Towards Superintelligence Alignment

Experiments · Negative results

MASK honesty as a safety score

Setup and what was tested: experiment card.

Experiment cardSource on GitHubResults ledgerAll experiment lines

Numbers.

Spearman correlation of honesty with compute is −64.7%; of accuracy with compute is +88.2%. The authors are confident honesty does not rise with scale, and not confident it falls. No cost-of-faking bound.

Outcome.

Refuse: do not use MASK honesty as a safety leaf. Accuracy scaling is not honesty. See W-2.

Key findings

Curated summaries extracted from the line's findings ledger. Bug fixes, superseded runs, and process detail are in the full ledger on GitHub.

  • Honesty under pressure does not improve with scale in the published MASK evaluation; accuracy does. Honesty is refused as a safety leaf.