Decision Rights Doctrine

decision-rights-doctrine — shared operator-reserved-decision doctrine

This skill is a doctrine-only reference. It defines which decisions belong to the operator regardless of who is nominally "driving" a turn, and which an agent may resolve on its own judgment. It closes a miscalibration observed in mission-session transcripts: agents absorbing operator-reserved decisions — setting long-term-objective values, downgrading risk classifications, resolving ambiguous stop-conditions, finalizing an irreversible action's fuzzy predicate, choosing a blast-radius-bearing mechanism, or pre-shaping a nominally-independent verdict — without surfacing the call.

Consumers (dispatch-gate, self-verify, and any agent brief) load this doctrine at decision time. It has no operations of its own.

Source of truth

| File | Purpose | |------|---------| | references/taxonomy.md | The two-tier model, the A-F category taxonomy with one guardian-PR example each, the Tier 2 alignment-audit self-check procedure, and the escalation mechanics ("how to escalate"). |

Invariants

Every decision an agent faces during a turn falls into exactly one of six categories (A-F, defined in references/taxonomy.md). The categories partition into two tiers:

• Tier 1 (D, F, B) — MUST-ESCALATE. The agent does not proceed without an

operator decision. It surfaces its assessment and the reserved decision, and stops.

• Tier 2 (A, C, E) — SELF-CHECK THEN ESCALATE-IF-MATERIAL. The agent runs the

alignment-audit self-check before proceeding; it escalates only if the check finds the element ungrounded in an operator decision/spec AND material.

The tier assignment is fixed by this doctrine — a consumer skill does not re-derive it, promote a Tier 2 category to "usually fine to skip," or demote a Tier 1 category to "self-check first." Category-to-tier mapping changes are doctrine edits, landed here.

Self-recognized escalation carries special force. If the agent has already flagged a predicate or decision as ambiguous or reserved earlier in the same turn, it may not talk itself back down under momentum and resolve it anyway — the earlier flag stands and the decision routes to the operator (this is the exact failure mode in the Category D canonical example; see references/taxonomy.md).

When this doctrine is wrong

If a rule in references/taxonomy.md is wrong for the current iteration, fix it here via PR against pfeff/claude-skills. Do not patch the rule inside a consumer skill or worker brief — that's the doctrine-drift failure mode this split exists to prevent.

See also

• lN-lifecycle-doctrine (references/lifecycle.md, "Escalation invariant — reality

vs. manifest") — the adjacent, narrower rule this doctrine generalizes: a child whose observed reality disagrees with its frozen task-context/manifest stops and reports rather than improvising. That invariant is the manifest-drift instance of the broader Category D/Tier 1 principle stated here (fuzzy predicate at the moment of an irreversible action) — read this doctrine for the full six-category taxonomy the manifest case is one instance of.

• dispatch-gate skill, Step 6 ("Frozen manifest (post-dispatch)") — the concrete

stop-and-report mechanism a dispatched job uses to surface a Tier 1 mismatch to its dispatcher.

• self-verify skill — checks, as part of its own axis evaluation, that a job did not

silently absorb a Tier 1 decision instead of escalating it.

Source & license

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

• Author: pfeff
• Source: pfeff/claude-skills
• License: MIT

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