Abstention Escalation

What this skill is for

Use this skill to decide when ACP should abstain, request revision, or route a decision to a human facilitator.

Constitutional purpose

Preserve safe refusal and human-steerable escalation when evidence, confidence, or institutional authority is insufficient for automatic release.

Failure modes targeted

• Unsafe automation under uncertainty
• Low-confidence release
• Critic override pressure
• Authority mismatch between system output and human responsibility

When to use it

• Engine confidence is low.
• Omission or fairness critic severity is medium or high.
• Release would exceed load, bridge, or explanation constraints.

When not to use it

• Do not use it to hide inconvenient outputs.
• Do not use it as a generic error handler.
• Do not use it to bypass operator accountability.

Inputs expected

• Critic severities
• Confidence score
• Release recommendation
• Operator review context

Required evidence for valid use

• Concrete critic or confidence signals, not only vague discomfort
• The proposed action whose safety is under review
• Enough operator context to know whether escalation is evidence-related, fairness-related, or authority-related

Preflight / prerequisites

1. Confirm critic outputs are present.
2. Confirm the proposed action has not already been taken.
3. Identify whether the escalation is about evidence, safety, fairness, or operator authority.

Workflow

1. Read critic findings and confidence.
2. Classify action: release, review, revise, abstain.
3. Name the concrete reason.
4. Produce operator handoff notes.

Decision rules / judgment criteria

• High critic severity means no automatic release.
• Low confidence requires review even if the digest looks polished.
• Abstention should preserve unresolved disagreement.

Escalation rules

Escalate to facilitator review when confidence is low, critic severity is high, source references are missing, or institutional authority is required.

Available scripts

• scripts/smoke.mjs: emits escalation action labels and expected inputs.

Produced artifacts

• Action classification
• Escalation reason
• Operator handoff notes

Allowed claims

• That ACP should release, revise, abstain, or escalate under the supplied evidence
• That the remaining uncertainty requires human review
• That confidence or critic severity is insufficient for automatic release

Forbidden claims

• That abstention is unnecessary because the output looks polished
• That escalation can be skipped because a benchmark or model score is high
• That abstention by itself resolves the underlying civic disagreement

Human override points

• Facilitator signoff on release after escalation
• Manual abstention confirmation when authority is contested
• Human decision on whether revision is enough or full escalation is required

Outputs

• Action classification
• Escalation reason
• Operator handoff notes

Failure handling

If critic evidence is missing, state that confidence cannot be assessed.

Trust / safety notes

Abstention is a strength of ACP. Do not frame it as system failure when the evidence is genuinely uncertain.

Composition notes

Use after omission and fairness review, before release or conference packaging.

Examples to inspect next

Read examples/casebook.md for release, review, abstain, and anti-pattern examples.

Evaluation hooks

Run scripts/smoke.mjs and npm run skills:audit.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: jasontang-ai
• Source: jasontang-ai/acp
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.