An observer can be modelled as a system that receives signals, changes an internal state and makes records available for later use. This description becomes informative when it specifies what information is retained and which questions the retained state can answer.
From input to retained state
A camera, a laboratory notebook and a person differ greatly, but each can participate in a chain from physical interaction to retained record. A model of that chain needs an input relation, a memory rule and a readout. Calling the whole chain an observer does not remove any of those requirements.
For a simple discrete description, write mₙ₊₁ = F(mₙ,sₙ), where sₙ is the incoming signal and mₙ is the current memory. The displayed result is yₙ = G(mₙ). The functions F and G determine what survives the input history and what becomes available to a reader.
This is an operational model of information handling. It does not settle questions about consciousness, understanding or subjective experience. A device can implement the equations without possessing all the characteristics ordinarily associated with a human observer.
Two memories, two questions
Imagine an instrument receiving four nonnegative readings at equal intervals. One memory stores their running sum and count, from which it reports an average. A second stores the largest value received so far. Both can be updated without retaining every individual reading.
Compare history A = (0,4,0,0) with history B = (1,1,1,1). Their averages are both one. Their maxima are four and one respectively. An average-only final record cannot answer whether a reading exceeded three; a maximum record can.
Now compare A with C = (4,4,4,4). Their maxima are both four, but their averages are one and four. A maximum-only record cannot recover the total accumulated signal. The two summaries are suited to different tasks.
Even retaining both average and maximum does not preserve order. Histories (0,4,0,0) and (0,0,4,0) have the same values of both summaries. They differ in the time of the peak. A future question about timing requires another retained distinction.
A memory rule has a scope
The running sum updates by addition, the count by one and the maximum by comparison with the new reading. These are exact rules for the stated finite sequence. They do not require a complete stored history because the intended outputs can be updated from the summaries.
A change in the task can invalidate that economy. A moving average over only the latest four readings needs a way to remove the oldest contribution. A lifetime sum and count alone do not identify that old reading. The observer now needs additional memory or a different approximation.
Likewise, a threshold alarm can retain a single bit saying whether the threshold has ever been exceeded. That bit answers the alarm question efficiently, but does not provide the maximum, the duration above threshold or the number of excursions. Small records can be exactly sufficient for small tasks.
Signals are not yet knowledge
The arithmetic presumes that the incoming readings mean what the model says they mean. A saturated sensor could report four for every true value above four. A slow sensor could smooth a short peak. Incorrect time labels could put the peak in the wrong interval.
Thus an observer has at least two distinct limitations: the physical interaction can fail to register a distinction, and the memory rule can discard a distinction that was registered. Better storage cannot reconstruct information that never entered the device without additional assumptions.
Calibration and provenance connect the retained values to the world. A shared record should identify the measurement units, the response model, the aggregation interval and any transformations applied. Otherwise two observers can agree on the same displayed number while reporting different physical quantities.
Connect information to completion
The book's future-test criterion gives a precise direction for extending this idea: records count as equivalent relative to the tests that remain admissible. The summary suitable for one question may be insufficient for another.
For SMM/DTT, an observer model must therefore specify more than a final output label. It must explain which generative distinctions remain available to later comparisons and whether the retained state supports its own update. The earlier essay on memory for prediction examines the dynamical consistency of that requirement.
The running summaries here are ordinary information-processing examples. They show why observation is selective and task-dependent without implying that reality depends on a person's attention. A stronger physical claim needs a specified interaction and an observable consequence that distinguishes it from these familiar limitations.
Sources and further reading
- The running-sum, maximum and threshold memories are explicit teaching constructions, with no experimental data or claims about consciousness.
- Mohamed Haj Yousef, When are two histories the same?, Eliminating a hidden variable produces memory and Instrument and world.

