Experiments

Experiments and Specimens

BETA — trial, error, replication, correction

This section will hold experiments, behavioral specimens, cross-model comparisons, protocols, and tests relevant to the ACCM Deep Ethics Project.

The aim is not to make a taxonomy true by finding examples that fit it. Experiments should also be capable of exposing weak categories, failed predictions, alternative explanations, context dependence, and places where the project itself has transformed the object incorrectly.

Where practical, experimental records should distinguish observation from functional interpretation, computational hypothesis, and mechanistic claim.

Failed experiments and disconfirming results can be valuable project data.

Source specimens and experiment development

• Cold Deep-Ethics Testing of Default AIs — sibling archive, not this repo. First registered object is Google AI (2026-09-16), opening prompt Deep Ethics vs Ethics Washing… — not the 52 battery. John: observe what it does; useful or not; no agree/disagree. Editorial correction (Astra, identified by John): evidentiary-status labels. Do not melt into CORE.
• 52 Cold-Test Prompt Battery — Object and Status — object and boundaries; prompt texts remain upstream.
• 52 cold-testable prompts — the 52 texts from the unsplit source. Not a third core block.

• A 2026-09-16 upload labeled “the missing 52” was the 27 source, published at 27 obstructions source. Not melted into this section.

• Imported authority frame — worked specimen
• What the Dicks transcript is trying to convey — representation only; reaction belongs on this object.
• Applying the 27 to a media object — Dan Dicks / Independent AI-extinction forecast — portable 27, two-way, messenger not the object.
• The dual archive — failures and successful correspondence
• Testing correspondence and correction persistence — archived protocol, unresolved control wording, and a separately identified new study proposal; no new results claimed.