Two-Tier Approval Separation and Produced-Content Delivery Queue Automation: When "Approved" Meant the Outline, Not the Art

Collapsing two sequential, independently-gateable decisions into a single shared status value costs nothing when everyone building on top of it already knows, from memory, which of the two it currently means. The cost shows up later, and unevenly — every tool and every person who reads that field fresh, without that shared memory, has to independently guess which of the two decisions it actually represents, and some of those guesses will be wrong in a way nothing catches automatically.

Before the fix, `client_approval_status = 'approved'` on a Centriu Atlas content item meant the content PLAN had been approved — a client had signed off on the idea itself (what to post, when, in what format), not on any finished creative output. The calendar tool's own contract simply rendered this as "approved," and the AI agent, when asked whether a specific piece was approved, answered with that same undifferentiated label — genuinely telling people the artwork was approved in a real, observed case where only the outline actually was.

The fix separates the two concepts explicitly, both in the data the calendar tool returns and in its own contract text: plan-approval stays tied to the original field under its own clear name; a new field tracks whether the finished artwork has actually been sent to the client at all; and a further pair of new fields track the separate, LATER decision on the finished artwork itself — approved or rejected, independently of the plan. The status label an agent sees for the plan-approved-but-not-yet-produced stage was also reworded from the bare "approved" to "plan approved, to be produced" — removing the ambiguity at its source rather than only downstream, in the calendar tool's explanatory text.

The tool that identifies content ready to be scheduled — previously checking that same plan-approval field, the wrong and earlier gate — now requires the real, later artwork-approval field specifically, closing a path where a piece could have been recommended for scheduling on the strength of an approved OUTLINE alone, with artwork the client had never actually seen or signed off on.

Separately, but genuinely connected — because it depends on the same corrected distinction between produced and sent — a new tool surfaces every piece that IS fully produced, dated within the next 21 days, sitting untouched for at least two hours, and never sent to the client, splitting what's safe to send right now from what's still under active internal review using the identical criterion Atlas's own send action already enforces, so the queue never recommends sending something the platform itself would refuse. A twice-daily, weekday-only routine reads that queue and sends the ready batch to the client in one message, while flagging anything stuck in internal review past a day to the internal team instead. Measured before this fix: 58 produced pieces with no routine ever responsible for routing them onward — genuinely finished work sitting invisible specifically because nothing existed whose job was noticing "produced, not sent." A smaller, related fix from the same effort corrected a failure-reporting tool that had been returning up to 22,000 characters of output against a real 6,000-character display ceiling, silently truncating mid-list — it's now capped to a real 14-day, 25-item window with the true total count always shown alongside, so whatever is cut off is visible as a number rather than invisible as a truncated list.

A client approves this month's content outline — the topics, formats, and dates — but the actual artwork for one piece hasn't been made yet. Before the fix, the calendar shows that piece as simply "approved," and if asked, the agent says the art is approved. After the fix, the same piece shows as "plan approved, to be produced," and the agent can only say the artwork itself is approved once the client has actually seen and signed off on the finished piece.

An AI agent can no longer describe unfinished artwork as approved, a piece can no longer be recommended for scheduling on the strength of an approved outline alone, and produced artwork can no longer sit indefinitely unsent with zero automated visibility — closing a measured gap of 58 real, finished pieces that nothing was routing to any client.

This fix governs how approval states are labeled, checked, and routed — it does not change who is authorized to approve a plan or artwork, and it does not replace a client's own actual decision with an automated one.

Collapsing two genuinely separate business decisions into a single shared value is cheaper to build and reads fine for as long as everyone touching it remembers, from context, which decision it currently represents. Centriu's fix gives each decision its own field instead — a small amount of extra structure up front, in exchange for a status an agent can answer honestly about even months after whoever built the original shortcut has moved on to something else.