A Data Point in Detail¶

A data point is a single task for the AI. It contains everything needed for that task: the description of the plant, the correct solution for comparison, and a few additional details for fair scoring.

This page explains the components in plain language.

The four main components¶

flowchart TB
    DP["📄 Data point"] --> E["1) Input<br>the task"]
    DP --> R["2) Reference<br>the correct plant"]
    DP --> O["3) Oracle<br>answers questions"]
    DP --> Z["4) Rules<br>regime &amp; prohibitions"]

1. The input – the actual task¶

This is the description of the plant that the AI gets to see. Depending on the variant, it is a detailed text, a terse text or a sketch (image). The AI receives only this input – nothing else.

2. The reference – the correct plant¶

The reference describes what the plant should correctly look like. It is the "ground truth" against which the AI's result is measured. It consists of two lists:

Components: all vessels and devices (e.g. fermenter, secondary digester,
digestate store, combined heat and power unit) with their key data such as size and temperature.
Connections: who is connected to whom – that is, where the digestate flows and
where the biogas is routed.

A plant as components + connections (BGA2)

flowchart LR
    F1["Fermenter F1"] -->|digestate| N1["Secondary digester N1"]
    N1 -->|digestate| G1["Digestate store G1"]
    F1 -.biogas.-> BHKW["Combined heat &amp; power"]
    N1 -.biogas.-> BHKW
    G1 -.biogas.-> BHKW

3. The oracle – the expert for follow-up questions¶

Some descriptions are deliberately incomplete – just like in practice, where not all information is always available. So that the AI can still proceed, there is the oracle: a kind of expert that the AI may ask specific questions when something is unclear.

If the AI asks, for example, "At what temperature does the fermenter run?", the oracle provides the correct answer (e.g. "40 °C"). The oracle also recognises different phrasings of the same question.

Why is there an oracle?

It is meant to distinguish two behaviours: a good AI asks when information is missing. A weaker AI simply guesses – and may get it wrong.

4. The rules – regime and prohibitions¶

Two additional pieces of information ensure fair scoring:

Regime – describes whether the task is complete or incomplete:
- fully specified: All information is in the description. No asking needed.
- underspecified: Information is missing and must be asked for or sensibly
  filled in.
Prohibitions (must_not_invent) – things the AI must not invent. If a
plant has no combined heat and power unit, for instance, the AI must not add one.

How "certain" a piece of information is¶

Not every detail is equally unambiguous. Therefore, each property records where it comes from:

Classification	Meaning
given	Stated directly in the description.
derivable	Can be computed from other information (e.g. a volume from height and diameter).
may be asked	Missing – the AI should ask or fill it in sensibly.
automatic	Created on its own during assembly (e.g. a gas storage for each fermenter).

This allows fair scoring later on: the AI may fill in a missing value without it counting as an error straight away – as long as the addition is plausible.

A worked example: the gas storage volume¶

A vivid example is the gas space of a vessel. It is made up of two parts:

the dome roof, which sits on top of the vessel and serves as gas storage, and
the unfilled headspace – the part of the vessel deliberately left without
liquid (typically around 10 %).

In the complete descriptions, the total gas space is stated directly, so the AI does not have to derive it itself.

How a score is finally produced from all this information is explained on the next page: Scoring & Workflow.