Invariant Discovery with Honest Refusal (v3.2)

Q: What does this research mean for the field?

The CV→0 protocol for invariant discovery with honest refusal achieves a zero false positive rate in null calibration, enables significant cross-domain transfer, and successfully refutes two previously accepted universal claims in battery degradation. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.CHALLENGES_CONSENSUS.

Evgeniy Dolgov

doi:10.5281/zenodo.22286637

September 3, 2026Open Access

Invariant Discovery with Honest Refusal (v3.2)

View Full Paper

Ask AI

Bookmark

Share

Key Points

To establish and evaluate an invariant discovery protocol that uses mechanical refusal mechanisms to prevent false mathematical discoveries.
Implemented a CV→0 protocol incorporating held-out validation, null calibration testing, compression tests, and a mechanical refusal taxonomy.
Evaluated performance against PySR across a 12-task benchmark and tested cross-domain language transfer.
Applied the framework to a battery case study examining 273 cells to test universal scaling claims via an α-hierarchy.
Achieved a false positive rate (FPR) of 0/100 under null calibration testing.
Recorded a 6/3/0 performance profile against PySR on a 12-task benchmark alongside a 3.8-fold increase in cross-domain transfer.
Refuted two universal claims within the 273-cell battery dataset using the α-hierarchy framework.

Implication

Computational benchmark demonstrates calibrated invariant discovery with zero false positives across physical datasets, suggesting improved reliability in scientific machine learning.

Abstract

v3.2: adds Ethics statement, 7 new references (SINDy, DSR, SRBench, Severson 2019, Attia 2020, Co-Scientist 2026), citations embedded in Related Work and §6. CV→0 protocol: held-out validation, null calibration (FPR 0/100), compression tests, mechanical refusal taxonomy. 12-task benchmark vs PySR (6/3/0). Language ontogeny: word born from data, cross-domain transfer ×3.8. Battery case study: 273 cells, α-hierarchy, 2 universal claims refuted.

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

Evgeniy Dolgov (2026) studied this question.

synapsesocial.com/papers/6a9bd4536b95aff0620ebe4e https://doi.org/10.5281/zenodo.22286637

View Full Paper

Ask AI

Bookmark

Share

Invariant Discovery with Honest Refusal (v3.2)

Key Points

Implication

Abstract

Discussion

Cite This Study

Invariant Discovery with Honest Refusal (v3.2)

Key Points

Implication

Abstract

Discussion

Cite This Study

Keep reading

More related papers

Keep reading

More related papers