Towards Computational Provenance: Carrying Causal-State Evidence in Generated Text
A study on computational provenance tested whether generated text can carry detectable evidence of which causally relevant internal state occurred. Two controlled architectures, a modular feed-forward neural network and a transformer-based model, were trained on an arithmetic task with a mandatory pathway through two discrete intermediate states. Both architectures passed all 128 matched pairs in public and sealed protected end-to-end evaluations, with the detector recovering the signal associated with the authenticated internal state. The required causal computation reproduced across five independently trained feed-forward models and three independently trained transformers.
The paper investigates whether language model outputs can provide verifiable evidence about internal computation. Using controlled architectures trained on an arithmetic task with two discrete intermediate states, the authors show that a subtle statistical pattern in generated text can be detected to authenticate the internal state used. Both feed-forward and transformer models passed all 128 matched pairs in public and sealed evaluations, and the causal computation reproduced across multiple independently trained models.
The approach demonstrates a bounded form of computational provenance by embedding a detectable statistical pattern in generated text that is causally linked to a verified internal state. The use of sealed protected evaluations suggests robustness against overfitting to the public test set. Reproducibility across five feed-forward and three transformer models indicates the causal computation is not an artifact of a single training run.
This research could inform future methods for auditing or verifying AI model behavior, potentially relevant for compliance, safety, and trust in high-stakes applications. However, the controlled setting with discrete intermediate states is far from real-world language models, so immediate commercial impact is limited.
If extended, computational provenance could enable new verification services for AI outputs, supporting regulatory compliance and quality assurance. Currently, the business value is primarily in research IP and potential future tooling for model auditing.
Next signals to watch include whether the technique scales to larger models or more complex tasks, and whether it can be applied to open-ended generation rather than arithmetic. Further research may explore adversarial robustness of the detectable pattern and integration with existing interpretability methods.