BEFORE THE OUTPUT

The Reverse Centaur

That order keeps the machine in service. Reverse it and approval can conceal a reverse centaur.

Named human owner
Success, non-goals, and stop conditions
Dated before the first output

THREE ROLES, ONE COMMITMENT

One premise travels through four hands

01Owner seals the target first
02Model drafts inside it
03Analyst tests against it
04Director can reopen it before spend

The org chart matters less than the control right: who can change the definition of success before commitment?

THE QUIET SWAP

A KPI is not a fact

Factual premise: how the world is
Value premise: what counts as success
Failure mode: AI ships both, welded together

A KPI looks like a fact but carries a decision premise folded inside. Herbert Simon separates factual and value premises in Administrative Behavior, pp. 131, 177; AI may supply an estimate, but the value choice still needs a named human owner.

Success = ROAS. Chosen by whom?

Treating ROAS as success is a value choice in a technical costume: this quarter's revenue over brand, a hard-won segment, or less dependence on a middleman.

TWO LIMBS, NOT TWO GUARDS

A chain of ratifiers may move nothing

Written before output: human-owned premise
Arrives inside output: system-supplied premise
Same humans, same approvals
Practical authority follows the premise

No whip, just a captured frame. The approvals remain; the authority moved to whoever supplied the premise.

A human loop can still be empty

Ben Green's The Flaws of Policies Requiring Human Oversight of Government Algorithms names the false comfort here (Computer Law & Security Review 45, 2022; DOI: 10.1016/j.clsr.2022.105681). Two ratifiers are not three lines of defense when both inherit the same premise.

THE SKILL THAT QUIETLY LEAVES

Verification can decay into ritual

Budzyń et al., Lancet Gastroenterology & Hepatology 10 (2025), 896-903. Four-centre retrospective observational study; 1,443 non-AI colonoscopies. DOI: 10.1016/S2468-1253(25)00133-5. Published correction: DOI 10.1016/S2468-1253(25)00294-8; the corrected PubMed record retains the design, sample, and result reported here.

The observed deskilling risk is a warning against treating downstream review as a permanent substitute for upstream premise ownership.

Four Polish endoscopy centres; retrospective observational study
1,443 non-AI colonoscopies, three months before vs three months after AI introduction
Unaided ADR: 28.4% to 22.4%; absolute difference -6.0 points
Association, not proof that AI caused the decline

GREEN LOOP, RED FIRM

Green loop, red firm

A green verification loop can wrap a failing firm when no named owner wrote the success premise before output appeared.

System proposes the target
Humans verify against that target
Every check can pass
The missing control sits before generation

WHO OWNS THE SENTENCE

The operator test

Keep the record with the decision log. Then ask the next upstream question: who drew the menu?.

Owner: one accountable human name
Premise: success, non-goals, stop condition
Timing: sealed before first output
Authority: who can reopen it before commitment

NODE 007 · Agenda ControlWho Drew the Menu?Agenda control is the power to draw the option set before the chooser arrives; the visible pick can only run over what reached the menu.

01 · BEFORE THE OUTPUT

Five minutes before an AI drafts a campaign, the team opens a blank premise record. One named human writes what success means, what must not be traded away, and what would stop the work. The record is dated before the model sees the brief. Only then does output appear. That order is the control. If the first complete definition of success arrives inside the model's answer, every approval that follows is already downstream.

02 · THREE ROLES, ONE COMMITMENT

Now let the work move through the firm. The owner seals the target. The model drafts inside it. The analyst measures the campaign against it. The director approves: launch, stop, scale. But the director also retains one control that matters more than the signature: the right to reopen the premise before money moves. Four hands can touch the work. Human ownership depends on when the success sentence was written, and who can still change it.

03 · THE QUIET SWAP

Now remove the premise record. Ask the AI for the campaign, the targeting, and the KPIs in one pass. Success arrives as click-through above one number and cost per acquisition below another. Nothing looks dramatic. The team runs on, faster and cheaper. But Herbert Simon showed that a decision is a conclusion drawn from premises. The system has supplied a value premise as well as factual estimates, and nobody owned that value choice before the output appeared.

05 · TWO LIMBS, NOT TWO GUARDS

Both humans stayed. Neither has a place to challenge the premise. The analyst answers one question well, did we hit the KPI, but that question already grants the KPI is right, and the place to challenge it sits outside her frame. The director approves against a number he never set. Green means go, red means stop. Two ratifiers in a row are not automatically a double safety. Neither owns the premise, so practical authority shifts to the system that supplied it.

06 · THE EMPTY APPROVAL

Add a second approver and you feel safer. You may not be. The approval step exists and can still be empty, because both people inherit the same target. Ben Green's 2022 analysis of human oversight challenged the assumption that putting a person in the loop necessarily gives them a meaningful way to intervene. The practical test is not whether someone can reject an output. It is whether a named person can reopen the premise, record why, and stop the commitment before it moves.

07 · Advertisement · ElevenLabs Studio

The last mile of content is often sound: the right voice, the clean take, the version that works in another market. ElevenLabs brings AI voice, music, dubbing, and sound tools into one production flow, so teams can shape audio with the same care they bring to the story. Create the version people can actually hear.

08 · THE SKILL THAT QUIETLY LEAVES

There is another reason not to make downstream review carry the whole control. In a 2025 Lancet Gastroenterology and Hepatology study, researchers compared non-AI colonoscopies at four Polish centres during the three months before and after AI detection tools were introduced. Across one thousand four hundred and forty-three procedures, unaided adenoma detection fell from twenty-eight point four to twenty-two point four percent. This was retrospective and observational, so it does not prove that AI caused the decline. It does show why a downstream human check should not be assumed to stay constant while the workflow changes.

09 · GREEN LOOP, RED FIRM

The whole machine sits at rest. The system proposes the target. The analyst checks the result against it. The director signs. The KPI loop lights green: target set, target hit, target confirmed, around again. Every downstream check can pass while the missing control sits before generation. Who wrote the success sentence before the system spoke? No one. The numbers are green. The firm is bleeding.

10 · WHO OWNS THE SENTENCE

The operator test is a record, not a feeling. Before the first output, write one accountable human name. Write the success criterion, the non-goals, and the condition that stops the work. Add the time. Add who can reopen the premise before commitment. Keep that record beside the decision log. If it was blank until the system answered, downstream approval cannot repair the missing ownership. Keep the org chart. Keep the approvers. Put the premise first, then ask who drew the menu.