility": the systemic offloading of maintenance work onto others.
- Administrative burden theory (Pamela Herd, Donald Moynihan): Empirical measurement of the learning, compliance, and psychological costs imposed on citizens by state and corporate processes.
- Republican political theory (Philip Pettit): Freedom as non-domination, demonstrating that making lives easier does not make people free if they remain subject to unchecked, arbitrary institutional discretion.
- Resilience engineering (David Woods, Erik Hollnagel, Richard Cook): Describing how human operators constantly adapt under resource constraints to bridge the gap between "work-as-imagined" and "work-as-done."
None of these moral propositions is individually original to Ethotechnics. Nor is the mechanism at its center. The fields closest to it each see part of it and stop at a different point:
| Field | What it sees | Where it stops |
|---|---|---|
| Safety science Charles Perrow, Normal Accidents (1984); Diane Vaughan, The Challenger Launch Decision (1996); Sidney Dekker, Drift into Failure (2011); Nancy Leveson, Engineering a Safer World (2012) | Warnings get reclassified as normal until they stop registering, and systems drift toward failure in steps that each look reasonable. Tightly coupled, interactively complex systems propagate an upstream error faster than the people downstream can chase it. | It studies an organization's own operators and its rare catastrophes. Its remedy is a learning culture, not a duty owed to the people a system decides about. An engineering tradition |
| Responsibility in automation Raja Parasuraman and Dietrich Manzey, "Complacency and Bias in Human Use of Automation" (2010); Madeleine Clare Elish, "Moral Crumple Zones" (2019) | When an automated system fails, the blame lands on the nearest human operator, and trust in the automation grows precisely as vigilance over its output decays. | It follows the blame after a failure. It does not follow the evidence that the operator's everyday corrections keep from the rule. Dependence without standing |
| Hidden labor Mary L. Gray and Siddharth Suri, Ghost Work (2019); Hamid Ekbia and Bonnie Nardi, Heteromation (2017) | "Automated" systems run on human work that goes unrecorded and is paid little or nothing. | It treats that work as a question of pay and recognition. It does not treat it as error signals the institution never receives. Absorption as concealment |
| Administrative burden Pamela Herd and Donald Moynihan, Administrative Burden (2018); Cass Sunstein, Sludge (2021) | The learning, compliance, and psychological costs a process imposes on the people it serves, with the waiting itself acting as a rationing device. | It measures the cost. It does not trace how the people paying it keep the rule that causes it from changing. What outcomes hide |
| Due process and contestability Danielle Keats Citron, "Technological Due Process" (2008); Kars Alfrink and colleagues, "Contestable AI by Design" (2023); GDPR, Article 22 (2016) | A person decided about by a machine is owed notice, a hearing, and a way to contest the decision. | It attaches the remedy to one decision. The individual appeal then becomes the channel that absorbs the evidence: each decision is reversed and the rule stays. Exception learning |
| AI risk frameworks EU AI Act (2024); NIST AI RMF (2023); ISO/IEC 42001 (2023) | Risk assessed before deployment, human oversight required for high-risk systems, and the system monitored after it ships. | They assess a system before it runs and again after a major change. Monitoring after release is left to the organizations responsible for the system, which then produce the evidence that it still complies. Endogenous authorization |
| Republican political theory Philip Pettit, Republicanism (1997) | Power that is unchecked and arbitrary dominates people even when it is kind. | It looks for a person or office that holds arbitrary power. It has no account of power spread across systems, where no one acts arbitrarily and the people affected have no way back in. Democratic vs. coercive governability |
| Social reproduction theory and care ethics Nancy Fraser, "Contradictions of Capital and Care" (2016); Joan Tronto, Moral Boundaries (1993) | Paid work and formal institutions depend on unpaid care, the work that keeps people fed, rested, and able to cope, and they use it up without counting it. | It describes that drain across a whole economy. It does not ask a particular system to record how much of a person's time and stamina it uses up. The consumption of adaptive capacity |
| Social-cost economics and environmental justice K. William Kapp, The Social Costs of Private Enterprise (1950); Marilyn Waring, If Women Counted (1988); Robert D. Bullard, Dumping in Dixie (1990); Jason W. Moore, Capitalism in the Web of Life (2015) | Costs a firm does not pay are shifted onto workers, households, places, and the future as a regular feature of how accounts are kept. National accounts leave unpaid household work outside production, and hazards settle on the communities least able to refuse them. | It explains why costs land where they do and argues for pricing or regulating them. It does not specify the record an institution keeps of the saving it retains, or the procedure by which the people carrying the cost can send it back. Insulation |
| Labor process theory and organizational sociology Michael Burawoy, Manufacturing Consent (1979); Arlie Russell Hochschild, The Managed Heart (1983) | Workplaces take emotional work and running repairs from staff without crediting them, and win their consent to it. Coping with the job's defects comes to look like a personal virtue. | It criticizes what that does to workers. It does not show how their unrecorded fixes keep the institution's own figures from showing the defect, or ask the institution to record the work. The human subsidy to institutional continuity |
Read together, the stops trace one gap. Safety science follows the lost signal, but inside the organization and after a catastrophe. Due process follows the person, but only as far as their own case. Labor and burden studies count the cost, but not the signal the cost hides. Risk frameworks monitor, but leave the monitoring to the operator. None asks what an institution owes when the person who absorbed the error is outside it, the harm is routine rather than catastrophic, and the rule itself has to be corrected.
Ethotechnics starts from that question. Its claim: institutions absorb the evidence that their automated systems are wrong, and the absorption keeps the error from anyone who could change the rule. The essay on insulation states it in full.
Why the question is urgent now:
- Scale. When Robodebt was automated, debt notices went from about 20,000 a year to 20,000 a week. At that volume, settling errors one case at a time cannot keep up with the rule that produces them.
- Records. A decision log now costs almost nothing to keep. An obligation stated at the level of the record, such as a grant's expiry date or an objection's answer, is a reasonable thing to ask of any operator.
- Law. The EU AI Act requires human oversight of high-risk systems. The GDPR requires a way to contest a solely automated decision with legal or similarly significant effects, where the decision is necessary for a contract or rests on the person's explicit consent. Neither law states a test a running system must pass to show the oversight or the contest works.
The theory's names for the mechanism:
- Capacity turned into entitlement: An institution notices that people can cover for its deficiencies, builds its operations on the assumption that they will, and turns a one-off act of endurance into a standing requirement.
- Non-expropriation of resilience: That people can adapt to a deficiency does not entitle the institution to their adaptation. When a fixable deficiency is routinely covered by unrecorded effort, good outcomes do not show the arrangement works.
- Intrinsic and compensated performance: Performance a system achieves on its own, set against performance that depends on unrecorded human correction. A monitored buffer that absorbs genuine uncertainty is resilience. Unrecorded work that covers a fixable defect is a subsidy.
- Masking: When people cover for a broken system, its error rate looks low. The institution reads the low rate as health and fixes nothing.
- Evaluation by burdens removed: A system or a reform is judged by whether it removes preventable correction work from people, not only by how much it produces.
- Non-domination, burden accounting, and error-correcting authority: Three commitments a legitimate system has to keep. No one is subject to unchecked power. The time and stamina people spend keeping the arrangement working is counted. And authority lasts only while the errors it causes can still reach someone who will correct them. None of the three follows from the others.
- Corrective standing as epistemic admissibility: Corrective standing is not a veto that lets one person stall a system. It is a rule about evidence. An institution that shuts out the evidence from the people who bear its failures loses its claim to be right, and its warrant to keep running the system.
- Stewardship versus counterfeit buffering: A person kept in the loop is a steward when the role is voluntary and they can override the system without penalty. They are a counterfeit buffer when the work is compelled and unrecorded: they catch the system's errors and take the blame, and the errors never reach anyone who would redesign it.
The claims meet established fields, and the fields push back. Each claim's strongest counter, stated in its own terms, the answer, and the measurement that would settle it:
| Claim | Strongest counter | The answer | What would settle it |
|---|---|---|---|
| Apparent institutional performance often rests on unmeasured human compensation. IT productivity economics (Erik Brynjolfsson and colleagues) | Enterprise digitization yields genuine productivity gains once an adoption lag passes; workarounds are transient user adaptation, not permanent extraction. | Whether adaptation is transient or structural is an empirical question nobody can settle yet, because nothing measures the adaptation. A system that does not record the labor it consumes cannot tell a learning curve from a subsidy. The dispute is falsifiable; what is missing is the instrument. | the burden dashboards |
| Generated output shifts the cost of verifying truth onto its reader, and no metric that counts generation will ever see it. Ambient clinical documentation trials | Trials report high clinician satisfaction, reduced subjective administrative burden, and low observed error rates in routine primary care encounters. | Satisfaction is self-reported and would not detect a cost paid after hours; low error rates in routine encounters bound the claim without touching high-ambiguity work. The honest response is measurement rather than rebuttal — instrument the verification cost and let the number decide. | the claim-anchoring eval |
| Errors propagate through tightly coupled systems faster than correction can chase them, and nominal reversibility is not operational reversibility. Medication-safety research on interception | Multi-tiered human redundancy intercepts a large share of prescribing and transcription errors before they reach the patient, so informal checks function as a resilient defense. | Interception is real, and it is the finding the framework starts from: safety that depends on unmeasured vigilance is the human subsidy, and the interceptors are unbilled. The question is not whether interception works but whether it is recorded as evidence about the system or absorbed as evidence about the people. | the rescue register |
| Delay functions as an unlegislated rationing device that suppresses claims more effectively than an explicit denial. Health economics (moral hazard and demand management) | Prior authorization and review queues are legitimate cost control: they deter moral hazard and require providers to demonstrate medical necessity before funds are committed. | A queue may ration legitimately, but only as a declared policy with evidence, an owner, and a clock. Choosing not to fund is a decision with reasons and a date. What the framework refuses is rationing that arrives as friction nobody authored — and delay falls hardest on the least persistent, which makes undeclared rationing regressive. | the expiry default and deliberate non-use |
| Institutions enforce rules whose supersession they hold somewhere in the organization. Evidence-based medicine and regulatory science | Slow adoption is an epistemological safeguard: many trial endpoints rely on surrogate markers that never translate to survival benefit, so conservatism protects patients from commercial overhype. | The clause forces the review, not the adoption. Held evidence fires a trigger, and a payer may reject the new evidence with recorded reasons and confirm the grant. What it forbids is skepticism nobody performed. Slow uptake with a record is judgment; slow uptake without one is an outdated rulebook. | the unwired-evidence term |
| Empathy displays substitute for structural repair and absorb the pressure to change. Occupational health psychology (Christina Maslach, Michael Leiter) | Secondary interventions — peer support, mindfulness, institutional recognition — produce measurable reductions in exhaustion, depersonalization, and turnover intent. | The distinction is substitution, not existence. Peer support that complements redesign is not theater; a wellness program offered in place of changing the workload is. The test is whether the intervention alters the system that produces the burden or only helps people survive it. | the ethics-theater term |
Its departure is the instrument. The claim becomes obligations a running system can fail, records an auditor can read, and public cases scored against both:
- Obligations with clause numbers: a grant that expires unless renewed on evidence stated in advance (STD-08 §1.1); a widening of scope that needs its own authorization (STD-08 §1.3); an objection that owes a revision or a reasoned refusal on a clock (STD-07 §4.2); a published measure of the time a system costs people (STD-01 §7.1); and standing for whoever absorbs its errors (STD-02 §8.4).
- Records: the revisable delegation record and the authority grant schema (JSON), so a reader can check whether each obligation was met.
- Public cases: the casebook scores 5 failures on the six safeguards and on what the institution did afterward: 3 left the process unchanged, 1 partly changed it, and 1 changed the rules.
In short: the theory says why people's endurance does not make an arrangement legitimate. The method gives the records that show when an institution is running on that endurance while reporting success.
The qualitative phenomenology behind these failure modes is developed across The Crumple Zone; the formal research program on burden accounting and performance decomposition is hosted at kanav.net/research/.
Citing this work
How to cite Ethotechnics
Ethotechnics Institute materials are published under CC BY-SA 4.0 . To cite the framework:
Jain, K. (2026). Ethotechnics: An open framework for the operational accountability of high-stakes automated systems. Ethotechnics Institute. https://ethotechnics.org
Individual glossary entries and patterns include their own citation metadata.