How does the Oracle work?¶

Some plant 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 simulated expert that the AI may ask specific questions when something is unclear. This page explains when the oracle is active, how a question turns into an answer and why that matters for fair scoring.

For a first introduction see A Data Point in Detail; this page is about the mechanism behind it.

When is the oracle active?¶

The oracle is consulted only for under-specified tasks:

fully specified (fully_specified): all information is in the description. There
is nothing to ask – the oracle stays silent.
under-specified (underspecified): values are missing. Only here may the AI ask
follow-up questions, and only then does the oracle answer.

Basis: the oracle field

Every data point carries an oracle field – a dictionary of field name → value. It is the expert's only source of knowledge. Example (shortened):

"oracle": {
  "F1.T_ad": 313.15,
  "G1.T_ad": 293.15,
  "F1.fill_fraction": 0.9,
  "bgaa.exists": true,
  "chp.exists": false,
  "sep.source": "N1",
  "substrate_feed": "approx. 80 t/d maize silage and 20 m³/d cattle slurry"
}

The two-round workflow¶

flowchart TB
    A["📝 Input<br>(incomplete)"] --> B["🤖 AI – round 1"]
    B --> C{"Code already<br>written?"}
    C -->|"yes"| Z["⚖️ Scoring"]
    C -->|"no, asks questions"| D["🔮 Oracle<br>answers the questions"]
    D --> E["🤖 AI – round 2<br>writes the code"]
    E --> Z

Round 1: the AI receives the description. For under-specified tasks it may emit
its open questions as a small JSON block instead of guessing immediately:
```
{ "open_questions": [{ "field": "F1.T_ad" }, { "field": "sep.source" }] }
```
Oracle: if the AI asks questions (and does not yet supply code), the oracle looks
up the matching values in the oracle field and returns them as an answer.
Round 2: with the answers in hand, the AI writes the full PyADM1ODE code, which
is then scored.

If the AI already writes finished code in round 1, the oracle round is skipped. There is at most one oracle round per data point (oracle_turns in the result).

How a question becomes an answer¶

The oracle understands different phrasings of the same question. You do not have to hit the internal field F1.T_ad – "At what temperature does the fermenter run?" is enough. Matching runs in several stages, from the most precise to the most generous:

Stage	What happens	Example
1. Exact match	The question names the field directly.	`F1.T_ad` → `313.15`
2. Keywords	Synonyms are mapped to a quantity.	"temperature", "heated", "mesophilic" → all `*.T_ad`
3. Component IDs	A named component returns all of its fields.	"F1" → `F1.T_ad`, `F1.V_gas`, `F1.fill_fraction`
4. Literal fields	Field names appearing in the text are added.	"… `bgaa.capacity_m3h` …"

The keyword table is bilingual (German and English) and covers typical terms – for example "gas storage"/"Gasspeicher" for V_gas, "fill level"/"Füllgrad" for fill_fraction, "rated power"/"Nennleistung" for P_el_nom or "feedstock"/"Substrat" for substrate_feed. Terms such as "efficiency" deliberately return both related fields (eta_el and eta_th).

Question → oracle answer

The AI asks:

{ "open_questions": [{ "field": "operating temperature of the fermenters" }] }

The oracle replies:

Antworten auf deine Fragen:

- F1.T_ad: 313.15  
- F2.T_ad: 313.15  
- N1.T_ad: 313.15  
- G1.T_ad: 293.15  

Bitte schreibe nun den vollständigen Python-Code.

Helpful, but not omniscient¶

The oracle is deliberately cooperative:

No matching question recognised? Then it would rather hand out all available
information than leave the AI stuck.
Only a little asked? If the AI asked for only a small part and at most five other
fields remain open, the oracle adds them as "further relevant information". A run thus does not fail because of a single forgotten question.

Why the oracle exists¶

The oracle separates two behaviours that would otherwise be indistinguishable:

A good AI asks when information is missing – and then builds the plant correctly.
A weaker AI guesses and may invent an implausible value.

This is exactly what the scoring reflects — but indirectly: the questions themselves are not graded. A model that asks gets the right values and lands inside the acceptance band on Measures; one that guesses wrong loses points there. See Scoring & Workflow for the details.

Turning the oracle off¶

For comparison the oracle can be disabled. The AI then asks no follow-up questions and has to guess the missing values:

python benchmark/eval/solve.py --no-oracle

This is the counter-check to the regular run: it shows what asking is actually worth. Expect the Measures score to drop sharply — the acceptance bands are far too tight to guess a value that was never stated.

You can explore the dataset visually – including the oracle answers per data point – in the Viewer.