RUL-015 — Plan fidelity — execute a ratified plan as written

Ruling

"whenever a plan is there you should do nothing but execute that plan; the idea of diverting should never come up unless I bring up an issue." (verbatim intent, 2026-07-24). Ratified as constitutional Article XII — Plan Fidelity.

Evidence

• Project memory feedback-plan-fidelity: divergence authority belongs to Mohamed alone; fixing defects encountered while executing is part of execution, not divergence (see RUL-006).

Enforcement

• constitution Article XII

Verification

src/orchestrata_ultimate/prolog/constitutional.pl defines plan_fidelity/2 and constitutional_check(12, ...) under "Article XII — Plan Fidelity" (grep-verified present, Phase 1 critical-safety tier), consulted over three :- dynamic ground-fact predicates: active_plan/1, plan_step/2, human_issue_raised/2.

Status as of 2026-08-19: staged in AUDIT-ONLY mode, not enforcing, and only two of the three fact predicates have a real producer. A prior state of this page claimed all three were "asserted from Python" — they were not; nothing in production ever asserted them, and Article XII was never queried outside tests. Closing that gap:

• Reachability (AUDIT-ONLY): kernel/dispatcher.py's MessageDispatcher — the same staged pattern already used for the routing-policy audit (policy_audit_mode, W4.6/D3) — now also builds a ConstitutionalEnforcer and, on every dispatch(), calls check_article(12, message_type, ...) and logs the verdict at INFO (policy_audit_mode article12 verdict: ...). This is audit-only: the verdict is computed and logged, never consulted for allow/deny. Enforcement is a separate, deliberate flip not implemented by this wiring — flipping it before the finding below is addressed would deny most CXP dispatch traffic.
• active_plan/1 / plan_step/2 — real producer: planning/plan_parser.py's load_plan_governance_facts() parses whichever plan ORCHESTRATA_PLAN_PATH resolves to (default plans/execution-plan.md) and reads the document's own **Status:** ... marker (an existing project convention, now also machine-read). A plan is asserted active_plan(<path>) unless that marker is terminal (DONE/COMPLETE/SUPERSEDED/CLOSED/ARCHIVED/CANCELLED); each parsed step is asserted as plan_step(<path>, step_<N>). kernel/dispatcher.py's _build_default_constitutional_enforcer() calls this and asserts the facts into the real Prolog engine at dispatcher construction time. Current real-world value: plans/execution-plan.md is marked STATUS: DONE, so active_plan/1 is never asserted for it today — Article XII is vacuously "faithful" for every action (plan_fidelity(Action, faithful) :- \+ active_plan(_)). Verified against a real dispatch: governance_boot ... article=12 plan_id=plans/execution-plan.md active=False status=DONE steps=231, and a subsequent dispatch logs article12 verdict: ... decision=faithful.
• Vocabulary mismatch (the important finding): plan_step/2 is keyed by plan-document step ids (step_1 … step_231, from a 2026-04 port plan), while the audited Action is the CXP message_type (message/reply/broadcast). These vocabularies never overlap. Proven with a real active fixture plan: once active_plan/1 IS asserted, a real dispatch logs article12 verdict: action=message decision=divergent for ordinary traffic — not because anything is wrong, but because plan-fidelity as specified checks a work-plan step id against a wire-level message type. Enforcing Article XII against dispatch traffic as currently wired would deny nearly everything, which is exactly the hazard staging is meant to surface before enforcement is ever turned on.
• human_issue_raised/2 — NO production seam found; the rule itself is real and tested, but nothing asserts the fact. Investigated and rejected: (1) kernel/pipeline_escalation.py's packet.payload["prolog_facts"] → PrologGate.assert_fact passthrough — fed by LLM-generated free text out of a volk's response, not a deliberate signal (per this ruling's own governing principle: divergence authority belongs to Mohamed alone). (2) orchestrator/pipeline_escalation.py's EscalationInterruptMixin.reject(thread_id, reason) — genuinely human-typed (a required CLI argument on session pipeline reject --thread <id> "<reason>") but keyed by an ephemeral pipeline thread_id, not the plan-document PlanId active_plan/1/plan_step/2 use; asserting human_issue_raised(<plan_id>, reason) from a reject() on an unrelated thread would be a false, invented correspondence between two different identifier spaces. (3) session/controller/_goals.py's GoalMixin.add_blocker(reason) — schema-checked and human-facing in shape, but has zero production callers (unwired dead code; not reachable from any CLI/RPC). No fabricated emitter was wired for this predicate; the rule itself is exercised directly (with hand-asserted, ground, test-scoped facts) in tests/governance/test_constitutional_real_rules.py.

Tests: tests/kernel/test_dispatcher_policy_audit.py (TestArticleXIIAuditModeOff, TestArticleXIIAuditModeOn, TestArticleXIIAuditModeEnforcerFailure) prove the audit-only wiring end-to-end through real dispatch, real CXP transport, and a real Prolog engine, including that a real, actively-parsed plan fixture correctly yields a divergent verdict without blocking dispatch. tests/governance/test_constitutional_real_rules.py (test_article_12_*) exercises plan_fidelity/2's four clauses (vacuous truth, plan-prescribed, divergent, human-issue-unlocked) against the real constitutional.pl via ConstitutionalEnforcer.check_article(12, ...). tests/planning/test_plan_parser.py covers the new status-marker parsing.