1. Declare the decision, not just the level of detail
A fictional team must allocate its next work period to North, South, or a balanced package. It knows that demand at each location may be low or high. A weather factor can shift every option's benefit equally; an optional coupling also penalises the balanced package. Benefits, penalties, feature costs and available actions below are stipulated, not estimated.
What does the current description tell us?
Controls only change this page. Nothing is saved or sent. Reload restores the declared starting example. There is no hidden “true world” used to select a favourable question.
2. See what the abstraction is hiding
Loss is the benefit of the best action in that world minus the benefit of the action chosen from the coarse description. Each action receives its worst loss over all worlds still compatible with that description.
| Action | Worst loss | Worlds attaining it | Decision status |
|---|
Inspect every remaining world and its benefits
South = 10 + 8y + 20z
Balanced = 14 + 3(x+y) + 20z − coupling·z
x, y, z ∈ {0,1}; coupling ∈ {0,6}.
| World | North | South | Balanced | Best action(s) |
|---|
Why a small numerical error is not the only route to a useful abstraction
The common weather shift can make every predicted benefit inaccurate by the same amount while leaving every action comparison unchanged. A decision-specific certificate can ignore that distinction. When weather also changes the relative merits of Balanced, that permission may disappear.
3. Refine only if the certificate fails
Observation costs are North 2, South 2, Weather 1. These are separate units from decision loss: the search minimises observation cost subject to the loss tolerance. Costs are not subtracted from benefits. This is a fixed set of observations before the action, not an adaptive decision tree.
Compare all possible refinements
| New distinctions | Cost | Worst loss after every possible answer | Meets tolerance? |
|---|
Buy the weather observation first
With weather coupling 6 and a loss tolerance of 1, the cheapest fixed set costs 3. A sequential decision can stop early. Spend 1 to learn the weather, then decide whether a demand observation is worth its cost of 2.
The result: a stopping rule and a reopening rule
Stop adding detail when the current description admits an action whose worst possible regret is within the declared tolerance. Reopen the description when the tolerance, available actions, payoff relation or compatible worlds change enough to invalidate that certificate. A disagreement with a fully informed choice is a counterexample to perfect accuracy; it is not automatically a counterexample to usefulness.
At the starting settings, Balanced has worst loss 1. It is optimal in four of the eight worlds and loses 1 in the other four: the coarse description passes a tolerance of 1. Requiring zero loss makes both demand distinctions necessary, while weather stays irrelevant. Adding the six-unit coupling makes weather decision-relevant; the complete refinement search then finds two equally cheap solutions: Weather plus North, or Weather plus South, each costing 3. Neither requires all three features at tolerance 1. In weather state 0, Balanced already meets that tolerance; in state 1, knowing either demand is enough to select an optimal directed package. Try these three cases above.
The exact finite certificate and its proof
Let C be the nonempty set of worlds consistent with the current description, A the finite nonempty action set, and Q(w,a) the stipulated benefit. No distribution over worlds is assumed. Define
R(C) = min_a max_{w∈C} L(w,a).
An action with loss at most ε in every compatible world exists if and only if R(C)≤ε. This follows directly from the minimum over all possible deterministic actions. The witnesses in the table make both passing and failing claims inspectable. Ties are retained.
A refinement partitions C into nonempty cells C₁,…,Cₙ and permits a different action after each answer. Its exact certificate is maxᵢR(Cᵢ). Because Cᵢ⊆C, every old action remains available and its maximum loss cannot increase; hence refinement cannot worsen this certificate. Enumerating all subsets of available features therefore gives an exact minimum-cost fixed refinement in this finite vocabulary. It does not prove minimality among unmodelled observations, mixed actions, adaptive queries or new action designs.
The generic 2r bound. Let q̂ be a representative payoff vector, with |Q(w,a)−q̂(a)|≤r for every compatible world and action. Choosing an action maximising q̂ gives regret at most 2r: compare the true best action to its representative value, then the chosen representative value to its true value. Here q̂ is the componentwise arithmetic mean used as a geometric centre, not a probability forecast. The bound is valid but can be loose; the app also calculates the actual worst regret of that representative choice.
Sources → new construction → limits
Simon, The Architecture of Complexity (1962), read in full afresh, motivates compressed descriptions that preserve useful relations within a range. Near-decomposability is conditioned on interaction and time scale; it is not a licence to discard arbitrary distinctions.
Girard and Pappas (2007) quantify errors in observations and trajectories of transition systems. This essai takes the narrower, one-decision case and explicitly connects output error to decision regret. It does not implement or claim approximate bisimulation. The minimax-regret certificate, refinement search and synthetic example are this essai’s construction, not Simon's or Girard–Pappas's theorem.
The certificate is conditional on the world set, payoff model, action set and tolerance. An omitted world or unmodelled objective can invalidate it. Feature acquisition is perfect, immediate, and non-invasive here; no temporal or organisational safety guarantee follows. The next substantive extension would test noisy or delayed observations and sequential actions, where a static certificate is insufficient.