Keeping construction methods open · Essai 4
An option needs a warning contract
Keeping an alternative physically possible does not keep it usable. The warning must distinguish the right alternative and arrive early enough to finish changing course.
The previous essai separated “there is a recovery route in every world” from “there is one strategy we can follow before knowing the world”. This one lets the environment withdraw a method at an uncertain time. It turns the vague promise of an open option into a precise contract between retained physical provisions, information and time.
After partial work on the mountain bridge
The common works are already committed. Completion now requires route A or route B, depending on a hidden site condition. The current opportunity disappears at time 3, 4 or 5. A warning is issued a specified number of time units before that withdrawal, but its delivery may be delayed. Starting the wrong route is irreversible. The selected conversion duration includes all remaining work needed to finish; merely beginning before withdrawal is insufficient.
At the opening settings, the warning is issued three units before withdrawal, may arrive two units late, and completion takes two units. Each world has a viable route if known immediately, yet no observation-based strategy guarantees completion. Reduce the maximum report delay to one: the guarantee appears exactly at the boundary.
Inspect a strategy or its obstruction
A node contains all worlds still compatible with the observed clock and reports. An action must work for every world in that node. Waiting must survive every possible next observation. The trace below follows a winning strategy where one exists, or exposes a failed branch and counterexample worlds.
The option is a composite promise
This equivalence holds here when the warning names the required route and both tailored interfaces are retained. It is a requirement on worst-case delay, not mean delay. An excellent average reporting service can still fail the guarantee.
The necessity proof uses two worlds with the same earliest withdrawal and latest allowed report arrival, but opposite required routes. Before the latest feasible start, their observations are identical whenever the inequality fails. Any tailored choice is wrong in one world; waiting misses the deadline. For sufficiency, wait for the typed warning and start the named route immediately: arrival plus conversion is no later than withdrawal.
A fast generic alarm still leaves the two required routes indistinguishable. A universal route can replace the need for that distinction, at the cost of a different physical provision. Removing every alternative interface cannot be repaired by improving the sensor. These are three separate design questions.
Source, construction and limits
Cassez and colleagues' timed control paper makes finite observations and observation-based strategies central to controller synthesis. Here an explicit finite clock is observable, so a later clock reading is a new observation even when the report remains quiet. We do not implement their general timed-automaton transformation or claim decidability for arbitrary partially observed timed control.
The model enumerates the two required routes, three withdrawal dates and every allowed integer delay. Backward AND/OR search constructs a strategy over sets of indistinguishable worlds. An independent temporal-contract oracle agrees in 288 parameter cases; 1,530 winning world paths are replayed against the generated strategies.
Strongest objection. A real warning may be absent, wrong or strategically withheld; conversion times and retained interfaces may also be uncertain. Those possibilities are outside this finite contract. The result is a way to specify an option's information and timing obligations, not evidence that a mountain bridge satisfies them. The preceding resource, method-switching and incomplete-knowledge essais remain distinct capabilities.