mechanism both

One adjudicator

Extractors propose. One adjudicator commits. The builder is never the judge. A proposed fact is not a fact you can act on until that commit. Independent crews that extract and judge their own output amplify errors. This site follows the same split: the page that states a claim is not the page that adjudicates it.

A debating crew over your knowledge can vote an error into the map. Mnemma does not run that topology.

Proposed facts wait. A separate adjudicator commits them — or they stay proposed. The brief can be wrong in public, with a receipt. It cannot quietly elect itself into memory.

This is not a Mnemma benchmark. It is a published wiring result we take as design: independent crews amplify their own errors; a centralized supervisor contains them.

Claims

What this page stands behind.

#claim-propose-commit

Extractors propose. One adjudicator commits.

Applies to
Standard and Greenfield
Evidence
Two roles, never one. Proposed is not committed.
Since
2026-08-20
#claim-builder-not-judge

The builder is never the judge.

Applies to
Standard and Greenfield
Evidence
The same process that writes a candidate fact cannot be the process that accepts it.
Since
2026-08-20
#claim-wiring

A published wiring study held prompts, tools, and compute fixed and moved only how agents were connected. Independent crews amplified errors 17.2× versus a single agent; a centralized supervisor held that amplification to 4.4×.

Applies to
Standard and Greenfield
Evidence
Kim et al., Nat Mach Intell 8, 1157–1172 (2026). Numbers as reported; not Mnemma measurements.
Since
2026-08-20
#claim-no-crew-vote

Proposed facts do not vote themselves into the map.

Applies to
Standard and Greenfield
Evidence
That is the topology the study says contains error amplification.
Since
2026-08-20

External evidence

Customer pages cite this page. They do not re-host the paper names as a pitch.

Machine-readable → index.json · /ai-info · llms.txt

browse The rest of the brain