Skip to content

Trace→Generator: Inverse Prompting Toy Scaffold

Status: Scaffold Scope: Minimal experiment design for searching prompt/control configurations from desired output constraints. Epistemic status: Engineering prototype template; no claim of full inverse recovery. Related files: - theory/emergence/trace-to-generator.md - theory/ai/llms-as-probabilistic-automata.md - theory/core/simulation-theory-map.md Failure conditions: - Treating prompt search success as proof of unique generator recovery. - Ignoring runtime and evaluator bias.

Goal

Given target output constraints, search for prompt candidates that generate outputs scoring well under an explicit evaluator.

Components

  1. Target constraints (style, facts, structure)
  2. Candidate prompts
  3. Generator interface (stub)
  4. Evaluator
  5. Mutation/search loop

Limits

  • No API credentials required; generation backend is mocked by default.
  • Reconstruction is non-unique.
  • Evaluator choices strongly shape recovered candidates.

Run

python lab/experiments/trace_to_generator/prompt_search_toy.py