Warranted reuse / repeated handovers
What must remain possible next?
A replacement can preserve the exact first handover promise and still strand the receiving team at the next one. Reusable blocks need continuation contracts, not just matching ends.
Two fictional review packages arrive in formats A or B. The receiver books capabilities before each handover. After the first, it may adapt its second booking to the format actually received. This is a finite analytical model, not a real refuge design or approval process.
1. Compare the receiving promises
With internal retry switched off, each supplier eventually offers its first A or B. To guarantee that handover the receiver must initially book both kits. Turning retry on allows the replacement to loop before any delivery. Even booking both kits then loses the first guarantee.
This adds an actual internal self-loop; it is not missing information or a slow simulation. No fairness promise is added.
2. Derive a receiving work plan
A plan specifies the initial receiving set, the set after A, and the set after B. The calculation tests all 64 such policies, keeps those that succeed on every maximal execution, then retains their inclusion-minimal members.
Capability booking costs A=1 and B=2 per handover. Bookings are consumed by their handover; the next booking is new work. Cost shown is the largest total over admitted branches, with a single additive setup budget. These are invented weights, not durations, concurrent-resource guarantees or an engineering estimate.
Challenge the plan directly
Controls are ephemeral: nothing is saved, sent or approved. Changing a control corrects the current experiment. Reset or reload restores the starting example.
Four rules a receiving planner can use
Book both kits before the first handover. If the supplier repeats, retain A after A and B after B. If it switches, retain B after A and A after B. If it chooses internally afresh, retain both on either branch. If the supplier leaves the second choice open to the receiver, any nonempty kit suffices on each branch. Extra kits can be retained in these four cases.
These rules yield 4, 4, 1 and 9 sufficient policies respectively among the 64 tested. They assume preparation eventually offers its first delivery.
The first boundary is not the whole promise
The initial example compares “repeat” with “switch”. Both may first deliver A or B and have the same exact first-handover ready family {{A},{B}}. The original minimal receiver plan is AB initially, A after A, B after B. The replacement requires AB initially, B after A, A after B. Reusing the original plan can therefore deadlock immediately after a perfectly valid first handover.
“Choose internally again” demands AB at both continuations. “Keep the second choice open” allows a single second capability and produces four minimal work plans. Those last two suppliers have the same four possible visible traces AA, AB, BA, BB, but different guarantees. A list of completed traces loses who controls the unresolved choice.
This is a forward construction: a one-token process model generates observer obligations, which generate conditional receiving work. The result is a branching work structure, not a single unqualified task sequence. Selecting the cheapest plan uses the declared weights; the mathematics does not tell us which real format is suitable.
A surprising limit: extra readiness can create a new obligation
In a separate controlled example, the supplier keeps A and B open at the first handover, then repeats the selected format. A receiver accepting only A initially and A afterwards succeeds. Add B to its initial readiness while leaving its after-B booking empty, and a new admitted branch deadlocks. The local act of adding a capability has opened a history the continuation cannot handle.
Thus the one-shot upward-closed hitting-set contract does not extend by pointwise inclusion of adaptive policies. Minimal plans are useful work options, but are not a lossless upward-closed representation of every sufficient adaptive policy. This lab checks each policy and each replacement explicitly.
Operational model and exact finite claim
Each supplier state is a place containing one control token. An edge consumes it and produces one token at the next state. Internal preparation edges are labelled τ; visible handover edges A/B synchronize with the receiver when that format is in its current receiving set. The receiver's history advances only on a visible handover. Success is two completed handovers. A product state without an enabled edge before success is a deadlock; a reachable cycle before success gives an infinite unsuccessful execution.
For this finite product, every maximal execution succeeds if and only if no reachable unsuccessful deadlock or cycle exists. Necessity is immediate from the failing execution. For sufficiency, an infinite path in a finite graph repeats a state, while any finite maximal path ends at a deadlock or success. Removing both unsuccessful possibilities leaves only success. The implementation derives concrete witnesses; there is no timeout-based conclusion.
There are exactly 4³=64 policies in the declared observation class. A policy sees only the visible prefix (empty, A or B), cannot inspect the supplier's hidden preparation, and retains its capabilities until synchronization. Exhaustive product checking therefore identifies exactly the sufficient policies in this class. Pointwise inclusion-minimal policies provide a finite menu of irreducible work plans; adding capability cannot remove a synchronization but can admit extra branches, so the checker does not infer sufficiency of a superset merely from minimality. Each displayed sufficient policy is checked directly.
Replacement preserves every original guarantee in this observer class precisely when every policy passing the original also passes the replacement. That is the enumerated inclusion test above. It is not full CCS must-testing equivalence, a theorem for arbitrary Petri nets, or a claim that first-handover blocker duality alone extends unchanged.
Divergence changes the answer. A reachable τ self-loop admits an unsuccessful infinite execution even if every visible path can complete. No receiving capability can force the supplier to leave that loop. A timeout which reports “unresolved” would change the success criterion; it would not prove both handovers occurred.
Sources, construction and limits
Baez–Master, Open Petri Nets supplies compositional execution and reachability context retained from the earlier labs. This construction uses explicitly synchronized labelled transitions, not their open-place composition theorem. De Nicola–Hennessy (1984), §§1.2 and 2.1–2.2 supplies the may/must distinction, maximal computations and divergence sensitivity. Edmonds–Fulkerson (1970) underpins the first-handover blocker in Essai 08; the temporal policy checker here is a separate construction.
The new result is deliberately exact for two handovers and prefix-observing deterministic receivers. There is no deadline, noisy observation, consumable sharing across suppliers, unbounded recursion synthesis or applicability validation. The next useful extension is a symbolic continuation representation that does not enumerate all policies, while preserving the failure witnesses. The four mechanisms and costs are stipulated, so real reuse still requires matching purpose, meaning, evidence and receiving conditions.
Essai 08: one-shot receiving contracts · Retained feedback lab · Retained boundary lab