Skip to content

Physical AI optimization

Package Value
Distribution archetype-physical-ai
Import package archetype.physical_ai

Extension API. Build pure instruction-search workflows over an injected evaluator and strategy.

PerturbationStrategy

Propose instruction variants near an incumbent instruction.

TemplatePerturbation dataclass 

TemplatePerturbation(vocabulary)

Deterministic one-token toggle over a fixed vocabulary.

This dependency-free strategy is a mechanism-check fixture. Production search can provide a semantic or model-backed implementation of :class:PerturbationStrategy without changing the optimizer.

RoundRecord dataclass 

RoundRecord(
    round, best_instruction, best_success_rate, evaluated
)

Best-so-far instruction after one optimization round.

Field Type Default
round int required
best_instruction str required
best_success_rate float required
evaluated int required

OptimizationResult dataclass 

OptimizationResult(
    best_instruction, best_success_rate, trace=()
)

Final instruction and the monotonic best-so-far trace.

Field Type Default
best_instruction str required
best_success_rate float required
trace tuple[RoundRecord, ...] ()

optimize_instruction async 

optimize_instruction(
    *,
    evaluate,
    base,
    strategy,
    rounds,
    neighbors,
    patience=2,
)

Hill-climb an instruction against an injected success-rate evaluator.