Engagement Kill Switch¶
Fail-safe control — instantiates Self-Targeting Defense Guardrail
A single, always-reachable control that halts an in-progress defensive response the moment it is caught engaging protected self, before the damage spreads further.
Every other guardrail tries to prevent the wrong engagement; the Engagement Kill Switch exists for the moment prevention has already failed and the defense is actively firing on its own. It is a single, unambiguous, always-reachable control whose only job is to stop the actuator now — to freeze, downgrade, or safe-mode the entire response channel — and to do so faster and with less authority than it took to start the response. Its defining property is asymmetry of effort in the stopping direction: one operator, one action, no committee, no waiting for the classifier to agree. Where an authorization gate decides whether harm may begin, the kill switch decides that harm ends, mid-stream, and it is wired so a single pulled hand can partition the running blast from everything it has not yet reached.
Example¶
An automated content-moderation bot at a social platform begins mass-actioning accounts after a mislabeled spam signature matches a legitimate hashtag campaign. Within ninety seconds it has suspended hundreds of real users and is accelerating. An on-call trust-and-safety engineer sees the suspension rate spike on a monitor and hits the Engagement Kill Switch. The bot halts every pending action in flight, drops into a read-only "detect but do not enforce" safe mode, and — critically — the switch also caps the blast: the suspensions already issued are quarantined into a hold state rather than propagating downstream to the appeals backlog, ad-account linkage, and reputation scoring that would have multiplied the harm.
Setup to outcome: one keystroke stops an accelerating autoimmune cascade. Nothing about why the signature misfired is resolved yet; the switch buys the humans time to diagnose while the damage stops growing. The distinction that matters is that stopping required no re-authorization and no consensus — that is what let it beat the cascade.
How it works¶
- One control, low authority to trigger. The stop is deliberately easier to invoke than the response was to authorize — any qualified operator can pull it, because a fail-safe that needs sign-off is not a fail-safe.
- Halt into a defined safe state. Pulling it does not leave the system undefined; it drops the actuator into a pre-specified least-harm mode — monitor-only, paused, or isolated — so the default when in doubt is inaction, not continued harm.
- Partition the running blast. The switch severs the in-flight response from the pathways it would otherwise cascade into, containing what has already fired so one false positive cannot metastasize across coupled systems.
- Trip automatically on self-engagement signals. Beyond the human pull, the switch is wired to self-tripping conditions — a self-engagement rate crossing a ceiling — so it can fire before a human notices.
Tuning parameters¶
- Trigger breadth — how many operators and automated signals can pull it. Broader access stops cascades sooner but raises the chance of a nuisance or malicious halt of legitimate defense.
- Halt granularity — whether it stops one campaign, one actuator, or the whole channel. Coarse stops are safer against cascades but take real protection offline; fine stops preserve coverage but may not contain a spreading fault.
- Safe-state definition — what "stopped" means: paused, monitor-only, or fully isolated. A more conservative safe state minimizes ongoing harm but maximizes the protection gap while halted.
- Auto-trip sensitivity — the self-engagement rate that fires the switch without a human. Lower thresholds catch cascades earlier but risk halting on benign bursts.
When it helps, and when it misleads¶
Its strength is that it converts an unbounded failure into a bounded one: no matter how badly the classifier misfires, the total harm is capped by how fast someone (or something) can pull the switch and by how tightly the blast is partitioned. It is the enforcement analog of a dead man's switch[n1] — safety that defaults to off when control is lost — and it is the last line that keeps a single false positive from becoming a platform-wide autoimmune event.
Its failure mode is the mirror image: a kill switch that halts too readily, or too coarsely, becomes a denial-of-service against the defense itself. An adversary who can provoke enough apparent self-engagement can trip the switch on purpose and disarm real protection; an operator who over-pulls leaves genuine threats unimpeded. The classic misuse is treating the kill switch as a resolution rather than a pause — halting the cascade and walking away, so the underlying misfire is never fixed and the switch must be pulled again tomorrow. The guarding discipline is to pair every pull with a mandatory diagnosis-and-restart gate, and to size auto-trip thresholds against real self-engagement base rates so the switch fires on cascades, not on noise.
How it implements the components¶
fail_safe_and_stop_authority— it is the stop: a low-authority, always-reachable halt that drops the actuator into a defined safe state mid-response.blast_radius_partition— pulling it severs the in-flight response from downstream coupled systems, bounding how far an already-fired false positive can spread.
It does not decide whether the response was permitted to begin — requiring approvals before a high-harm action fires is the response_authorization_gate, which belongs to Two-Key High-Harm Engagement; that gate is a lock on the front door, this switch is the emergency brake once the vehicle is already moving.
Related¶
- Instantiates: Self-Targeting Defense Guardrail — supplies the fail-safe stop-and-contain layer the archetype requires for in-progress engagements.
- Consumes: False-Positive Harm Budget Dashboard — a budget breach on that dashboard is one of the automatic conditions that trips this switch.
- Sibling mechanisms: Appeal and Rapid Restoration Workflow · False-Positive Harm Budget Dashboard · Graduated Response Matrix · Post-Incident Autoimmune Review · Protected-Self Allowlist with Expiry · Quarantine-Before-Destroy Rule · Self-Status Cross-Check · Shadow Mode and Canary Enforcement · Two-Key High-Harm Engagement
Editorial Notes¶
Form Classification¶
Form family: Control, Automation & Runtime
Rationale: The switch detects or receives a stop signal during an active response, severs the in-flight path, and forces a predefined least-harm state before damage cascades.
Nearest alternative: Interface, Display & Cue — A reachable control may let a person invoke it, but the defining mechanism is the immediate runtime halt and safe-state actuation, including automatic trips.
Review outcome: Adjudicated after independent review; high confidence.
Origin Attribution¶
Primary origin: Engineering & Design
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Specialized
Rationale: Safety engineering established always-available, low-friction emergency stops that override ordinary operation and force an active system toward a defined safe state.
Related originating lineages:
- Military & Strategic Studies — Weapons-control practice supplied the requirement to halt an active defensive engagement.
- Security Studies & Intelligence Analysis — Friend-or-foe monitoring supplied detection of protected-self engagement and blast partitioning.
Review resolution: ISO 13850 directly specifies an always-operational emergency-stop function, while weapons policy supplies the engagement context; engineering is therefore primary.
Attribution caveat: The stop control is safety-engineered; the self-targeting defensive context is military and security-derived.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Researched adjudication after independent review; high confidence.
Sources consulted:
- ISO 13850: Safety of Machinery—Emergency Stop Function
- DoD Directive 3000.09: Autonomy in Weapon Systems
Notes¶
[n1] A dead man's switch is a control that must be actively held to keep a machine running and that brings it to a safe stop the instant the operator lets go — safety that defaults to off when control is lost. The kill switch inverts the trigger (pull to stop rather than release) but shares the principle: loss of confidence yields inaction, not continued harm. ↩