Who should lead each judgment?

A status update on the Meal Companion says that planning is autonomous and that purchases require household approval. It is a fair summary, and it is a description of a product, which is why it covers three of the five judgments that produce a week and misdescribes the other two.

The week it fails to describe is an ordinary one. A household rejects a meal on preference grounds, and the system that composed the plan is the wrong party to decide which replacement the household would rather eat. Later the same week acquires a dietary restriction nobody in the product has seen before, and planning becomes something the household leads while the software organises options. Neither link is an exception to the summary. Each of them has a leader, and the summary named a different one. Leadership attaches to a decision and never to a product, so every link already carries an allocation whether or not anyone chose it, and the ones nobody chose were made by whoever implemented them.

Helpful context: Established, mixed, and frontier regions supplies the terrain reading this page converts into an assignment, judgment chains supplies the links to assign, and authority is granted supplies the reason capability never settles the question. The Collaboration Contract (next in this track) covers the terms the sharing runs under once a leader has been named.

Five modes, and none of them is a level

Automated and human in the loop are the two phrases most products reach for, and neither can record an allocation, because both describe a workflow rather than a link. Recording needs a vocabulary fine enough to separate three questions that ordinary usage runs together: who forms the decision, who authorises it, and who performs the action that follows. Five combinations of those answers recur often enough to be worth naming, and they are the working vocabulary for everything below.

Mode	What the system contributes	What the person keeps
Thoughtware decides and acts	forms and executes a bounded decision under a standing grant	the grant itself, and a route to reverse
Thoughtware decides, person approves	a proposed decision with the evidence behind it	authority over each instance
Thoughtware recommends, person decides	comparison, consequences, and a preferred option	the choice and its consequence
Person decides using analysis	evidence, scenarios, and criticism	the framing and the decision
Person leads, Thoughtware assists	selected bounded steps on request	the whole problem

Read down the last column and the five look like a ladder, with a maturing product climbing toward the top row. They are not a ladder. Each mode is correct under conditions that make the other four wrong, so a link moving from person approves to decides and acts has improved only if the ground beneath it changed. Where the move happened because the approval queue was slow, nothing matured. A decision was reassigned on operational grounds while the record continued to claim otherwise.

The mode is only a name. What each one obliges the surface to provide belongs to the Collaboration Contract, and the two are worth keeping apart, because a mode named correctly and built badly fails in a way no allocation table can show.

Terrain does most of the assignment, then stops

Most of the assignment follows from the ground. Judgment terrain reads a link on six conditions, and a favourable reading across them supports system leadership while a hostile one returns it, which is what the three regions compress into a single word. Busy evening fit reads favourably on every condition and lands in the first mode, while an unfamiliar renal restriction reads hostile on every condition and lands in the fifth.

Two things then finish the assignment that are not properties of the ground at all, and the first is institutional policy. A household requires approval before its money is spent, and forty accurate weeks do not weaken the requirement, because the requirement was never a claim about accuracy in the first place. The same structure holds inside an organisation where a manager may approve a refund that a support agent may not. Policy of that kind holds a link with a person in thoroughly established terrain, which means an allocation derived from the terrain reading alone will over-assign to the system every time. Human judgment that should remain covers the decisions that stay with people for reasons stronger than policy.

The second factor is the expertise of the person on the other side, and it is the one that must not move the allocation. A novice needs more explanation, stronger challenge, and clearer comparison than a specialist who wants compact evidence and control of the framing. Both of those describe how the collaboration should be shaped rather than who owns the decision, so a system that quietly takes more authority because the user seems inexperienced has confused the two. User expertise changes the form of the sharing and leaves the owner where legitimacy put them.

One week, four leaders

One week of the Meal Companion is the cheapest way to see allocation working at link scale, because the product is constant across it and the leader is not.

Link	Why the ground reads that way	Who leads
Decide whether a meal fits a busy evening	dozens of comparable evenings behind it, reversible until someone cooks	AssessMealPracticality decides and acts inside limits
Recommend a replacement for a rejected meal	patterned work, but the preference belongs to the household	RecommendMealSubstitution recommends, the household decides
Clarify whether leftovers are acceptable	the missing fact is owned by the household	AskTargetedQuestion asks, the household answers
Interpret an unfamiliar medical diet	novel, consequential, weakly reversible	the household or a professional leads
Purchase the groceries	exact execution carrying financial consequence	the household approves, deterministic code acts

Five links, four leaders, one product. Nothing about the model changed between the first row and the last. What changed was the ground under the link, and at the purchase row, a grant the ground has no bearing on.

Human and machine judgment should be allocated decision by decision.

Introduction to Thoughtware · Ch. 14

Invoice intake produces the same shape in a different vocabulary. Deterministic code retrieves the invoice and the purchase order, a cognitive unit judges whether the difference is material, an agent investigates the likely cause, a reviewer approves any exception, and code posts the result. A product called AI invoice review describes none of that, because it names a screen, and there are five allocations behind the screen.

The allocation can be wrong halfway through

An allocation is made against the case as it was understood when the run began, and cases do not hold still. The Meal Companion starts an ordinary week in the first mode, composing around two busy evenings under a standing grant. Three messages later the household mentions a diet the product has no precedent for. The reading that licensed system leadership was accurate when it was taken and is wrong now, and the run is already half finished.

What a mature system does at that point is suspend the routine action, keep the work already completed, and state which reading changed. The two alternatives fail in opposite directions. Finishing in the old mode is false familiarity, since the strategy runs straight through the detail that made the case different. Discarding the run fails in the other direction, because it teaches the household to withhold exactly the kind of information that made the change of leadership correct.

Leadership also moves the other way, and designs that contain only an escalation path miss it. Once the household establishes which dietary guidance is authoritative, the novel part of the week is settled, and what remains is ordinary composition, portion arithmetic, and list assembly. A system with no way to resume leaves the person holding a run they never wanted to own, which is a slower way of failing than escalating too late.

An allocation nobody wrote down is still an allocation

Every link in that week behaves in some mode at runtime, so the choice was never between allocating and not allocating. It is between an allocation somebody authored and one that fell out of an implementation. The second kind is neither neutral nor cautious. It is a set of decisions with no author, and that is why it survives review so easily: there is nothing on the page to disagree with.

Writing the rows down makes disagreement local, which is most of their value. A reviewer who thinks the substitution link should recommend rather than decide can argue about one row without reopening the product. Legal asking where the human is gets a row rather than a reassurance. Support tracing a failure finds a named owner for the link that produced it instead of an architecture diagram covering the whole journey.

The record also decides what is worth measuring, because links in different modes do not fail in the same way. A system-led link fails by acting outside the ground it was evaluated on, so what matters is suite performance and how often it escalated. An approval link fails by making refusal impractical, so what matters is whether the evidence was complete and whether anyone has ever said no. A single accuracy figure averaged across a chain carrying three modes captures neither failure. Evaluation as engineering covers how the suites behind those numbers are built.

What to do next

The quickest test for whether a product has an allocation at all is to name one link and ask two people who work on it who leads there. Matching answers mean a row exists somewhere, even if nobody has drawn the table. Diverging answers mean the implementation has been deciding, and the disagreement is worth more than either answer, because it locates the link where the product and its description have come apart.

Naming a leader is where this page stops. It does not settle how far the system may reach before anyone asked, how far a household correction travels afterwards, or what has to be visible before an approval means anything. The Collaboration Contract covers those terms, and any of them can contradict a mode that was named correctly.

Read next: The Collaboration Contract.