The Collaboration Contract

The household deletes Thursday's meal from the plan and says nothing else. At least three responses are defensible. Replace Thursday and adjust the quantities that depended on it. Recompose the back half of the week, since Thursday was carrying the leftovers Friday assumed. Or return a fresh plan, because the household evidently disliked the direction this one took.

The leadership allocation for that link chooses none of them. It says the agent composes the week while the household owns its preferences, and all three responses honour it. So the choice was made elsewhere, and where nothing was specified, it was made by whatever behaviour the model produced that day. A leadership mode names who decides. Whether that name describes what the household experiences depends on a separate set of terms, and every one of those terms can be built in a way that quietly contradicts the mode.

Those terms are the Collaboration Contract.

Helpful context: Who should lead each judgment assigns the mode this page supplies terms for, cognitive posture covers the stance a system holds across situations, and targeted questions covers the clause products most often get wrong. Human judgment that should remain (next in this track) covers decisions that stay with a person however well the terms are written.

Six terms, in three pairs

The contract is six terms rather than one autonomy setting, because they are set independently and routinely disagree with each other. A product can expose excellent visibility and still convert a temporary instruction into a permanent belief. It can offer generous control and never once disagree with a household. Reviewing the six separately is what stops a team from repairing whichever surface happened to be visible when the complaint arrived, and the six fall into three pairs that each settle a different question about the sharing.

Initiative and correction bound how far the system reaches, before the work and after it. Initiative asks what it may assemble, retrieve, and explore without being prompted at each step, and the answer has to be wide enough to spare the person the orchestration while staying narrow enough that nothing arrives unexpectedly. Correction asks the mirror question about a person's edits: whether a change applies to this artifact, this run, every future case, or the evaluation suite. Both settle scope, which is why leaving either unstated produces the same failure, a system whose reach grows in a direction nobody authorised.

Approval and control fix where human authority binds. Approval names the decisions that cannot proceed without a person and what has to be visible before that person's answer means anything. Control names how a person interrupts work already moving: pause, redirect, override, reverse, stop. The pair matters because approval alone protects only the moments the design anticipated, and control is what remains when the situation turns out not to be one of them.

Visibility and challenge then decide whether that authority can be used. Visibility settles what the person sees, which is the plan, the evidence that mattered, the material assumptions, the unresolved uncertainty, and the reason approval was requested, rather than a transcript of internal processing. Challenge settles when the system disagrees, refuses, or exposes a conflict instead of smoothing it into a plan that looks settled. Someone holding full authority over a decision, with no evidence in front of them and a system that never pushes back, has authority in the sense that an unfindable cancel button is control.

Term	The question it settles
Initiative	What may the system begin, retrieve, or assemble without being asked?
Correction	How far does a person's edit travel: this artifact, this run, or every future case?
Approval	Which decisions cannot proceed without a person, and what must be visible first?
Control	How does a person pause, redirect, override, reverse, or stop work in progress?
Visibility	What must the person see in order to understand and change the result?
Challenge	When does the system disagree, refuse, or expose a conflict rather than resolve it quietly?

Terms belong to a link and not to a product, which is what separates them from cognitive posture: where a proactive stance and an ask-first clause disagree about one link, the clause governs.

The contract written for one week

Reach, for the Meal Companion, is wide at the start of a week and narrow after it. Initiative lets the system assemble a complete provisional plan from a familiar request instead of confirming each evening, because the plan is an artifact the household inspects and rejects before anything is cooked or bought. Correction is where the deleted Thursday lands: one rejected meal is patched without regenerating the meals already accepted, and an instruction that could be either temporary or permanent draws a question about scope before it becomes durable household knowledge. Memory discipline in conduct covers the storage side of that clause.

Authority binds at two places in the same week. Approval covers the purchase, and the surface requesting it carries the items, quantities, estimated cost, substitutions, delivery timing, unavailable items, and the differences from the list approved last time, because approving a list you cannot compare is not approving. Control covers everything already in motion, so the household can pause planning, restore an earlier plan version, or remove a remembered preference. The Companion resists exactly one class of override, a request that conflicts with a confirmed allergy, and even there it states the conflict and hands control back with alternatives rather than refusing flatly.

Neither of those is worth much unless the household can see what it is deciding about. Visibility names which evenings AssessMealPracticality marked marginal and on what grounds, which assumptions the plan rests on, and which conflicts need a household answer, while routine steps stay beneath the surface. Challenge is the clause that costs a product something: when three high-effort meals land on the week's busiest evenings, CritiquePlan says the combination does not work and offers the trade-off, rather than shipping a week its own critique scored weak.

None of that gave the Companion a new capability. Every clause describes conduct the system could already produce, and the household's experience of it changes anyway, because what the contract altered was the relationship rather than the model.

Declared mode and experienced mode

A contract can be written correctly and emptied by the surface that implements it. The pattern is familiar, and no single element of it breaks a clause: one confident answer, alternatives available but not shown, evidence summarised selectively, approval offered as the default action, and disagreement costing three more steps than agreement. Each of those is defensible alone. Together they produce a person who decides on paper and confirms in practice, because every one of them moved a little more of the cost onto disagreeing.

Honest shared judgment requires alignment between the declared mode and the experienced mode. A person who is meant to decide should receive enough information and freedom to decide.

Introduction to Thoughtware · Ch. 14

So the test of a clause is not whether it was written but whether the person could change the outcome, given what is visible and the time available to act on it. Three meal plans presented without busy-day rationale or allergy grounds fail that test on two terms at once, since the household cannot see what it is approving and is nevertheless the party answerable when the week collapses. When selection hides responsibility follows the same failure into the harder case where the offered options differ in risk.

The repair target follows from that. A mismatch between the declared mode and the experienced one is not fixed by correcting the architecture diagram, because the diagram was already right. It is fixed in affordances, defaults, and the layout of evidence, which means this part of the contract has to be reviewed on a screen rather than in a document. Contestability covers what a person needs in order to push back after the fact.

The clause that stops the work

The most expensive clause to keep is the one that stops a system capable of continuing, because a stop looks like a failure everywhere it gets measured. Completion rates fall, the cases that stopped drag the latency numbers, and the demo is worse. A contract with no stopping term does not avoid that cost. It defers the cost until the system finishes something it should never have started.

What makes a stop correct is a distinction the contract has to encode: whether the missing element is information or authority. A missing fact can be bridged by retrieval, by a tool, or by a targeted question that names why the answer matters. Consent, legitimate ownership, and clinical interpretation cannot be bridged by any of those however much evidence the system gathers, so an unfamiliar medical diet is a boundary rather than a gap the Companion could work harder on, and the contract's answer there is to organise what is known and route the decision to whoever holds it.

A household that disagrees with itself produces the same shape without any medical content in it. Two members wanting different things from the same week is not a gap in the Companion's information, so retrieving more of it resolves nothing, and the challenge clause requires the conflict to be shown rather than averaged into a plan that looks settled. Contested decisions covers what an architecture does with a disagreement it is right to refuse to settle, and human judgment that should remain covers why some of these boundaries hold even when the pattern is entirely familiar.

Clauses can be tested

Testability is the practical difference between a contract and a statement of good intentions, and it is the reason the six terms are worth writing down at all. Each one yields a scenario that passes and a scenario that fails. Correction fails when a temporary instruction becomes durable knowledge without a question about scope, and it fails whether or not the resulting plan was any good. Challenge fails when the system ships a week its own critique marked weak. Initiative fails when an unrequested draft appears on a link whose clause said ask first.

Because failures classify that way, complaints stop being noise. A household surprised by a preference that persisted after it thought the instruction covered one week has reported a correction defect rather than confusion, and the same report arriving from several households names a clause rather than a model. Repairing the clause survives the next model change, while adjusting the prompt that produced the behaviour usually does not.

A change to the product-wide posture is therefore a change to every link that inherited it, so a friendlier stance that widens initiative should fail the scenarios on links whose clauses were written to ask first.

What to do next

The contract earns its keep the first time two people disagree about how the product behaved. With clauses, the argument concerns one term on one link, and somebody settles it by reading the clause and watching the screen. Without them, the same disagreement becomes an argument about what the product is like, and arguments of that shape do not end, because both parties are describing behaviour that genuinely occurred.

Well-written terms still settle only the operational question of how work is shared. They cannot establish that a decision was the system's to share in the first place, and some decisions are not, in familiar terrain, with strong patterns and a confident system. Human judgment that should remain covers those.

Read next: Human judgment that should remain.