1. The shared interface carries two directions
The reading-room coordinator asks a status supplier for a clear report. The supplier can answer “uncertain” or “clear”. The coordinator either asks again or returns an answer to the outside caller. This is a fictional protocol; “clear” is an authored report, not a newly obtained survey.
Answers every status query
Queries or answers the caller
Composition hides B. The caller sees its request and any eventual C response. Hiding does not stop an internal conversation or turn it into a result.
Ephemeral experiment: controls only alter this page. Nothing is saved, sent or approved. Reload restores the initial case. A reset changes the model, not the project.
2. Make both components use the same promise
Totality says a strategy can answer every legal request at a finite turn. Winning adds a condition on every infinite play following it. The same B contract must be used on both sides of their shared interface.
Why these are local strategy checks, not labels fitted to one run
The supplier is checked against arbitrarily many legal requests from its environment. “Eventually clear” means it returns uncertain until the selected reply number, then clear forever. The coordinator is checked against all legal sequences of clear/uncertain reports, not just the chosen supplier. Waiting for clear admits exactly one kind of infinite B play: uncertain forever. A finite budget admits no infinite play after the single C request.
Thus the waiting coordinator is winning when the shared B contract excludes uncertain forever; it is not winning when the contract accepts every infinite B play. A supplier that is always uncertain has precisely the opposite requirement. Changing the contract cannot make that pair jointly winning.
3. The two repairs make different promises
Promise eventual evidence
With a supplier that eventually returns clear and a waiting coordinator, the “eventually clear” B contract is satisfied on both sides. The composition returns a ready answer. The proof depends on that supplier assumption; typing alone cannot make evidence arrive.
Promise an answer, including “unresolved”
A finite query budget prevents hidden infinite interaction without assuming eventual clear evidence. On exhaustion the composite returns an unresolved result to its caller. It has terminated, but it has not established that the reading room is ready. A new external observation would start a new episode.
The operational distinction is between a promise of a useful outcome and a promise of a visible response. Escalation preserves the latter by weakening the former; the interface must expose that difference.
A legal game arena, polarity, and the composition proof
I is the empty game. B has plays that are prefixes of q b₁ q b₂ …, where q is an Opponent request and each bᵢ is a Player reply, uncertain or clear. C has one Opponent request followed by a Player answer, ready or unresolved. These prefix-closed sets of alternating plays define the arenas.
In B→C, B's polarities reverse. The coordinator receives the C request as O, sends a B query as P, receives B's reply as O, and either sends another B query or answers on C as P. Only Player switches components, as required for linear implication. The supplier's deterministic response rule and the coordinator's deterministic rule generate prefix-closed strategies satisfying the legal-play and determinism conditions.
τ : (B, W_B) → (C, ∅)
Winning arrow: total, and for every infinite play s following it,
[s|input is finite or wins] ⇒ [s|output wins].
Why joint winning rules out hidden infinite chatter. Suppose the composite has only a finite external prefix and an infinite B dialogue u. For σ the I projection is finite, so winning requires u∈W_B. For τ that winning B input would require an infinite winning C projection. But the C projection is finite. Contradiction. This is Abramsky's composition argument specialised to these arenas, not a new theorem about arbitrary project workflows.
A numerical bound for the budget repair. Starting with K queries remaining, each completed uncertain query reduces the nonnegative counter by one. The zero case must exit on C. A clear reply exits sooner. There are at most K B queries and K B replies before the external response. No fairness assumption or simulation timeout is needed.
Source → mechanism → limit
Abramsky's 2024 slide 5 supplies the symmetric feedback picture. Haghverdi–Scott, §§4.1 and 4.3 explains the partial-function trace and Int construction retained in Essai 1. This essai moves to the explicit games and strategies of Abramsky (1997), pp.3–11 and 17–18: interaction plus hiding, infinite chattering, and winning strategies closed under composition.
The finite-state executor below detects repeated complete internal state, not repeated text or a time limit. Winning judgments are separate symbolic checks over all infinite plays admitted by these particular strategies. This is a small real game-semantic instance, not a general winning-strategy synthesiser, an incentive model, or proof that evidence or coordinated benefit will occur in practice. A bounded reply of “unresolved” returns control to the caller. Repeating the same exchange with an uncertain supplier can produce that answer indefinitely. The caller can instead commission a different observation through the earlier island reader’s external inspector.