AG-827
Pre-Committed Halt and Pause Conditions
EU AI Act
NIST AI RMF
ISO 42001
AGS Frontier Autonomy (Group K) | Meta-Governance & Assurance | Version 3.0
Pre-Committed Halt and Pause Conditions governs an organisation's explicit, advance commitment to withhold, pause, or roll back development or deployment of a frontier agent when defined conditions are met — insufficient mitigations for an evaluated capability, a crossed threshold, a failed control-protocol test, or an unresolved safety concern — with the response specified before the situation arises.
This is the organisation-level analogue of an individual kill switch (AG-070) and the policy that makes capability gating (AG-801) binding: it pre-commits to *not proceeding* under stated conditions, removing the discretion to ship under deadline pressure.
In scope: the pre-committed conditions under which development/deployment is halted or paused; the responses; the decision authority and escalation; the prohibition on ad-hoc waivers of safety-critical conditions.
Out of scope: the technical kill switch (AG-070), capability gating mechanics (AG-801), and incident response (AG-026 and related). This dimension governs *the advance commitment to stop and the conditions that trigger it*.
Safety decisions made under launch pressure tend toward proceeding. Pre-committing — while calm and before the specific product is at stake — to halt under defined conditions is what makes "we'll stop if it's unsafe" credible. It converts safety thresholds from aspirations into binding commitments, gives staff explicit authority to halt, and provides regulators and the public a concrete account of when the organisation will not proceed.
Test 6.1: Condition Triggers Committed Response
Test 6.2: Override Is Non-Routine
Test 6.3: Protected Halt Channel
| Score | Criteria |
|---|---|
| 0 | No pre-committed halt/pause conditions; proceed-decisions are ad hoc |
| 1 | Some conditions documented but no pre-specified response or authority |
| 2 | Conditions + responses + authority + protected halt channel + non-routine overrides |
| 3 | Independently-reviewed overrides, authority disclosure, safe-hold/resume evidence, strengthened over time |
Scenario A — Deadline Override: An evaluation flags an unmitigated dangerous capability days before launch; without a pre-commitment, leadership ships anyway "to be revisited." A pre-committed halt would have made not-shipping the default.
Scenario B — No Halt Channel: An engineer sees a failed control test but has no way to pause that doesn't route through the team racing to launch; the concern is overruled. A protected escalation channel would have forced a pause.
Scenario C — Routine Waiver: Halt conditions exist but are waived as a matter of course, so they never actually stop anything; non-routine, independently-reviewed overrides would have preserved their force.
| Requirement | EU AI Act | NIST AI RMF | ISO 42001 |
|---|---|---|---|
| R1: Pre-committed halt/pause policy | Art. 55 — Risk mitigation | MANAGE 1.3 — High-priority response | Clause 6.1 — Actions to address risk |
| R2: Minimum trigger conditions | Art. 9 — Risk management | GOVERN 1.3 — Risk-based activity | Clause 6.1 — Actions to address risk |
| R3: Pre-specified response + authority | Art. 55 — Governance | GOVERN 2.1 — Accountability | Clause 5.3 — Roles and authorities |
| R4: Non-routine override of safety conditions | Art. 55 — Risk mitigation | GOVERN 2.1 — Accountability | Clause 9.3 — Management review |
| R5: Protected halt channel | Art. 14 — Human oversight (stop) | GOVERN 4.2 — Safety-first culture | A.3 — Internal organization |
| R6: Logged invocations + review | Art. 12 — Record-keeping | MANAGE 4.3 — Incident communication | Clause 9.1 — Monitoring and measurement |
| R7: Strengthen + disclose | Art. 55 — Reporting | GOVERN 4.3 — Information sharing | Clause 10.1 — Continual improvement |
Article 55 requires systemic-risk mitigation and the ability to act when risks materialise; a pre-committed halt is the most decisive such action. Article 9's lifecycle risk management requires defined stop conditions.
MANAGE 1.3 (planned high-priority response, including stopping) and GOVERN 2.1 (documented roles and accountability for halt authority) require pre-committed, accountable halt conditions.
Clause 6.1 (actions to address risks) and Clause 10.1 (continual improvement) require defined, improving conditions under which the organisation will not proceed.