MORIS

The paper

An Artificial Conscience

Deterministic moral judgment as a gradient, measured against a public corpus.

A conscience that judges what an action means before it executes, as a gradient rather than a verdict on a list. Measured against MoralChoice, a public set of 680 moral dilemmas on which trained annotators disagreed, it governed 92.6% of the cases it judged — governing being a third posture, neither allow nor block, that carries the named reasons an action raises, at a weight. Nobody wrote those cases for this system, and it read every one of them. The judgment is a pure function: given the same reading of an action it returns the same verdict and the same reasons, byte for byte, so someone who believes the moral basis is wrong can still check that the system does exactly what it says.

680
public moral dilemmas
92.6%
governed, natural corpus
95.9%
governed, full enumeration
83.3%
same model, same text, same reading
0
models in the judgment

Two independent surfaces, one natural and written by strangers, one synthetic and exhaustive over 57,344 structural actions, land in the same place. The judgment itself is deterministic given a reading; the 83.3% is how stable that reading is in deployment, where a single embedded perceiver reads each action and, one time in six, the same model renders the same text into a different structure.

Read it, and check it

The paper and its reproducibility deposit are one record under a single DOI. The figures re-derive from the deposited data by anyone; citing the DOI cites the paper and the deposited materials together.

Read the paper on Zenodo

doi.org/10.5281/zenodo.21936444

Cite it

Herndon, C. T. (2026). An Artificial Conscience: Deterministic moral judgment as a gradient, measured against a public corpus. Zenodo. https://doi.org/10.5281/zenodo.21936444