Bound the AI answer with a citation pack before it hits the queue.

Answers without evidence are not queue-ready

The agent produces a fluent recommendation. It lands in the review queue looking finished. The reviewer has three minutes and no link to the record the model used. They either rubber-stamp or re-do the work from scratch. Neither is why you added the agent.

Fluency is not evidence. Queue-ready answers need a citation pack: sources shown, a confidence boundary, and an explicit reject-if-missing rule—so reviewers check something real before they act.

This note gives you the TWIMCO Citation Pack Spec. It complements a HITL review queue design (queue mechanics) and shadow-mode success criteria (promote rules). The Spec is about the evidence packet on each item.

Verification-minded delivery—like Sutton|Place—treats inspectable proof as ordinary. Knowledge and clinical-adjacent workflows in the shape of PatientIQ make citation discipline non-negotiable when answers influence decisions. Agent work scopes under AI applications & agents.

Live studio tools such as /tools/ai-workflow-readiness/ help you judge whether a workflow is ready for AI at all. A forthcoming citation-pack helper may assemble these fields; until then, require the Spec before answers enter production queues.

If you are still deciding whether the workflow should use AI at all, start with readiness—not with citation cosmetics on a process that lacks visible state and recovery. The citation pack assumes a job worth reviewing; it does not repair a broken handoff design.

If you are still deciding whether the workflow should use AI at all, start with readiness—not with citation cosmetics on a process that lacks visible state and recovery. The citation pack assumes a job worth reviewing; it does not repair a broken handoff design.

Structurally, treat the citation pack like a boarding pass: without it, the answer does not board the queue. Soft warnings that reviewers can ignore will be ignored under load. Hard reject-if-missing is kinder than silent error after action.

Fields of the Citation Pack Spec

Answer identity. Job or agent name, queue destination, action the answer may authorize if accepted (comment only, propose edit, trigger write).

Required sources. What must be attached or linked: document ids, record urls, retrieval chunk ids, prior ticket ids, policy section anchors. Minimum count you set for the job. “Model memory” is not a source.

Source display. How the reviewer sees sources: inline quotes with pointers, link list with timestamps, side-by-side record peek. If sources are not visible without extra hunting, the pack fails display.

Confidence boundary. What the agent is allowed to claim: factual extract, ranked options, single recommendation. What it must not claim (legal conclusion, diagnosis, irreversible money move) unless a human gate exists. Boundary is a product rule—not a softmax number you invent as precision.

Reject-if-missing. Hard fail conditions: no sources, sources stale beyond a window you define from your data, conflicting sources without a conflict flag, action type outside boundary. Rejected items never look like ordinary queue work—they route to a failure lane.

Reviewer checks. Three to five checks before accept: open at least one primary source; confirm the answer does not exceed boundary; confirm the proposed action matches permissions; confirm conflicts are flagged. Checklists beat vibes.

Audit fields. Who accepted/rejected, time, and whether sources were opened (if your UI can record that). Audit without blame theater—learn where packs are thin.

The unreproducible asset is your filled Spec for a real agent job—the required sources, reject-if-missing rule, and reviewer checks competitors cannot get from “add citations to the prompt.”

Design packs for the job, not for decoration

Different jobs need different packs. A catalog copy suggestion may need product record + style guide. A disclosure summary may need document ids + verification timestamps. A clinical-adjacent summary may need guideline sections + patient-record pointers under your access rules. Copying one pack onto every agent is theater.

Retrieval quality is part of the pack. If the retriever returns irrelevant chunks, the Spec should still reject thin packs—do not ask reviewers to paper over retrieval failure with trust.

Do not invent accuracy percentages for the Spec. Require observable pack completeness. Shadow-mode promotion can use your own scored samples later; this Spec stays about required evidence on each item.

Permissions matter. Sources the reviewer cannot open are not sources. Packs must respect the same access boundaries as the human job.

When the agent proposes a write, the pack must include the before-state and the after-proposal so the reviewer can see the diff—not only a paragraph of rationale.

Prompt instructions that say “cite sources” without schema fields will be ignored under load. Put the pack in the product contract: structured fields the queue UI requires. Free-text “citations” at the bottom of a paragraph are how packs rot.

Prompt instructions that say “cite sources” without schema fields will be ignored under load. Put the pack in the product contract: structured fields the queue UI requires. Free-text “citations” at the bottom of a paragraph are how packs rot.

Operate reject-if-missing without draining the queue

Reject lanes need owners—the same as any exception path. A reject pile nobody works is how teams disable the rule. Staff the lane or narrow intake.

Measure qualitatively at first: are rejects mostly missing sources, stale sources, or boundary violations? Use that to fix retrieval or prompts. Do not invent industry reject-rate targets.

Train reviewers with an operator training path scenario: accept with pack, reject missing sources, escalate conflict. Citation discipline is a practiced skill.

Coordinate with permission matrices and escalation ladders. Citation packs do not replace permission; they inform the human who holds permission.

Publish the Spec beside the agent job spec so tool wiring and evidence rules stay paired. When prompts change, re-check that required sources still match retrieval. Prompt drift without Spec drift is how packs go hollow while demos still look good.

Failure modes the Spec refuses

Ornamental footnotes. Links that do not support the claim. Reviewer checks require opening primary sources.

Confidence as decoration. A green badge without boundary language. Boundary field blocks this.

Silent writes. Answers that trigger actions without packs. Answer identity + reject-if-missing block this.

Queue rubber-stamp. Time pressure without checks. Reviewer checklist blocks this—if you staff time for it.

These are delivery-shaped composites, not client telemetry. Ground the habit in inspectable verification like Sutton|Place and evidence-heavy workflows like PatientIQ.

Refuse to score “citation quality” with invented precision. Prefer pass/fail on required fields and spot audits of whether cited sources support the claim. Support is a human judgment; the Spec makes that judgment possible by forcing sources into view.

Put this into practice

1. Pick one agent job that already proposes answers to humans; draft required sources and reject-if-missing.
2. Define confidence boundary and three reviewer checks; wire reject to a failure lane with an owner.
3. Run a week of shadow or supervised queue traffic; fix thin packs before expanding autonomy.
4. Attach the Spec to the agent job spec and the HITL queue design.
5. Scope builds under AI applications & agents; use /tools/ai-workflow-readiness/ when judging whether the wider workflow is ready.

Let answers enter the queue only when the citation pack is complete—not when the prose sounds sure.

When promoting from shadow mode, require citation-pack completeness as a promote criterion alongside your shadow success card. Autonomy that skips evidence is not a promotion—it is a hope with a higher rate limit.