Skip to content

REALITY & TIME · EXISTENCE AND THE WORLD WE EXPERIENCE

When does a computation become an output?

Use a versioned two-component record to distinguish internal progress, consistent output and knowledge that an update has completed.

A calculation can have finished internally while its recipient still lacks a dependable result. Completion must specify who can use which record, under what conditions. A simple concurrent-reading example shows why a final-looking number is not enough.

Choose the boundary of the output

A sensor calculates a pair of values and publishes them to a reader. The old pair is (1,2); the new pair should be (3,4). If the writer replaces the first component before the second, a reader can encounter (3,2). Both components are individually valid numbers, yet the pair belongs to neither completed result.

This is an example of a torn observation: the chosen readout crosses an update boundary. It is not resolved by saying that the processor's instructions occur in a definite order. The reader also performs a sequence of actions, and the two sequences can interleave.

One possible specification is that an accepted read must return either the old complete pair or the new complete pair. Another might permit partial progress but label it explicitly. The appropriate rule depends on the application; it should be stated before correctness is assessed.

A worked version-check protocol

Use a version value initially equal to 2. In this idealized model, each individual register access is atomic, all accesses have one sequentially consistent order, and there is no version wraparound during a read. These are assumptions of the example, not guarantees supplied by an arbitrary programming language or processor.

Writer and reader operations, each in its own stated order
StepWriterReader
1Set version to 3: update in progressRead version as v₁
2Replace first component by 3Read first component
3Replace second component by 4Read second component
4Set version to 4: update completeRead version as v₂

The reader accepts only if v₁ = v₂ and that value is even. If it sees version 3, it rejects an in-progress result. If the version changes during its read, it retries. Under the stated assumptions and one update, every accepted pair is either (1,2) with version 2 or (3,4) with version 4.

There are 70 order-preserving interleavings of these two four-operation sequences. Enumerating them checks the claim directly. More complicated implementations require their own argument, including memory ordering, multiple writers, failures and version reuse. The example establishes a finite consistency property, not a universal protocol for physical observation.

Completion and knowledge of completion can differ

A sender may finish transmitting before the receiver has accepted the message. The receiver may accept it before the sender receives an acknowledgment. An acknowledgment may be delayed or lost even though the result is already available at the destination.

These are different events and different states of knowledge. A record should identify which one its status refers to. A local completion flag also does not prove that every participant in a network possesses a consistent global snapshot. Distributed snapshot methods explicitly account for local states and messages still in transit.

The distinction leads to records without a global clock. Agreement about valid dependencies does not require every event to share a perfectly synchronized physical timestamp.

A usable scientific output has provenance

For an experimental or numerical result, completion may require more than the numerical payload: an input version, calibration identifier, uncertainty estimate, units and acceptance criterion may be necessary for reuse. These fields make the result inspectable; they do not guarantee that its interpretation is correct.

A valid record can contain an incorrect model prediction. A perfectly synchronized publication can reproduce a systematic calibration error. Consistency, numerical correctness and empirical adequacy should therefore be checked separately.

Making outer completion operational

The book's discussion of when inner order becomes unobservable raises a related question: which differences between generating histories remain accessible after completion? An operational proposal must define both the completed record and the tests that can distinguish it.

The published completion-principles framework supplies mathematical structures once its hypotheses are specified. A computational consistency example can clarify those hypotheses, but it does not establish a universal physical completion event, a preferred global present or a measurement-collapse mechanism.

Sources and further reading

Continue reading

Essay · 5 min read

Observation and completed records

Test whether a total-only record remains sufficient when new operations become available, and connect this closure requirement to completion.

Interactive essay · 4 min read

Records without a global clock

Toggle messages in a three-process system, trace event dependencies, and discover why logical timestamps do not measure duration or prove causal order.

Essay · 4 min read

An event and a record of it

Follow an event into a physical memory, compare independent records with copied evidence, and distinguish classical recording from quantum cloning.

Follow a connected reading path →

SMONAD

Share this page

Share this page with its own title, summary and preview image.

More platforms