Can we prepare a response without committing to one package?
The client needs privacy in either case. It can offer five budget units with four weeks, or eight units with two weeks. These are the two offers on the table. Prepare the response now, then commit when the client confirms which offer it can make.
The candidate pool in the next panel also applies here. The comparison uses the same three package costs, durations and privacy conditions.
1. Resolve the context
Start with eight possible combinations of privacy, budget and time. Each control rules out worlds inconsistent with what is now stipulated. The three candidate packages and their declared meanings stay fixed.
These are ephemeral teaching controls. Nothing is saved, approved, commissioned or changed in the campaign. Reload resets this page. Changing the candidate relation starts another model comparison; it is not a base-context arrow.
| Package | Cost | Duration | New privacy condition |
|---|
Durations are elapsed package times; packages are alternatives, not a resource schedule. The existing 17-task campaign and its other policies remain in Essai 01. This finite model does not claim that these are all possible responses.
2. Watch interpretation compose
Each box counts accepted (world, package) pairs. It is not a count of tasks or completed milestones. Narrowing in either order yields the same final relation because intersection commutes for these independent stipulations.
Privacy → budget → time
Time → budget → privacy
Inspect every remaining world and its actual obligations
| Context world | Baseline | Consent patch | Aggregate reporting |
|---|
The direction matters. Here an arrow means “this more specific context is included in that less specific one”. Changing a known 8-unit budget into a known 5-unit budget is not such an inclusion. These incompatible points have a common less-specific context; reinterpret from that context. This example does not prove that real chronological changes commute.
3. Now—and only now—choose a response
A candidate must work in every remaining world to be eligible for this policy. This adds an explicit robust-choice rule after interpretation. The fibration neither ranks packages nor invents the consent patch or aggregate-reporting work.
A choice made too early need not survive new information
Hold privacy required and 4 weeks available. Before the budget is resolved, the only package that works at both 5 and 8 units is aggregate reporting. The shortest-robust-package policy therefore selects it. Now resolve the budget:
The practical distinction: transporting a previously selected plan preserves that choice; reconsidering after interpretation may select a different plan. A cached recommendation can remain feasible yet cease to be preferred. This comparison assumes the recommendation remains open to reconsideration; a committed or irreversible decision has different available actions and cannot simply be reopened by this calculation. The earlier campaign's minimal/strict/pivot policy menu was never a choice of cleavages: its policies add or choose work, while a cartesian interpretation must retain the factorisations above. This two-package witness makes that difference checkable.
4. The empty result is the important result
Try the tight context. All three candidates fail for a stated reason. The full model has an empty relation, so its lift exists and communicates failure. Removing that object leaves no lift for this particular source and context arrow. Declaring every fibre to contain only “successful plans” has silently broken the claimed fibration.
Even a nonempty relation does not establish a robust plan: different packages may work in different remaining worlds. Step 3 checks that stronger question separately.
What has actually been established?
Interpretation can be derived before repair is chosen. The construction below is a split fibration of relations over contexts. Its laws are ordinary set laws; the small campaign interpretation is ours. This does not make the earlier campaign hierarchy a fibration retroactively.
The construction and its universal property
Fix a finite set W of context worlds and a set P of candidate packages. Base objects are subsets C⊆W. There is one base arrow C′→C exactly when C′⊆C. A fibre object over C is any relation R⊆C×P, including the empty relation. Fibre arrows are relation inclusions.
(C′,R′) → (C,R) iff C′⊆C and R′⊆R.
Projection: (C,R) ↦ C.
Chosen lift over C′⊆C: (C′, R∩(C′×P)) → (C,R).
Factorisation proof. Let D⊆C′ and let (D,S) map to (C,R). Then S⊆R and every pair of S has its first coordinate in D⊆C′. Therefore S⊆R∩(C′×P), giving the required arrow to the chosen lift. There is at most one such arrow because the total category is a poset. Thus the lift is cartesian.
Split laws. Restriction to C changes nothing. For D⊆C′⊆C, first restricting to C′ and then to D is literally the same relation as restricting directly to D. Those identity and composition equations provide the chosen splitting. They hold for arbitrary sets, not just this eight-world calculation.
Why the failed version has no lift. Choose a nonempty R whose accepted pairs all lie outside a nonempty C′. Any arrow over C′→C must have relation S⊆R with first coordinates in C′; hence S=∅. If empty relations have been deleted, no candidate object can serve as a lift. This is an existence failure, not merely a failed choice of cleavage.
The ordinary baseline and what the extra language earns
A well-written table filter returns exactly the same relation. The app uses that direct filter as an independently checked baseline. Category theory has supplied an explicit contract for interpreting repeatedly, a universal property that prevents arbitrary repair being smuggled into a “lift”, and a precise counterexample to the successful-plans-only story. It has not supplied exclusive computational power or measured maintenance savings.
The narrowing model deliberately retains possible worlds. A single real programme may change assumptions, commitments or available resources in ways not representable by subset inclusion. That requires a richer base and an independently justified interpretation; this small fibration is not a general change-management theory.
Sources → construction → limits
Bénabou (1985), §9.1–9.4 develops reindexing, cartesian lifting and splitting. The Stacks Project, §4.33 supplies the modern definition. The relation family, campaign reading, failure witness and choice policy are this essai’s construction.
The model is checked against exhaustive small categories and an independent direct enumeration of all 27 filter settings and three candidate pools. Those checks support the implementation; the general proof is above. Cost, duration, adequacy of the privacy responses and the finite candidate set are supplied assumptions. No engineering, legal compliance, completed work or human benefit is established.