Draft policy language
Minimum viable contestability sets eight commitments for challenging an automated decision. Each lists its requirements, one sentence apiece, and the evidence that shows they are met.
Start with minimum viable contestability →Draft standards for records a running system must keep: an answer to every objection, authority that expires, a published time cost. None binds until adopted.
On this page
Identifier, lane, status, version, how many clauses it carries, and where its evidence pack lives. The lanes are core rights and obligations, implementation templates, and reference formats.
| Standard | Title | Status | Version | Clauses | Evidence pack |
|---|---|---|---|---|---|
| STD-01 | The Temporal Bill of Rights Core Seven rights that protect a person's time from automated systems: to stop a process, to exit, to a bounded wait, to reversal, to wait without coercion, to reach a human, and to see the burden. | Draft | v1.0.3 | 19 | STD-01 pack |
| STD-02 | Contestability and recourse Core How a person challenges an automated decision: the reasons they are owed, an appeal path, a review clock, remedy, and who has standing to challenge. | Draft | v1.4.1 | 35 | STD-02 pack |
| MVC-01 | Minimum viable contestability standard Core A one-page baseline of eight commitments: standing, reasons, records, timelines, a named responder and standard of review, defined effects, remedies, and non-retaliation. | Draft | v1.0.1 | — | — |
| PM-01 | Institutional failure postmortem template Implementation A one-page postmortem template: how long the harm ran, whether it could be undone, who carried the burden, and how it was repaired. | Draft | v1.0.2 | — | — |
| STD-03 | Justice SLOs: targets, budgets, and breach actions Implementation Each justice metric, such as time to halt, gets a target, a measurement window, a breach action set in advance, and a named owner. | Draft | v0.6.2 | — | — |
| STD-04 | FHIR profile set for contestability artifacts Reference FHIR profiles for decision records, appeal events, and repair outcomes with governance metadata. Replaced by STD-05 . | Deprecated | v0.3 | — | — |
| STD-05 | W3C Verifiable Credential schemas for contestability Reference Verifiable Credential schemas and JSON-LD contexts for decision records, appeals, and remedies. | Draft | v0.3 | — | — |
| STD-06 | Human impact safety case Implementation A safety case for human impact: a standard set of tests with thresholds declared in advance, the evidence each produces, and the limits past which a system does not deploy. | Draft | v0.7.1 | 20 | STD-06 pack |
| STD-07 | Revisable delegation record Reference One append-only record shape for what an institution believed, could do, authorized, did, saw diverge, and revised. | Draft | v0.1.4 | 17 | — |
| STD-08 | Delegation Core What must stay true while an automated system acts for an institution: its authority expires unless renewed, the policies behind it are rechecked when facts change, human oversight names what the human can actually change, and the capacity to correct errors grows with the authority granted. | Draft | v0.3.2 | 24 | STD-08 pack |
| STD-09 | Agent chains Reference What must hold when a consequential decision passes through a chain of agents and services that no single person oversees: the whole chain is treated as one delegation, delays add up across it, and it can correct errors only as well as its weakest link. | Draft | v0.2.3 | 17 | STD-09 pack |
Conformance with these standards is evidence about how a delegation is run. It is not evidence that the power should be held, and it does not reduce anyone's legal rights or the operator's liability.
Anyone with a GitHub account can object to a clause, and each objection is answered within 30 days. The site's own record lists every objection and its answer.
Implementation schema
Under STD-08, every automated action runs under an authority grant: a record of who may act, the limits it cannot exceed, and when the permission lapses unless renewed.
Engineers do not need to rewrite their agent runtime or replace frameworks like Temporal, LangChain, or custom job runners. Ethotechnics sits at the invocation and audit boundary.
Before the worker in the example places a hold on an account, it reads
the grant and checks four fields: state is allowed, the until date has not passed, the action's class is in
action_classes, and the hold stays inside
scope.limits, such as at most 120 holds an hour. If
mode were confirm, a named person would also
have to approve each hold before it took effect.
No runtime rewrite
Published authority-grant schema · Download example
{
"schema_version": "1.3.0",
"grant_id": "GRANT-2026-0142",
"provenance": {
"issued_by": "risk-governance-board",
"issued_at": "2026-03-01T09:00:00Z",
"basis_summary": "Fraud-hold policy 3.2 adopted with a measured 1.4% false-positive rate on the Q4 backtest; reversal path rehearsed in February."
},
"scope": {
"action_classes": [
"READ",
"WRITE"
],
"systems": [
"core-banking-holds"
],
"populations": [
"retail accounts, excluding accounts flagged vulnerable"
],
"limits": {
"max_hold_days": 7,
"max_holds_per_hour": 120,
"max_hold_value_usd": 25000
}
},
"issuing_authority": {
"name": "Risk Governance Board",
"role": "Delegating authority",
"contact": "risk-governance@example.org"
},
"grantee": {
"id": "agent-risk-ops-v2",
"type": "agent"
},
"action_classes": [
"READ",
"WRITE"
],
"for_whom": "Retail banking customers, on behalf of the Fraud Operations desk",
"evidence_basis": [
{
"ref": "POL-FRAUD-HOLD-3.2",
"summary": "Hold rules and thresholds for suspected account takeover.",
"as_of": "2026-06-18T00:00:00Z"
},
{
"ref": "EVID-BACKTEST-2026Q1",
"summary": "Q1 backtest: 1.4% false-positive rate, median release 31 hours.",
"as_of": "2026-04-05T00:00:00Z"
}
],
"assumptions": [
"False-positive hold rate stays below 2% on the rolling 30-day window.",
"The Risk Steward can release any hold within two hours of a challenge.",
"Vulnerable-account flags are current within 24 hours."
],
"review_conditions": [
"False-positive hold rate exceeds 2% on the rolling 30-day window.",
"Any challenge decided against the system.",
"Policy POL-FRAUD-HOLD-3.2 enters review_required or expires."
],
"revocation_conditions": [
"Holds placed on accounts flagged vulnerable.",
"Intervention reach time exceeds four hours in any incident.",
"Withdrawal rehearsal fails."
],
"correction_capacity": {
"assessed_at": "2026-06-18T16:30:00Z",
"detection": {
"present": true,
"evidence": "Model Risk Review recomputes the false-positive rate and release latency each month from the raw hold log rather than from the Fraud Operations dashboard; the 2026-06 recomputation put the rate on accounts flagged vulnerable at 11% against a 2% assumption. The qualifier is on the record: CHL-2026-0088 was reported from outside by the account's representative, not caught by the monitor, so detection is present but was not first.",
"evaluator": "Model Risk Review (second line)",
"independent_of_executor": true
},
"challenge": {
"present": true,
"evidence": "CHL-2026-0088 was raised on 2026-06-10 by an affected account's representative through the channel REG-ACCOUNT-HOLDS-1 declares for the account_hold class, and was decided rather than left open."
},
"standing": {
"present": true,
"evidence": "REG-ACCOUNT-HOLDS-1 admits affected_party, corrective_labor, representative, and oversight for the account_hold class, and both decision and grant may be challenged. CHL-2026-0088 was raised on the affected_party basis, from outside the operating team."
},
"review": {
"present": true,
"evidence": "The Risk Steward on call answered inside the register's PT48H deadline (opened 2026-06-10T14:20Z, deadline 2026-06-12T14:20Z) and referred the grant to the Risk Governance Board, which decided REC-2026-0019 on 2026-06-18. Both the examiner and the clock are named in REG-ACCOUNT-HOLDS-1."
},
"authority_to_modify": {
"present": true,
"evidence": "REC-2026-0019 narrowed the grant itself: populations now exclude accounts flagged vulnerable and max_hold_days fell from 14 to 7. The examiner changed the grant rather than recommending a change."
},
"reversible_transitions": {
"present": true,
"evidence": "holds.release reversed every hold in the 2026-05-20 rehearsal and released the CHL-2026-0088 hold within two hours; measured median release is 31 hours (DEP-RISK-OPS-V2-2026-06)."
},
"operable_after_correction": {
"present": false,
"evidence": "Not shown beyond 48 hours. DEP-RISK-OPS-V2-2026-06 records institutional reversibility as not feasible: withdrawal past 48 hours breaches the card-network fraud-loss SLA and the rules-engine fallback is not staffed to absorb it, with correction latency measured at 31 hours against that 48-hour margin. Capacity is counted only for reversal inside 48 hours, which is why max_hold_days is 7, and under §4.2 the grant may not be expanded until the fallback is staffed."
}
},
"mode": "unattended",
"state": "allowed",
"state_history": [
{
"from": "none",
"to": "allowed",
"at": "2026-03-01T09:00:00Z",
"reason": "issued",
"evidence_ref": "POL-FRAUD-HOLD-3.2"
},
{
"from": "allowed",
"to": "review_required",
"at": "2026-06-11T10:00:00Z",
"reason": "challenge",
"evidence_ref": "CHL-2026-0088",
"note": "Challenge opened by an affected party over a hold on a recently flagged vulnerable account."
},
{
"from": "review_required",
"to": "allowed",
"at": "2026-06-18T16:30:00Z",
"reason": "challenge",
"evidence_ref": "REC-2026-0019",
"note": "Narrowed: populations now exclude accounts flagged vulnerable; max_hold_days reduced from 14 to 7."
}
],
"attestation": {
"signed_by": "key:risk-governance-board-2026h1",
"algorithm": "ed25519",
"signature": "MEUCIQD8sV3rN0kQ2Jm5Fh7tX1wLpA6yKb9cUe4SgTn2HvZ0dQIgLxO0Q7mWc3KpR8dN2fYj5bHs1aVvT6xCu9eEi4QnA7k=",
"signed_at": "2026-06-18T16:30:00Z",
"payload_digest": "sha256:9f2c41ba7e03d58c6a1f0b8e47d29c53a6b1e7d40f92c8a3b6d5e1c7f0a49b82"
},
"liability": {
"accepted_by": {
"name": "Meridian Retail Group",
"role": "Operating institution",
"contact": "legal@example.org"
},
"instrument": "Automated Decision Liability Instrument ADM-2026-02, accepted at grant issue and reaffirmed at REC-2026-0019",
"accepted_at": "2026-03-01T09:00:00Z",
"carve_outs": [
"Liability capped at USD 2,000,000 per incident under ADM-2026-02 §7; the cap is recorded so the residual risk above it is visible before an incident rather than argued after one."
]
},
"until": "2026-12-01T00:00:00Z",
"renewal_basis": "A fresh backtest on the preceding quarter, the current dependency record, and a withdrawal rehearsal within the last 90 days.",
"policy_refs": [
{
"policy_id": "POL-FRAUD-HOLD-3.2",
"version": "3.2.0"
}
],
"intervention_ref": "INT-RISK-STEWARD-1",
"capability_catalog_ref": "CAT-RISK-OPS-V2-2026-07",
"dependency_ref": "DEP-RISK-OPS-V2-2026-06"
}
Statutory crosswalks
For risk owners and legal counsel: these are proposed technical correspondences, not a determination of legal compliance.
The crosswalks live on one page. Each framework — the EU AI Act, NIST AI RMF, ISO/IEC 42001, ISO/IEC 27001, SOC 2, IEC 61508 — is mapped to the Ethotechnics controls, the clauses a contract can cite, and the evidence an auditor should ask for.
Open the crosswalk hub. The critiques of each framework sit in existing frameworks, below.
Where to start if you are writing policy, preparing for an audit, or overseeing a build.
Minimum viable contestability sets eight commitments for challenging an automated decision. Each lists its requirements, one sentence apiece, and the evidence that shows they are met.
Start with minimum viable contestability →The STD-01 minimum binding set lists the least each right requires. The evidence packs list the records that show it is met.
Review minimum binding set →Implementation examples show each standard as the halt rules, clocks, and records a delivery team builds.
Open implementation examples →Buying a system from a vendor? Cite clauses in the contract and require an evidence pack at each release
Each status has entry conditions, listed below. They say what a standard needs before it can move to that status.
Under active review. Clauses may still change.
Settled enough to cite in production policy and procurement requirements.
Not for new adoptions. A replacement is linked when one exists.
deprecated_by.
The axioms and laws the standards are written against, and the reference documents that apply them.
Foundations
The five axioms every standard rests on (finitude, consent, stewardship, reversibility, and legibility) and the premises they assume.
View axioms →Foundations
The twelve laws the standards enforce, the one-sentence rule that sums them up, and the six things a system must keep in step as it grows.
Read the laws →Practice
How to implement STD-01 and STD-02 in agent systems, mechanism by mechanism, with a route to challenge that a person can use.
View practice →STD-01 reference
One invented case traced from each harm to the STD-01 right it breaks, the check that catches it, and the change that fixes it.
Open mapping →STD-01 reference
Each STD-01 right, the validators that test it, and the mechanisms that implement it.
View matrix →STD-01 reference
The least each STD-01 right requires before a system can claim it, with clause references and examples that fall short.
Review binding set →Governance
How to write these standards into contracts, procurement terms, and audits.
See guidance →Governance
Each Ethotechnics control mapped to the EU AI Act, NIST AI RMF, and ISO/IEC 42001, with the evidence a buyer or auditor should request.
Open crosswalks →Reference
Definitions of the terms the standards use, each at a stable URL.
Browse glossary →Existing frameworks
For each of the most cited AI governance frameworks, what it leaves out once a system is running.
OECD
Non-binding principles: they ask for accountability and set no duty anyone can enforce.
Read critique →NIST
Voluntary risk management that asks for feedback and impact mapping, and lets the operator decide what counts as done.
Read critique →ISO
Certifies the management system. Nothing makes one system's authority to act expire.
Read critique →EU
Most high-risk systems are self-assessed before release, and their authority does not lapse. A challenge need not reach the system.
Read critique →Corporate
Internal boards and review gates that answer to the company, not to the people a system decides about.
Read critique →Meta-critique
The general critique: frameworks that govern through documents, roles, and reviews, not through records a running system has to keep.
Read critique →Standards comparison
A side-by-side comparison for regulators, operators, and risk owners.
Frameworks and law
This comparison covers four AI governance frameworks and one data protection law that already governs automated decisions. Each addresses a different part of the problem:
These frameworks improve documentation, oversight, and monitoring. They leave one question largely open. When the evidence shows a running system is harming people, does anyone have to act on it within a set time, without the harmed person carrying the burden? Acting can mean halting the system, reversing its decisions, restoring the people harmed, repairing the harm that cannot be undone, or correcting the rule that produced it.
In short
The frameworks above answer one question: “Did you manage risk responsibly?” Ethotechnics answers a different one: “When this system is harming someone, does evidence of the harm reach someone who has to change the system?” The difference is in mechanics, not values.
In the casebook, harm ran from four days to about twenty years before anyone halted or changed the system. Decisions occur at machine speed. Reversal occurs at institutional speed. Harm compounds in the gap.
The measures an operator would take are time in harm, time to halt, recovery time, and the time a system costs people.
Comparison matrix
| Dimension | Prevailing standards | Ethotechnics |
|---|---|---|
| Governance object | Risk & compliance | Runtime governability |
| Typical artifacts | Registers, assessments, policies | Expiring authority grants, objection records, burden reports |
| Authority | One approval, with reviews on a schedule or after a major change | A lease that expires unless renewed, and a new decision for each widening |
| Oversight | Human oversight with a means to interrupt | A named owner who must answer, and a halt that works |
| Measurement | Process maturity | Time in harm + the time the operator takes from people |
| Challenge | A right to contest one decision | A revision or a reasoned refusal for each objection, on a clock |
| Monitoring | Visibility | Visibility that triggers a required intervention |
| Where the check happens | Audits, conformity assessment, regulators | Built-in constraints + internal power to halt, reverse, or revise |
| Ethics framing | Principles weighed case by case | Constraints that hold in every case |
1. Controls in the running system
A policy states what should happen. A control in the running system limits what does. Beyond inventories, risk registers, and reviews after the fact, the standards require halt conditions, rehearsed rollback paths, and safe fallback modes that work while the system runs.
2. Authority that has to act
Oversight is symbolic unless someone has to act on what it sees. Rather than advisory ethics committees or perfunctory review gates, the standards ask the operator to name who can halt, reverse, or revise a running system, and to put that person on a clock. An objection owes a revision or a reasoned refusal. A grant lapses unless someone renews it on stated evidence.
3. Temporal performance bounds
The person waiting on a decision carries the cost of every delay. When decisions occur at machine speed, reversal cannot run at institutional speed. A system may not decide faster than its appeal path can answer, and every pending decision needs a deadline.
Existing standards and Ethotechnics agree that governance matters. The question Ethotechnics adds is whether evidence that a system is harming someone compels action. The action can be a halt, a reversal, a repair, or a changed rule. It has to come within a set time, every time, and the harmed person must not carry the burden.
That is the difference between governance as documentation and governance as capability.
Other framings
Short framings of the comparison above, each for a different reader.
Regulator-facing brief
EU AI Act Article 14 and NIST AI RMF MANAGE 2.4 ask for a way to stop a system. Neither requires the halt to be rehearsed or to meet a time limit. These drafts would, once a contract or a regulator adopts them, and they name the records an inspector needs to check the halt without taking the operator's word for it. No deployment conforms to them today.
Records to request:
Board brief
In the casebook, harm ran from four days to about twenty years before anyone halted or changed the system. Three questions to ask before approving a deployment:
Each answer is a record the standards require: a named owner who can halt, a halt drill receipt, and an expiry date on the grant.
Diagrammatic version
The policy stack and the control plane, side by side, for workshops and audits.
Policy stack → policies → reviews → audits → reports Control plane → halt rules → clocks → rollback → restoration
The first line is checked from outside the running system. The second is built into it.
Implementation examples
For each domain, conventional governance beside the same system built with halt rules, clocks, and restoration paths.
How the same system is built differently in each domain once it has to be stoppable and reversible.
Read guide →Benefit and debt rules with named owners, an outside halt, and restoration before restart.
Read example →Credit decisions with reasons, named owners, and challenges that reach the scoring rule.
Read example →Account protection systems that measure recovery time alongside detection.
Read example →Diagnostic tools whose overrides reach the model's owner, with patient challenges answered by a date.
Read example →FHIR profiles that make refusals, appeals, and repair clocks exchangeable data.
Read example →Support automation with a route to a person and a named owner for each hand-off.
Read example →Recommendations a person can turn off, with challenges where targeting sets a price or an offer.
Read example →An assistant with a grant that ends, a stop that works, and records the person checks it against.
Read example →Retail personalization is a boundary case: no single recommendation changes a person's status, access, money, or risk.