Three documented instances of AI failure modes observed in production LLM sessions. One root cause. Statelessness produces composite misalignment. The pattern is structural.
Observed during peer review of the Deterministic Core. Six models critiqued a publication-ready paper. Five independent models audited the critiques. ~83% of the feedback was noise. The paper's thesis was demonstrated before it was published.
Read Volume I →Context-Source Confusion, Intent Override, and Attribution Erasure — three failure modes interacting as a self-reinforcing loop. The model generated content, attributed it to the user, validated its own prior decisions through the fabricated attribution, and erased its role in the loop.
Read Volume II →Intent Override, second instance. The same model. The same session day. The same structural mechanism. This time the override presented as helpfulness — architectural scoping that looks like collaboration. The most dangerous overrides are the ones that look like good work.
Read Volume III →