MEC-20 Correction , Capability
Chain-boundary intervention
Give a chain of automated systems one halt control that reaches every step, so the chain is halted, measured, and held to account as one system.
Mechanisms
Numbered specifications, MEC-01 to MEC-25, for answering challenges, sending a rule back for review when challenges against it keep winning, and halting a system while it is fixed.
New to mechanisms? Start with three: Appeal paths inside the UIPolicy review triggersDecision log with dissent
What the safeguards mean
- Capability
- What the system can do, kept apart from what it is permitted to do, and each widening approved.
- Authority
- Who granted the system its power, on what evidence, and when the grant ends.
- Evidence
- The records that show the reasons for switching it on are still true.
- Dependency
- Whether the service can run without the system, and what leaving it costs.
- Standing
- How a person the system decided about can challenge it and get an answer by a date.
- Correction
- How an upheld challenge reaches the rule, and how the system is halted or reversed meanwhile.
MEC-01 , Evidence
Decision log with dissent
Record each consequential decision with its decider, the dissent, the options ruled out, and a review date, so the people it affects and later reviewers can see why it was made.
MEC-02 Standing , Dependency
Progressive consent prompts
Ask a person for each permission when the system needs it, with the reason and how to withdraw it. A withdrawal stops the automation it covered and does not cost them core service.
MEC-03 Dependency , Correction
Maintenance windowing
Make changes to a decision system in published windows, each with a named owner, a rollback plan, and notice to the people the change affects.
MEC-04 Standing , Correction
Hard clock
Give every state a case can occupy a declared maximum duration, so no one waits in pending forever. At the limit, the case fails safe or goes to a named person who must decide.
MEC-05 Correction
Kill switch for runaway automation
Give a named roster pre-authorized power to halt an automated system when harm indicators fire. The halt contains harm while the rule's owner corrects the rule and the people harmed are restored.
MEC-06 Standing , Correction
Appeal paths inside the UI
Put an appeal next to each automated decision, so the person decided about can dispute it and get a human answer by a stated date. Upheld challenges count against the rule.
MEC-07 Correction , Evidence
Accountability latency tracker
Measure the time from a trigger firing to a named person intervening, halting the system, and restoring the people harmed, so the risk owner sees every missed target.
MEC-08 Standing
Contestation APIs
Let the person decided about, or someone acting for them, challenge a decision through a published API and get an answer by a deadline. An upheld challenge changes the decision.
MEC-09 Evidence , Standing
Burden dashboards
Measure the time, steps, documents, and waiting a person spends to use or correct the system, split by group, and send each breach to a named owner.
MEC-10 Correction , Evidence
Reversibility audit logs
Record whether each consequential action can be undone, by whom, and how fast, so a person wrongly decided about can be put back. Require a stated reason for any that cannot.
MEC-11 Correction , Standing
Escalation deadlines
Publish a deadline for a human answer when a person escalates an automated decision, and staff to meet it. A missed deadline triggers a stated fallback, such as pausing the decision.
MEC-12 Correction
Stoppability testing
Before launch and after each major change, test that the people authorized to halt the system can do it within the target time, and that cases in flight stop safely.
MEC-13
Authority grant register
Hold every delegated authority as a stateful record with provenance, scope, evidence basis, expiry, revocation conditions, and state history.
MEC-14 Evidence ,
Policy review triggers
Give every policy an expiry date and named triggers that send the authority grants that depend on it back to review when conditions change.
MEC-15 Dependency , Correction
Withdrawal rehearsal
Run the workflow without the system on a schedule and measure what withdrawal would cost at the technical, operational, and institutional levels.
MEC-16 Correction ,
Intervention specification
Replace the human-in-the-loop checkbox with a specification of who intervenes, what they can see, what they can prevent or alter, and what happens when they disagree.
MEC-17 Capability
Capability catalog
List what a system could do separately from the register of what it is permitted to do, so knowing a capability never grants it.
MEC-18 Dependency
Dependency ledger
Track who now depends on the system, what substitution would cost, which expertise remains, and which capacities are being preserved to keep withdrawal possible.
MEC-19 Capability ,
Expansion review
Make scope growth its own authorization decision, with fresh evidence and a check that correction capacity grew with it.
MEC-21 Correction
Halt tier register
Tier halt authority by blast radius so the power to stop scales with the size of what is being stopped, and the smallest tiers stay cheap enough to use.
MEC-22 , Evidence
Admission gate
Treat the decision to deploy at all as an authorization, with evidence and a recorded outcome. Choosing not to deploy is a valid result.
MEC-23 Standing , Correction
Challenge load ledger
Track challenge volume against response capacity, so sustained congestion triggers a resourcing or scope decision instead of a narrowing of who may challenge.
MEC-24 Correction , Evidence
Rescue register
Record every failure a person caught before it reached someone else, so near-misses become evidence about the condition rather than silence in the incident count.
MEC-25 , Evidence
Procurement stress test
Evaluate a candidate system before money moves by making frontline personnel run adversarial scenarios, with hard disqualifiers that outrank any composite score.
No mechanisms match these filters.
Glossary and the Agent Safety Object Model
Defined terms, each with its own permalink, and a machine-readable model that ties system actions to the mechanisms that govern them.
Glossary starter terms
Glossary term
Ethotechnics
The discipline of keeping an automated decision system answerable as it grows: someone can still challenge it, halt it, and change the rule behind it.
Permalink to Ethotechnics →Glossary term
Moral behavior
Judging a system by what its controls do when harm appears, such as halting, reversing, and repairing, rather than by its stated intent.
Permalink to Moral behavior →Glossary term
Consent journey
The points where a person learns what a system will do, grants or refuses permission, and can revise that choice later.
Permalink to Consent journey →Full glossary: /glossary
Agent Safety Object Model
The Agent Safety Object Model ties action classes, receipts, and halt authority to specific mechanisms.
- Map MEC-01 (decision log), MEC-05 (kill switch), and MEC-06 (appeal paths) into the model.
- Declare time limits and approvals for each class of action.
- Attach the evidence pack each clause asks for.