What Can the System Know at Decision Time?

At minute 10, Botie must decide whether to inspect a pump. One sensor recorded a vibration spike at minute 8, but its network connection delayed the message until minute 12. A historical table now contains that spike. The running system could not have used it at minute 10.

This distinction also affects software. A support article may have been written yesterday but added to a search index only this afternoon. We need to design around the information the service could actually retrieve.

Move the decision time or delay a record. Both clocks must be at or before the decision.

A sensor samples the world at one time and delivers the record at another. Clocks, buffering, and network interruptions can separate them. For an offline training table, the delay is easy to overlook because all the records are now present.

A decision-time reconstruction needs the timestamp when the consuming system could actually access the record. If a later repair changes a historical value, preserve the earlier version too. The final corrected table may describe a different information state.

Document which clock each timestamp uses and how clock disagreement is handled. A precise-looking timestamp is not proof of a synchronised clock.

A field called temperature could mean Celsius, Fahrenheit, or a moving average calculated over a window. A feature contract should state units, shape, missing-value behaviour, and the aggregation window. It should also say which entity the record describes.

For text retrieval, include when a document version entered the index and what permissions the requester had. The existence of a document somewhere does not mean this request could retrieve it.

These are interface requirements. A fitted model cannot reliably repair a changed meaning that the pipeline does not identify.

Start with two yes-or-no questions for each record. Did its event happen by the cutoff? Was the record available by the cutoff? Both must hold. We use inclusive comparisons, so a record available exactly at the decision time is accepted.

In the workshop, a Threshold gate reports 1 when its first input reaches its second. Connect the cutoff to the Score input and the record timestamps to Cutoff. This tests whether the decision time has reached each timestamp. Multiplying the two binary masks implements “both”.

Event times = [8, 8, 11]
Availability times = [9, 12, 11]
Decision time = 10
Event-time mask = [1, 1, 0]
Availability mask = [1, 0, 0]
Eligible = [1, 1, 0] × [1, 0, 0] = [1, 0, 0]

The labels on the wires matter. Reversing a comparison would admit future information.

import numpy as np
arrival = np.array([4, 9, 6])
reading = np.array([20., 99., 24.])
known = arrival <= 6
print(known.tolist())
print(reading[known].tolist())

[True, False, True]
[20.0, 24.0]

A reading from last week can pass both comparisons. That does not make it a useful description of a pump now. Add a maximum permitted age when the task requires recent evidence, and say which timestamp the age uses.

Do not confuse a freshly delivered old reading with a fresh measurement. A freshness check usually needs event time. A separate transport monitor can detect long delivery delays.

The workshop deliberately checks eligibility alone. A complete pump service would also choose a freshness rule and a response when no acceptable reading remains.

A fault may happen after the prediction and be confirmed several days later. The outcome window tells us what we want to predict. Label availability tells us when we can train or evaluate using the confirmed result.

For a retrospective training simulation, include only labels that would have been known at that training date. A final dataset with every later confirmation can overstate how quickly the system could have adapted.

Keep the prediction identity so a later outcome can be joined to the correct earlier decision. An outcome must never be inserted into its own input features.

If no eligible reading exists, the pipeline needs a response. It might request a manual check or keep using a previous reading within a defined age limit. Filling a missing value with zero quietly assigns a numerical meaning to absence.

Record the branch taken so later evaluation includes the cases where the model could not run. Reporting accuracy only on successfully served cases hides coverage failures.

Our reusable Python function returns an eligibility mask. The caller can then decide how to select records or handle an all-False result without pretending the evidence existed.

Include records just before the cutoff, exactly at it, and just after it. Include a record with an early event and a late arrival. Those cases exercise the design rule more directly than a large random table.

This lesson assumes arrival cannot precede the event on our shared clock. A violation raises an error. Real systems with uncertain clocks may need an explicit tolerance or interval model instead of silently accepting impossible ordering.

A contract is useful when another component can test it. The challenge’s invalid-input behaviour is part of that contract.