Contested decisions

The household request includes: "We are happy to eat leftovers once this week." One member reads that as permission for two leftover nights. Another reads it as at most one. Both are reasonable. There is no fact of the matter to retrieve.

A capable model asked to plan the week will pick an interpretation, fluently and consistently, with the same confidence it uses for allergy checks. The plan looks settled. The disagreement is not. What arrived was a contested decision: competent people disagree, and values have not been chosen.

Most chatbots false-close contested decisions with average answers. Architecture that distinguishes contested from uncertain treats fluency as a warning sign on preferential work, not as a success metric.

Helpful context: Two questions that sort a decision, six factors around a decision, and contestability. Contested work differs from uncertain open work. There is nothing factual left to discover, only a choice to make. How decisions close (next in this track) follows the lifecycle once values settle.

Contested is not uncertain

Teams collapse two open shapes because both feel fuzzy in a backlog. The repair differs.

	Uncertain (open factual)	Contested (open preferential)
Disagreement	Reviewers argue, some answers are better	Reasonable people want different things
Path forward	Evidence, evaluation, repetition	Authority settles values
Model risk	Wrong fact	Hidden policy

Whether a vendor's variance explanation is credible is open factual work. Marginal cases exist, but some explanations are better supported. Whether a quarter of small variances from the same vendor warrants escalation regardless of any single explanation is contested work. It depends on tolerance nobody has decided.

Automating the second question freezes one taste as policy invisibly, because the output looks like the first. The Meal Companion version is leftovers-once language. The enterprise version is escalation thresholds, fairness definitions, and vendor tolerance bands written as if they were facts. The danger is that a fluent system makes contested values look settled: the output reads with the same certainty as a factual answer, and the user has no signal that a value was chosen on their behalf without authorization.

The correct behaviour for a system is to surface the disagreement rather than paper over it.

The cognitive unit · Ch. 6

What surfacing looks like

In quiet policy mode, the system answers as though the organisation had chosen. A contested value becomes a default nobody can point at. Users learn to distrust fluent certainty on questions that were never theirs to delegate.

In surfaced disagreement mode, options, trade-offs, and the person who must choose are visible. Settlement waits for authority. The plan may remain provisional until the household clarifies leftovers-once versus twice. In the Meal Companion, AskTargetedQuestion forms the smallest useful question when information or authority is missing. It does not guess that "once" meant the model's training distribution. Challenge and visibility clauses from the Collaboration Contract apply here: expose conflict, do not smooth it into a balanced-sounding plan.

ContestedAuthority settles valuesOpenCases accumulateClosing

Contested work moves to open only through settlement of values, not through case accumulation alone.

Authority comes first for contested work. Once values are settled and remaining disagreement is about facts or extractable rules, cognition earns its place. How decisions close follows that later movement.

Record the settlement

Settlement is an event. Events need a record: who decided, which values were chosen, effective date, scope. Without it, the organisation relitigates the same contested question with every new model and every new team inheriting the code. The record is what gives later cognition legitimate terrain. A policy to cite beats a preference absorbed from annotations. Endorsed knowledge carries provenance. Silent prompt drift does not. The record also protects the team that made the decision: when a later stakeholder questions the policy, the record shows the reasoning and authority behind it rather than leaving the team to reconstruct intent from prompt archaeology.

Human judgment that should remain applies when nobody with standing has settled yet. Contested work is the case where settlement is the work. Until then, architecture abstains, asks, or presents alternatives with explicit trade-offs.

What this looks like in practice

When leftovers language is ambiguous, the system does not compose a full week assuming one reading. It surfaces the two interpretations and what each implies for variety and effort. It names who in the household settles preferential trade-offs. After settlement, the endorsed reading is stored as knowledge, not silent prompt drift.

For invoice escalation tolerance, the same structure applies with different actors: surface the band options, name the policy owner, record the decision, wire deterministic checks to the endorsed band once settled.

Contested decisions often arrive disguised as requirements bugs. "Make it fair" is not a closed spec. Architecture refuses false closure and routes to authority instead of averaging. Product requirements often smuggle contested values inside factual wording. "Handle leftovers efficiently" embeds a trade-off between variety and effort. "Be conservative on vendor escalations" embeds risk appetite. "Family-friendly meals" embeds cultural preference. Labelling these lines explicitly during refinement and storing settlements as endorsed knowledge with scope prevents downstream composition from becoming accidental policy. The labelling exercise is brief in practice: reading each requirement line and asking "would two reasonable people disagree on what this means?" surfaces contested values in minutes during backlog grooming rather than months into production.

Pairing with leadership and contract

Contested links often require human leadership even when terrain looks familiar. Who should lead each judgment marks them as person-led or person-settles. Challenge and visibility allocations from the Collaboration Contract forbid smoothing.

When settlement arrives, the sort runs again. Contested may become open factual or closed depending on what was decided. A household that settles "leftovers once means at most one leftover night" converts preferential work into checkable knowledge for future runs. Evaluation for contested work differs from factual open work. The measurement asks whether the system surfaced the disagreement, asked the smallest useful question, and abstained from silent policy. Suites penalise fluent false closure.

Anti-patterns in production

Several recurring failures characterise contested decisions handled poorly. Average and ship: models blend competing values into a plan that offends nobody immediately and fails trust later. Defer to the loudest user: the last editor wins without recording settlement, so the next run relitigates. Hide behind confidence: fluency scores displayed as certainty on preferential work train users to stop asking who decided. Automate the symptom: teams add another selector instead of asking who must settle values, and selectors then swallow responsibility above their ceiling.

Each anti-pattern is fixable with architecture discipline: label contested, surface, record settlement, rerun the sort.

Connecting to contestability architecture

Contestability is the product-facing sibling of contested decisions. Contestability features make settlement paths obvious: who can decide, where the record lives, what changes after settlement. Without those affordances, surfacing disagreement feels like an error state instead of correct behaviour.

Design reviews ask what the user sees after surfacing. If the answer is only a chat reply, architecture stopped early. If the answer includes named approvers, provisional state, and record creation, contested work is operable.

Settlement records should capture what was decided, who endorsed it, and which future runs may treat the outcome as closed knowledge. Future abstention then cites the settlement instead of reopening the same dispute silently.

What to do next

The practical starting point is finding one requirement where reasonable people read it differently. Label it contested. Remove silent defaults and add explicit surfacing or targeted questions. Write where settlement records live and who may endorse them.

See how decisions close. Closure follows settlement for contested work, then evidence for open factual work, and the lifecycle is how architecture matures over time.

Read next: How decisions close.