FWD-1 Stopping an AI decision

Fijishi Working Draft 1 – 14 October 2026

Stage 1 of 4: working draft. Not yet open for public comment, not a numbered version, and not the criteria Fijishi certifies against. This draft replaces, as a new version, the earlier paper "Sovereign Algorithmic Immunity Doctrine (SAID)" (2023-2026); the earlier text is available on request. Standards in development

The question

When an AI system is making decisions, can the firm stop them – deliberately, quickly and on record – and show afterwards who could do so and what happened?

Why it matters to firms

AI systems now decide or shape claims, prices, underwriting and fraud flags. Rules in several jurisdictions give people a right to have some automated decisions reviewed by a person, and supervisors increasingly ask who can override or suspend an AI system. A defect in a model is a technical problem. Being unable to show who could stop it, and when it was stopped, turns it into a governance failure. Jurisdiction by jurisdiction: the Index.

Principles

  1. Named authority – a named role, not a team, may stop the system.
  2. Proportionate stop – the firm can stop one kind of decision without switching off everything else.
  3. Tested, not assumed – the stop has been shown to work.
  4. Recorded – every stop, override and restart leaves a record.
  5. The vendor does not hold the only switch – where a vendor runs the system, the firm keeps the right and the practical means to stop it.
  6. Controlled return – restarting is a decision by a named authority, and decisions made while a defect went undetected can be found again.

What a firm should be able to show (draft criteria)

  • FWD-1.1 An inventory of the AI systems that make or materially shape customer-affecting decisions, and the decision types each affects.
  • FWD-1.2 For each system, a named role and a deputy with documented authority to suspend it.
  • FWD-1.3 The ability to suspend one decision type or segment without halting unrelated processes.
  • FWD-1.4 A target time for a suspension to take effect, set by the firm, and evidence that it was met in the most recent test.
  • FWD-1.5 A test of suspension at an interval the firm sets, at least once a year, with the result recorded.
  • FWD-1.6 A record of every suspension, override and restart – who, when, why and which decisions were affected – that cannot be altered without trace.
  • FWD-1.7 Where a vendor operates the system or model, a contract that gives the firm the right and the practical means to suspend it without the vendor's consent, and notice of vendor changes that affect it.
  • FWD-1.8 A way to identify, and re-examine, decisions made while a defect went undetected.

How it relates to existing rules and standards

  • It is meant to operate inside an AI management system such as one certified to ISO/IEC 42001; it does not replace one.
  • It sets out how a human-oversight, override or human-review requirement can be shown in evidence. Where such a requirement binds a firm, the requirement governs; these draft criteria do not add to it.
  • One published draft names this directly: paragraph 60(ii) of the Reserve Bank of India's draft guidance on model risk management calls for override, suspension or deactivation mechanisms, "including kill-switch arrangements" – consultation draft; comments closed 24 July 2026; no date set. Status: Regulatory Deadline Calendar.

Where it is used now

It informs the Kill-Switch Compliance Certificate, which measures one system against one named requirement.

What happens next

FWD-1 is the first draft we will open for public comment. Comments are welcome now: protocol@fijishi.com