Shared waters · Essai 05 · a finite organisational rehearsal
Keep service running while capability changes.
A plan can leave a route open and still permit endless delay. Here we separate a possible route, a policy that never strands it, and a policy under which every maximal sequence of events finishes.
The offshore depot wants a newly trained capability adopted after independent approval. Its existing promise is modest and explicit: at least one service team remains available throughout the change. Training normally removes the incumbent team from service. Temporary cover, or a separately supplied shadow-training method, can bridge that conflict. Neither supplies approval.
Start with a route that can wait forever.
The default reviewer may approve or ask another question, indefinitely. You can arrange cover and finish training, but you cannot disable the reviewer’s questions. Compare the three claims below.
Some sequence reaches the target without losing service capacity. This existential claim can select favourable external decisions.
Every permitted prefix still has some safe continuation to the target. Uncontrollable events remain permitted. An infinite non-completing sequence can remain.
A finite decreasing rank certifies progress under every admitted event choice. This is stronger than nonblocking; it is not a time bound.
Three modules, then one controller.
People and competence
Ready → training → capable → operational
Training completion is not controlled by the supervisor. Applying for review requires capable people. Adoption requires their capability and the review’s grant.Temporary service cover
Off ⇄ on
One commissioned cover team supplies the service capacity while the incumbent trains. Releasing it during ordinary training breaks the promise.Independent review
Unasked → pending → granted / refused
“Apply” synchronises with people being capable; “adopt” synchronises with a grant. Questions and decisions remain outside managerial control.Composition: a shared event must be allowed by every module that names it. Other modules retain their state. Thus “adopt” joins capable people with a granted review; it cannot manufacture either. The service promise observes the resulting joint state: ordinary training needs cover. The supervisor then disables selected controllable joint events. It cannot erase a physically admitted completion, question or decision.
Inspect the local relations and the actual product
| Joint state | Service promise | Safe / nonblocking | Progress rank |
|---|
Try the policy as a sequence.
Settings above define a new fictional model and clear this rehearsal. Below, choose which permitted event occurs. This illustrates the supervisor’s admitted possibilities; clicking a review outcome does not give the manager that authority.
This page’s trace
Ephemeral model exploration: no real work, permission, saved record or file. Reset or reload clears the trace and restores defaults. There is no submission, background process or durable receipt to find.
What is the smallest supplied enabling change?
These are four combinations of just two authored options. The costs are illustrative units: cover 1, shadow training 2. They are neither prices nor estimates of effort. The search derives a choice within this library; it does not discover the library or validate either method.
| Available intervention | Cost | Safe route | Nonblocking | All maximal sequences |
|---|
In the bounded-positive-review model, temporary cover is the cheapest supplied way to make progress possible. In the endless-question model, both internal options still fail to guarantee completion. With possible final refusal, no nonblocking supervisor can promise adoption: allowing the request exposes a terminal refusal, while avoiding it prevents adoption. A different promise or an independently justified review assumption is needed; a controller cannot procure a favourable decision by deleting it from the model.
Who chooses the outcome of training?
Suppose management can authorise “begin” and training can then either succeed or stall permanently. A successful path exists. Enabling “begin” admits the stalled outcome too. The supervisor therefore cannot guarantee completion or keep a route to completion from every resulting state. If the second outcome returns to the start instead, another attempt stays possible, although the retries could continue forever.
The solver enables or disables each event as a whole. Grant, refusal and external questions remain outside management control. The twelve rehearsal settings above have a single successor for each state and event.
What this essai adds.
Essai 04 checked whether local service steps joined and whether a request could manifest a standing capability. This essai represents work that changes capability, while exposing a continuing service constraint and another party’s choices. It derives the temporary support sequence and, more significantly, tells us when that sequence is insufficient.
The useful boundary is now a controlled interface: which events may management prevent, which events the environment may take, and which service condition must survive their composition. A plain product-state table computes exactly the same answer. The module presentation earns its place by making the source of each constraint and decision explicit; no additional category-theoretic computational advantage is claimed.
The two fixed points and the progress certificate
Let S be safe joint states and G the target states. First repeatedly remove states with an uncontrollable successor outside S. This greatest safe region permits disabling all controlled exits; it allows deadlock. For nonblocking, repeatedly retain only states with a path to G inside the current region and with every uncontrollable successor still inside it. Enable every controlled edge that remains inside. This greatest region guarantees an available completion after every prefix, not inevitable completion.
For progress, start W₀ = G. Add a safe state at stage k+1 when every uncontrollable successor is already in Wₖ, and either an uncontrollable successor exists or a controllable edge enters Wₖ. Retain all mandatory edges and only controlled edges to lower ranks. Every admitted step strictly lowers a nonnegative integer, so every maximal event execution reaches G. Outside the final W, either an uncontrollable edge stays outside W, or there is no controlled escape into W. That complement is a checkable obstruction to any such supervisor.
The finite model uses full observation. Goals terminate the episode; their capacity promise remains true, but later work is outside scope. Infinite event histories count as executions; a finite prefix is maximal only when no event is enabled. If the real environment can stop producing events while an event is enabled, this event-sequence theorem gives no deadline or eventual real-world response. No fairness assumption is silently added.
Source → concept → construction → limit.
| Source | What carries into this construction | Boundary |
|---|---|---|
| Willems (2007), printed pp.70, 80–82, 94–95 | Behaviour as allowed histories; control restricts histories through an exposed interface. State retains memory that changes continuations. | The bounded reviewer retains its remaining questions. We do not apply the linear differential-system implementability theorem to organisations. |
| Libkind (2020), §2.2; Baez, part 8 | Compatible local transitions; states and tasks for a team. | A finite synchronised product is written explicitly. No implementation of the full resource-sharing-machine algebra or network-model theorem is claimed. |
| Ramadge & Wonham (1987), §§2, 5–7 | Supervision disables controllable events; uncontrollable behaviour stays possible. Nonblocking means each permitted prefix can complete. | The all-execution requirement is our stronger finite rank construction, proved above. A permissive event model supplies no wall-clock guarantee. |
| Baez, part 5; part 9 | Omitted constraints can change composition; reusable participants do not supply free parallel capacity. | Training capacity, cover, costs and review alternatives are stipulated here. No source establishes their organisational effectiveness. |
All ten mandatory records were freshly reread for this attempt, including Willems’s 54 pages and the six Baez items. Mathematical illustrations were inspected where extraction omitted notation. The saved part 4 page-count discrepancy remains: 15 accessible pages, no separately recovered sixteenth page; the substantive author webpage was cross-checked. Ramadge–Wonham §§2,5–7 were read in addition. Detailed coverage and independent checks accompany the private native result.