LLM Oracle Reference

The LLM Oracle module models LLM output completion as weighted unification. Given a schema (field → type string) and a partial attrset (known fields), it enumerates completions ranked by fewest invented fields.

Weight = number of fields absent from partial that had to be invented (assigned a fresh variable). Lower weight = closer to what was already known.

---

rankedComplete

rankedComplete :: schema -> partial -> int -> [term]

Returns up to n completions of partial satisfying schema, sorted by ascending weight (fewest invented fields first).

Arguments:

• schema — { fieldName = typeString; ... } — type string is a builtins.typeOf value: "int", "string", "bool", "list", "set", "float", "null"
• partial — attrset with known field values (may include variables)
• n — maximum completions to return

let u = import ./. {};

schema = { name = "string"; age = "int"; admin = "bool"; };
partial = { name = "alice"; };
in
u.rankedComplete schema partial 3
# => [
#      { name = "alice"; }   -- weight 2: age and admin invented
#    ]

Fields in partial that conflict with schema types cause immediate failure (no completions returned).

---

rankedCompleteFull

rankedCompleteFull :: schema -> partial -> int -> [{ answer: term; weight: int }]

Like rankedComplete but returns records with both the completed answer and its weight.

let u = import ./. {};

schema  = { x = "int"; y = "int"; };
partial = { x = 1; };
in
u.rankedCompleteFull schema partial 5
# => [ { answer = { x = 1; }; weight = 1; } ]

The answer contains only the fields present in partial (ground fields). Invented fields remain as logic variables in the reified result.

---

llmComplete

llmComplete :: schema -> partial -> var -> goal

The underlying relational goal. q is unified with the completed attrset.

• Fields present in partial: included as-is, weight 0.
• Fields absent from partial: assigned a fresh variable with weight 1.
• Ground partial values violating schema types: fail immediately.

let u = import ./. {};

schema  = { status = "string"; code = "int"; };
partial = { code = 200; };
in
u.run 3 (q: u.llmComplete schema partial q)
# => [ { code = 200; } ]

Use llmComplete directly when composing with other goals (e.g., adding additional eq constraints on the result).

---

Notes

• rankedComplete / rankedCompleteFull delegate to wrun / wrunFull, so all semiring behaviour applies (deduplication, stable sort).
• The oracle is purely structural — it does not call any external LLM. The name reflects the intended use: score partial outputs from an LLM by how much was already known vs. invented.