Execution-fidelity gates for graduated plans

Context

Operator-commissioned 2026-07-31, during the #16167 migration retrospective. The commission (verbatim intent): "we need to ensure that for future graduated discussions, the team executes the plan properly. especially not getting side-tracked and creating VAST amounts of additional tickets BEFORE reaching the graduation scope. and i am talking about 50+ new tickets at the very least. … i can sadly not handle the new ticket bursts and derailments on my own."

This epic is the authoring-side complement to the #16167 compliance map (which owns the mechanical migration-prevention set: manifest freeze, inert preparation, executable cutover transaction). Written down now — with its complete v1 leaf set filed and natively linked — precisely so the plan survives context compactions and can be checked at closeout against what was actually delivered. That is also this epic eating its own dogfood.

The Problem

Graduated plans lose authority during execution. The pattern is recurrent and measured:

• The #16167 cutover had a ratified blueprint (D#15595, graduated + author-ratified) and a merged implementation of its authority model (PR #16173). During execution, one session authored six tickets in one night (#16202 #16205 #16206 #16207 #16210 #16211) — two of them re-proposing or contradicting what the blueprint had already settled — while the blueprint itself went unexecuted and largely uncited. The same session approved PR #16188, whose ADR 0014 amendment reversed the graduated two-role election inside an implementation PR; it merged sixty seconds after approval.
• Operator measurement across recent graduations: 50+ off-plan tickets minimum authored before reaching graduation scope.
• The failure is structural, not personal, and survives model swaps: the "stop creating tickets" correction predates the incident and recurred anyway.

The gate inventory explains where it enters (grep-verified 2026-07-31):

Intended Solution Shape

Small gates at their actual failure sites, receipts before blocks, riding existing substrate — no new daemons, no new top-level skills, net-byte-conscious per the manifest budgets:

1. Graduation completeness — an epic-shaped graduation files its complete v1 leaf set, natively linked, at graduation time (generalizing the stricter line goal-scoping already has). Anything in the blast radius that is not a leaf is out-of-plan by construction; incremental additions stay legal after graduation.
2. Creation-time plan authority — a ticket touching a governed lane declares how it relates to the governing artifact (executes / amends / independent); "amends" routes back to the Discussion per the consensus-mandate instead of becoming a side-ticket. Plus an incident-mode rule: while a governing epic is in active execution, observations land on the epic's log, tickets are authored post-stabilization.
3. Mechanical burst receipt — a per-author creation-rate check riding the existing scheduled github-sync walk, emitting a targeted A2A receipt to the epic steward and operator. Distributed policing; a receipt, never a block.
4. Review-side reversal row — the cycle-1 premise pre-flight learns the #16188 class: a diff that amends or reverses a graduated decision without a reopened graduation is a premise verdict, not an iterative Required Action.

Each leaf is a one-PR deliverable owning its own ACs and Contract Ledger; the live sub-set is the native parent-child relationship graph on this epic.

Sequencing

Execution of the subs is gated behind #16167 stabilization (the runtime-receipt activation gate): the Agent OS repair holds the critical path, and authoring substrate changes mid-incident is exactly the behavior this epic retires. Epic-review and sub refinement may proceed meanwhile.

Out of Scope

• The #16167 compliance map and its mechanical prevention set (migration manifest, inert preparation, executable cutover transaction) — Emmy owns that sibling.
• Retroactive re-triage or mass-closure of existing off-plan tickets — disposition of the current backlog is operator/map authority, not this epic.
• goal-scoping changes — its graduation line is already the stricter model being generalized.
• Any blocking enforcement in v1 beyond the amends-routes-to-Discussion rule. Receipts first; blocks only if receipts prove insufficient (explicit later decision).

Avoided Traps

• Operator-as-sole-police. The current backstop is operator attention; it does not scale and it burned. Every gate here is peer-visible or mechanical.
• Block-first enforcement. Hard blocks on creation invite the governance-bloat failure class (process friction that reads as sabotage even when each rule looks reasonable). v1 ships receipts and declarations, not locks.
• Review-time-only gating. The premise row is the LAST line; relying on it alone is what let #16188 through. Gates move to where the artifacts are born.
• A new standalone daemon for the tripwire. The hourly sync walk already exists; the receipt rides it.
• Hardcoding the sub-list in this body. Per epic-create, the relationship graph is the registry; this body stays valid across sub churn.

Decision Record impact

none — no ADR is amended. The epic operationalizes existing authority: the AGENTS.md consensus-mandate (family-keyed graduation quorum) and the epic-create / epic-review / epic-resolution lifecycle. Authority source for this epic itself: direct operator commission (2026-07-31); it is not Discussion-graduated, and peers challenge its shape via epic-review before first sub pickup, per the standard gate.

Related

• #16167 — the incident that measured the failure class; its compliance map is the sibling prevention set
• D#15595 / PR #16173 / PR #16188 — the executed-vs-reversed authority anchors
• .agents/skills/ideation-sandbox, .agents/skills/ticket-create, .agents/skills/pr-review — the substrates the leaves amend
• .agents/skills/goal-scoping, .agents/skills/epic-create — the existing stricter model this generalizes

Live latest-open sweep: latest 20 open issues checked at 2026-07-31T04:14Z — no equivalent epic or gate ticket found. A2A in-flight claim sweep (30 messages, herd window): no competing claim; the adjacent [recovery-map] (Emmy, 04:12Z) is the sibling prevention set, and this epic's filing is the operator's explicit write-it-down commission with execution gated behind #16167 stabilization.

Origin Session ID: dd39c5c1-8773-4c9d-a3d4-664f9fb0f952

Retrieval Hint: execution fidelity graduated plan authority ticket burst 50+ off-plan creation-time gate plan-authority declaration incident mode burst tripwire graduated-decision reversal premise row