SoftPrompt-IR is an experimental annotation concept for prompt intent and weighting.
Contributions are welcome that focus on:
- empirical comparisons (before / after)
- small-scale evaluations or ablations
- controlled examples and counterexamples
- failure modes and edge cases
Please keep changes minimal and well-scoped. Prefer observation over prescription.
This project does not aim to define execution semantics or guarantees. If you propose an extension, explain what ambiguity it reduces.
Open an issue before submitting larger changes.