Human judgment that should remain

A promotion committee receives an analysis it cannot fault. Performance evidence is assembled, the work is compared against declared expectations, inconsistencies across three review cycles are flagged, and similar cases from previous years are surfaced. The analysis is more thorough and more consistent than anyone in the room could have produced alone. Nobody in the room thinks the system should sign the decision.

That reluctance is not a complaint about accuracy, which is what makes it worth examining. The terrain here is familiar: the pattern is dense, the evidence is structured, and the organisation has handled hundreds of these. A decision can sit in exactly that terrain and still belong to a person, because terrain answers which participant is best positioned to do the work, while ownership turns on who holds the standing to be answerable for the result. Capability does not accumulate into standing, which is why this class of decision does not move when the next model ships.

Helpful context: Who should lead each judgment? allocates leadership link by link, and judgment terrain owns the six conditions that allocation reads. This page covers the conditions that can override the reading. Human judgment remains part of the system covers how the resulting human step is then placed in the specification.

Best placed is not the same as entitled

Leadership allocation asks a performance question. For each link in a chain it weighs novelty, evidence coverage, consequence, and reversibility, then names the participant most likely to produce a good result, which is what makes who should lead each judgment the first pass over any chain. That allocation stays necessary. The promotion room shows that it is not sufficient, because a second question can override its answer: who has the right and the responsibility to own this decision.

The two questions disagree in both directions, and the second direction is the one teams miss. A system may be better placed to weigh a large body of evidence while a person still has to decide, because the decision expresses what the institution values. Equally, a person may apply a policy less consistently than software would, and the institution still remains accountable for that policy, for its exceptions, and for the consequences of enforcing it. Consistency is therefore an argument about quality and never an argument about ownership.

The Meal Companion carries the same split in miniature. After forty weeks it judges busy-evening fit better than a tired parent manages on a Tuesday, and it still may not buy the groceries. Purchase is exact execution with financial consequence, and what the household granted was permission to prepare a basket rather than permission to spend. Authority is granted develops that separation, and its consequence here is narrow: permission arrives through an explicit act, so no accumulation of demonstrated skill issues it.

What legitimacy is made of

Two of the conditions concern who stands behind the decision once it is made. Moral authority is the right to make a call about what ought to happen, and institutional accountability is the requirement that a named person or body can explain that call, defend it, correct it, and absorb its consequences. An employee who was passed over is owed someone who can be argued with, who can change the criteria, and who can provide a remedy. A Thoughtware system can record an explanation, produce an apology, and perform a corrective action, and none of that amounts to holding responsibility for the original choice.

Three further conditions concern the people a decision lands on rather than the people behind it. Consent asks whether those affected meaningfully agreed to be decided about this way, and accuracy does not supply it, because a model may predict an outcome correctly while the person retains the right to contest the evidence or refuse the action. Contested values appear where reasonable people disagree about which outcome is fair, which leaves accuracy nothing to be accurate about, and contested decisions develops what the architecture does then. Dignity rules out the move the other conditions invite, since a decision about someone's standing cannot be settled by a sufficiently complete record of their behaviour.

Fairness deserves separate treatment because it answers the strongest objection to all of the above, which is that software applies rules more evenly than a tired manager does on a Friday. It does. A rule can also be applied with perfect consistency and remain unfair in its purpose, in the evidence it relies on, or in how its effects fall across people. Consistency is evidence about execution rather than evidence about fairness, and once that is clear, waiting for a better model stops being a plan for this class of decision, because the missing element was never intelligence.

A high-quality recommendation does not acquire moral authority by being accurate.

Introduction to Thoughtware · Ch. 36

Reversibility has to be measured in the person, not the transaction

Consequence and reversibility already appear among the terrain conditions that judgment terrain owns, and reversibility is the one teams most often believe they have handled, because they have built for it. Plan versions roll back. A proposed knowledge record can be rejected. A remembered preference can be corrected or forgotten. An unactioned purchase can be cancelled. All of that is real engineering, and it does resolve a large class of mistakes.

The trouble is where the measurement gets taken. A rejected opportunity, a disclosed secret, a harmful message, a medical action, or a public accusation keeps acting on people after the database has been restored to its prior state, so a reversible transaction is not evidence that the human consequence is reversible. Measured in the person instead, weak reversibility argues for human leadership before the action rather than for better recovery after it, which is the opposite of what a working rollback capability tends to persuade a team of.

The Meal Companion's medical case is the clean instance. A plan version rolls back cleanly and a week already eaten under a mistaken dietary interpretation does not, and no audit trail changes that asymmetry. So the Companion may organise approved guidance, name the terms it cannot resolve, and adapt the schedule around what is known, while the restriction list stays with the household or a clinician. Refuse and defer covers how that stop is expressed as conduct.

Deep support without a transfer of ownership

None of this argues for a thinner system, and the promotion case is where that is clearest. The system may assemble the evidence, surface inconsistencies between reviewers, compare the work against stated expectations, and name what it could not resolve. The committee signs. Support of that depth is what the architecture is for, and it can keep deepening for years without moving the line, because capability never drew the line in the first place.

The Meal Companion's human-led links show the same shape at household scale. Before a purchase the system prepares the basket, explains substitutions, and shows price changes, and the household approves. On an unfamiliar medical restriction it organises what is endorsed and identifies unresolved terms without diagnosing anything. When one member wants dinners built around a training goal and another wants meals a child will reliably eat, it can expose the conflict, propose shared meals with optional additions, compare the practical effects, and ask which priority should govern this week. What it must not do is decide whose preference weighs more.

Link	What support may do	What stays with people
Purchase groceries	Prepare the basket, explain substitutions, show price changes	Approving the spend
Interpret an unfamiliar medical diet	Organise endorsed guidance, name unresolved terms, adapt the schedule	Owning the restriction list
Settle a household disagreement	Expose the conflict, propose shared options, compare effects	Deciding whose priority governs

AskTargetedQuestion is where the distinction becomes executable rather than editorial. It forms the smallest question whose answer would change the plan, which means it first has to establish whether an answer would help at all, and targeted questions develops that test. Where leftover acceptance is unknown, a question closes the gap. Where the missing element is authority, no question closes anything and the decision routes to whoever holds it. How that routing is declared, and what the receiving human step has to carry, is human judgment remains part of the system.

Where support quietly becomes ownership

The realistic failure is drift, not a team deciding to automate promotions, and it happens because excellent support becomes hard to tell apart from ownership. When the analysis is right nearly every time, the approver's practical options narrow to agreeing quickly, and a recommendation that cannot be contested has become a decision whatever the interface calls it. Contestability covers what must be visible for disagreement to be possible, and passive acceptance theatre covers the interface pattern that removes it.

Two routes carry ownership across the line without anyone deciding to move it. The first is capability treated as permission: the system can execute the financial action, so execution gets wired to the analysis, which is why architecture separates proposal, approval, and execution even where the execution is exact code. The second is policy by accumulation, where preferences inferred from behaviour harden into rules that bind people who never endorsed them. A temporary "avoid pasta, we had it recently" becoming a standing household rule is the small version, and knowledge is endorsed is the gate that stops it.

Presentation does the same work more quietly, because a preselected default can carry a decision nobody recorded making, which is the pattern when selection hides responsibility takes apart.

Because the drift is gradual, the defence is a written reason rather than continued vigilance. One sentence naming the condition is reviewable: purchase stays with the household because the spend is consequential and consent is required, which can be checked against the architecture in a way that "high risk" cannot. The sentence also states its own expiry, since it tells a future team what would have to change, and if a grant is later issued for spending under a threshold, the sentence narrows rather than disappearing. Grants need named owners and review dates for the same reason, because the constraint is a policy fact and policy moves without any model moving.

What to do next

A product can pass every evaluation it owns and still be taking decisions it has no standing to take, because nothing in an evaluation suite measures standing. The check that catches it runs one link at a time. For each decision the roadmap plans to automate, ask who would be answerable afterwards, whether that person exists, and whether they would have what they need to disagree. Where the answer is nobody, capability is not what blocks the automation, and no model release will unblock it.

Contested decisions takes the hardest case, where the disagreement is about values and no quantity of evidence settles it. The Collaboration Contract specifies the terms under which support and approval alternate across a chain, and human judgment remains part of the system places the human step on the architecture.

Read next: Contested decisions.