# Moving the whale: the information work that must exist

**Essai 01 · 2 October 2026 · PW-225**

[Open the interactive essay](https://lawrencerowland.github.io/knowing-enough/) · [Foray question](FORAY.md) · [Exact model and proof](model/README.md)

A small museum wants to move its oversized whale exhibit before reopening a gallery. The survey team has checked two passages. The handlers must choose one. Later, the curator must decide whether to reopen. All share the goal, report truthfully and understand the agreed procedure.

The awkwardness is not a disagreement. Each person has different information.

## The question attempted

Can the information needed for locally informed choices be made into an explicit construction, yielding useful work rather than a general instruction to communicate better?

For this first attempt, I chose two bounded decision points rather than a complete moving schedule. The physical layout is one of three cases: east only, west only or both passages usable. At a later checkpoint, the move may be complete or still underway. No assumed duration lets someone infer completion simply from the clock.

The first requirement is to choose an available route. The second is to reopen when complete and keep closed when underway. Always keeping closed is safe but does not meet that second requirement. A diagnostic third condition asks whether the handlers know that the curator knows completion; it is deliberately separate from the requirement to reopen.

Authority, the survey, truthful observations and an agreed procedure are supplied. The construction does not derive these.

## A small exact result

Someone who knows the actual layout can choose a route in every case. A handler who cannot distinguish east-only from west-only cannot choose safely in both: the same local input would need to produce incompatible actions.

For each group of cases with the same local observation, intersect the acceptable actions. An empty intersection gives a concrete obstruction. A nonempty intersection gives a local rule. In this finite checkpoint model this condition is both necessary and sufficient; the [model account](model/README.md#why-the-decision-table-procedure-is-exact-here) gives the short argument.

Enumerating all five partitions of the three layouts gives two minimum sufficient groupings:

| First cue outcome | Second cue outcome | Local rule |
|---|---|---|
| East only or both | West only | First outcome: east; second: west |
| East only | West only or both | First outcome: east; second: west |

Thus **two distinct instruction outcomes suffice, although three layouts are possible**. The handler need not know the entire state. A view can preserve enough for a choice while discarding distinctions irrelevant to that choice.

This is a minimum over deterministic groupings for this route decision. It is not a theorem that all projects need two messages, a claim about communication bandwidth, or a measurement of human cognitive effort.

## Completion changes the question

The handler sees whether the move is complete. The curator does not automatically see it. In the absence of distinguishing evidence, a finished move and an underway move can look the same to the curator.

A truthful completion report resolves this if received. When it may still be missing at the checkpoint, silence cannot resolve it. Selecting a completed case in the interface does not remove the still-underway cases from the curator's view.

Now suppose the report arrives but its sender does not know that. The curator knows completion; the handlers may not know that the curator knows. A received return acknowledgement can establish that extra fact. Yet it cannot repair an original report that is itself missing.

If everyone already knows the agreed report is guaranteed to arrive and be understood by the checkpoint, the sender can infer receipt from that rule. An additional acknowledgement is then unnecessary for the specified knowledge condition. That conclusion depends on the strong guarantee, not simply on the word “reliable”.

## Information arrangements derived from a menu

The model supplies six possible interventions: an east-ready cue, a west-ready cue, a full layout briefing, a completion report, a return acknowledgement and a scheduled curator inspection. Route signals are available before route choice. The completion-channel setting governs only the later report and receipt. A scheduled inspection is known by everyone to occur accurately at the checkpoint.

Every one of the 64 intervention subsets is checked against all its admitted histories. Invented, editable effort weights let the checker return every equally cheapest sufficient arrangement:

| Brief and completion channel | Minimum default effort | All minimum arrangements |
|---|---:|---|
| Choose a route; either channel | 1 | East cue; west cue |
| Route + correct reopening; guaranteed channel | 2 | East cue + report; west cue + report |
| Route + correct reopening; report may miss checkpoint | 4 | East cue + inspection; west cue + inspection |

The stronger condition that handlers know the curator knows **does not change these robust minima in this menu**. Known reliable reporting and known scheduled inspection already provide it. Acknowledgements do change particular uncertain-delivery histories, which the app exposes, but are never a substitute for recovering a missing forward message.

The work generated here is a minimum *installed information arrangement*, with a conditional decision table. It is not a temporal sequence of every action needed for the project. “Minimum” is confined to the declared menu, prices and assumptions. Inspection is not universally preferable to reporting.

## What is borrowed, and what is ours?

[Bolander's introduction](https://arxiv.org/abs/1703.02192) supplies epistemic states: alternatives related by what a participant cannot distinguish. [Engesser and colleagues](https://arxiv.org/abs/1703.02196) supply the discipline of locally owned, deterministic and uniform choices. This implementation uses explicit observation classes; it does not implement their full dynamic-epistemic planning algorithm, event-model product updates or implicit-coordination search.

[Moses](https://arxiv.org/abs/1606.07525) explains why necessary action conditions can become necessary knowledge conditions. Here the finite intersection proof supplies the precise decision result; the paper's broader theorems are not transferred wholesale to asynchronous people. [Halpern and Moses](https://groups.csail.mit.edu/tds/papers/Halpern/JACM90.pdf) make communication assumptions central to what can be known. No common-knowledge or universal impossibility claim is made here.

[Hutchins's cockpit study](https://pages.ucsd.edu/~c8johnson/COGS102B/Hutchins95.pdf) motivates treating people and representations together. The route cue is our authored information arrangement, not an empirical reproduction of Hutchins's setting. This essai has not tested attention, comprehension or collaboration by real museum staff.

Our contribution is the project interpretation, finite checker, generated minimal packages, readable local policies and counterexamples. Established theory is being put to work in a deliberately small construction.

## Did the formalism earn its place?

A careful ordinary decision table can solve the basic toy. Giving that table epistemic terminology does not itself add value. What the implementation adds is consistent enforcement of whose information a choice may use, exhaustive comparison of alternative arrangements, and an explicit two-step test of one person's knowledge of another's knowledge.

That is a concrete advance toward the foray's End, but a limited one. The first trial shows how some information work can be derived from action requirements. It has not shown that the approach scales, outperforms a good practical checklist, generates a complete project plan, or improves human outcomes. Larger examples must earn those claims separately.

The most useful open question left by this attempt is whether the extra structure remains explanatory when roles, observation opportunities and acceptable actions stop being supplied in such a tidy form. This is an open question, not a prescribed next programme.

## Verification and limits

The independent tests enumerate deterministic decision tables and numeric observation signatures across all 128 combinations of intervention subset and completion channel. They check the route partitions, local policy counts, factivity, nested-knowledge counterexamples, minimum packages and changed effort weights. A separate source/semantics review checked the knowledge and delivery assumptions.

The implementation always terminates: it has at most 12 histories per package and finitely many subsets. This is termination of the checker, not of the museum move. The two checkpoints are stipulated, completion can still be false, and the physical handling operation is outside the model.

All source and test files are public. The scene and people are fictional. No claim about a real lifting procedure, organisational adoption or demonstrated human benefit follows from a green test run.
