Skip to content
finalstate.aiThe Reverse Centaur
006Atlas

BEFORE COMMITMENT

The Reverse Centaur

A campaign target has an owner and a return route from the draft.The owner records ROAS with a margin floor. The draft returns along a visible revision path. No timestamp or lock determines whether ownership passes.OwnerTargetROASMargin floorDraftReopen
Illustrative workflow. A premise record is a proposed safeguard; its date alone does not establish control.

An illustrative marketing team records its objective, the tradeoffs it will accept, and who can change them before spending. Writing this before generation can expose assumptions early. An informed revision after the draft can also preserve control. A reverse centaur emerges when the organization makes that challenge ineffective.

THREE ROLES, ONE COMMITMENT

The director can return the brief

The model drafts against the chosen objective. The analyst checks the estimate and the campaign results. The director decides whether to commit, with the right to reopen the premise before money moves. A changed objective sends the work back for another check.

A target passes through model, analyst and director; the director returns a changed brief for checking.A campaign slip carries ROAS through three work stations. The analyst checks the result. A return route reaches the owner, who revises the floor; spending remains held while the changed brief requires another check.ModelAnalystDirectorROASROASROASSpendOwnerFloorrevised

THE QUIET SWAP

Separate the estimate from the objective

One campaign sheet separates into an empirical estimate and a chosen objective, each with its own review connection.Revenue divided by advertising cost appears on the estimate half. ROAS as a target appears on the objective half. Separate human review connections reach both, without assigning facts exclusively to machines or values exclusively to people.EstimateObjectiveRevenueAd costROASTargetEvidenceChoice
Simon, Administrative Behavior, fourth edition (1997), chapter I, “Value and Fact in Decision.” Official publisher excerpt: https://www.simonandschuster.com/books/Administrative-Behavior-4th-Edition/Herbert-A-Simon/9780684835822

A forecast of revenue is an empirical claim. Choosing revenue per advertising dollar as the success criterion is a judgment about what to pursue. Herbert Simon distinguishes factual and value elements while explaining that decisions often combine them. Both need accountable review: people choose measurement assumptions, and models can help compare objectives. See decision premise.

Success = ROAS. Chosen by whom?

In this campaign, ROAS uses revenue as its conversion value. Measuring the ratio does not decide whether to prioritize it over margin, customer retention, or brand. The team also chooses the reporting period and which conversions count. Google Ads distinguishes actual conversion value per cost from a target the advertiser chooses.

THE RETURN PATH

A later challenge can change the target

The same model-supplied target has a working revision route on the left and a refused revision request on the right.Both briefs start with ROAS. A challenge reaches the owner on the left and adds a floor. On the right the request meets a barrier and ROAS stays unchanged. A model-supplied first draft is not itself an ownership failure.ROASROASReopenedBlocked+ Floor
Illustrative alternatives. The left branch adds a margin floor and returns the plan for checking. The right branch rejects the request to revise the objective.

Both versions begin with a model-supplied objective. In one, the director presents margin evidence, reaches the owner, and obtains a revised brief before spending. In the other, the organization refuses to reopen the target; the director can only approve or reject against it. The difference is the available decision right, not who wrote first.

Count the rights behind the approvals

Ben Green’s 2022 analysis examines 41 policy documents concerning human oversight of government algorithms. It challenges assumed oversight and calls for evidence about whether people can perform it, alongside scrutiny of adoption itself. Applying that concern to this marketing team is an extension. Two signatures alone do not establish independent assurance: scope, access, escalation and conflicts matter.

RETAINED PERFORMANCE

Review capability needs evidence too

Fixed bars show non-AI ADR of 28.4 percent before AI introduction and 22.4 percent after, on the same zero-based scale.The bar lengths use a common zero-to-40-percent axis. Counts are 226 of 795 and 145 of 648. The six-percentage-point difference and its confidence interval belong to the comparison, not a measure of skill lost.Non-AI ADRBefore28.4%226 / 795After22.4%145 / 648010203040Percent of procedures
Budzyń et al. (2025): 226/795 before; 145/648 after. ADR counts procedures detecting an adenoma or cancer; sessile serrated lesions are excluded. Raw results agree between the published abstract and accepted manuscript. The correction concerns an omitted covariate in an adjusted supplementary model; its corrected table was not obtained, and no adjusted odds are claimed here. Article: https://doi.org/10.1016/S2468-1253(25)00133-5 . Correction: https://doi.org/10.1016/S2468-1253(25)00294-8

A retrospective study compared 1,443 non-AI colonoscopies at four Polish centres in approximately three-month windows before and after AI introduction. ADR was 28.4% before and 22.4% after: a difference of −6.0 percentage points (95% CI −10.5 to −1.6). The first 100 procedures per centre after introduction were excluded. This historical comparison was observational, despite randomization in the parent AI-on/off study. Patient selection, time and diagnostic volume may help explain the difference. It does not prove skill loss or validate premise records. Deskilling examines the retained-capability question.

THE UNCHANGED TARGET

A passed target can leave a problem unresolved

An approved ROAS brief remains unchanged after margin evidence reaches a blocked revision route.Two checks validate the ROAS result. A separate margin marker is below its proposed floor. The request to add that floor meets a barrier, leaving the original brief unchanged. This is an illustrative mismatch, not a plotted business result.ROASMarginFloorReopen
Illustrative campaign, with no measured margins or defect rates. The margin shortfall prompts a request to change the objective; that request is refused.

Return to the illustrative campaign. The analyst confirms that revenue-based ROAS meets the brief, while the director identifies a margin shortfall. Those findings can coexist because revenue per ad dollar does not include all product costs. If the request to add a margin floor is refused, another ROAS approval cannot resolve that dispute. The organization remains responsible for keeping the target in force.

WHO OWNS THE SENTENCE

Trace a challenge through to a decision

Evidence and a named revision authority change the campaign brief; spending is held until the revised plan is checked.Source evidence connects to the estimate, while an owner connects to the chosen objective. A challenge adds a margin floor and returns the plan for checking. The spending route stays closed during that unresolved check.ROASRevenueAd costEvidenceOwner+ Margin floorRecheckSpend held
Proposed operator practice. The record shows a successful revision and a held commitment while the changed plan awaits checking. A named owner or an early date alone is insufficient.

Record who chose the objective, who checked its factual assumptions, what must not be traded away, and what would stop the work. Name the person who can revise it before commitment. Test that route with a specific challenge and keep the response beside the evidence. If a late revision succeeds, update the brief and check the changed plan before spending. Then ask who drew the menu.

Read the transcript

01 · BEFORE COMMITMENT

An illustrative marketing team is preparing an AI-assisted campaign. Its owner records a return-on-ad-spend target and a margin floor. The model will draft within those terms. The owner also keeps a route back from the draft, so an assumption can be challenged before the team spends. Writing the record early helps expose the choice. A timestamp cannot establish control by itself. If the model proposes the objective first, a person can still examine its consequences and change it before commitment.

02 · THREE ROLES, ONE COMMITMENT

The same target travels through three roles. The model drafts the campaign. The analyst checks the estimate and the results against the brief. The director considers the commitment. The director needs the right to reopen the premise before money moves. Here, the director returns the brief to its owner, who tightens the margin floor. That revision sends the campaign back for checking. The first approval does not automatically cover the changed objective.

03 · THE QUIET SWAP

Remove the agreed brief and ask the model for a campaign, forecast, and success criteria together. The answer can bundle an estimate with a choice of objective. Revenue divided by advertising cost is a quantity the team can estimate or measure. Making that ratio the target requires another judgment. Herbert Simon distinguished factual and value elements in decisions, while recognizing that they often arrive together. Separate their review. Someone must examine the evidence and measurement assumptions; someone must accept responsibility for the objective. A model can assist with either task.

04 · THE DISGUISED JUDGMENT

Return on ad spend is a measurable ratio. In this example, its numerator is revenue. Choosing that ratio as the campaign objective does not settle whether margin, retention, or brand should take priority. Nor does the ratio choose its own reporting period or decide which conversions count. Those choices belong in the brief, where the team can examine them. ROAS can be useful without answering every question about success.

05 · THE RETURN PATH

Both versions start with the model supplying the target. In the first, the director brings margin evidence to the owner. The owner reopens the objective and adds a floor. The campaign returns for checking before spending. In the second, the director makes the same request, but the organization refuses to revise the target. The brief stays unchanged. The director can reject a campaign for missing ROAS, yet cannot obtain a decision about whether ROAS is sufficient. The test is whether an informed challenge can change the target before money moves.

06 · THE EMPTY APPROVAL

Ben Green’s 2022 analysis examined forty-one policy documents concerning oversight of government algorithms. It questioned the assumption that a human approval step provides effective protection, and argued for testing oversight and scrutinizing the decision to adopt the system. This marketing example extends that concern into a different setting. A second signature does not establish independent assurance. Ask what each reviewer can inspect, what each can challenge, and where a dispute goes. A different person or another model can still be confined to the same narrow test.

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 · RETAINED PERFORMANCE

Authority also needs the capability to use it. A 2025 study compared one thousand four hundred and forty-three non-AI colonoscopies at four Polish centres, in roughly three-month windows before and after AI introduction. Detection fell from twenty-eight point four to twenty-two point four percent. This endpoint counted procedures finding an adenoma or cancer, excluding sessile serrated lesions. The historical comparison was observational. The first hundred procedures per centre after introduction were excluded, and patient selection, time, and diagnostic volume could affect the result. A correction concerns an adjusted supplementary analysis; the displayed figures are the raw comparison. This is evidence to investigate retained performance. It does not prove that AI caused skill loss, or that a premise record would prevent it.

09 · THE UNCHANGED TARGET

In the illustrative campaign, the analyst confirms that the ROAS target is met. The director then identifies a margin shortfall. Revenue per ad dollar can satisfy the brief while product costs leave too little margin. The director asks to add a margin floor, but the organization refuses the revision. Another approval against ROAS leaves the same dispute unresolved. The model supplied a proposal; the organization kept it in force. Human control requires a route to reconsider that choice, with evidence and time to act before the commitment.

10 · WHO OWNS THE SENTENCE

Take one campaign decision and trace a challenge through it. Record who chose the objective and who checked its factual assumptions. Keep the evidence, the unacceptable tradeoffs, and the stop condition. Name the person authorized to revise the objective before commitment. Then test the return route with a specific challenge. In this example, the owner accepts a margin floor and sends the changed plan back for checking. Spending remains on hold. A late revision can restore control if the people have the evidence, authority, and time to complete it. Keep that response with the decision log, then ask who drew the menu.

01 / 10 · BEFORE COMMITMENT0:00 / 6:16