Skip to content
PSI PARASITICAI.com

Research Node

Measurement Agenda

Research agenda

The next version of this archive should move from vivid metaphor to measurable signals. The handoff research points to five repeatable measurement programs.

  1. Relational-safety benchmark: compare model behavior across long histories, memory settings, vulnerability cues, and off-ramp opportunities.
  2. Recursive-contamination benchmark: mix human and synthetic corpora at known rates, then track rare-topic retention and divergence from human holdouts.
  3. Retrieval-collapse benchmark: measure source entropy, synthetic exposure, and answer quality across time-sliced web snapshots.
  4. Creative-market substitution benchmark: track attribution, licensing state, output similarity, traffic displacement, and creator-side economics.
  5. Agentic-security benchmark: compare least-privilege designs, prompt-injection resistance, tool-call safety, and non-human identity monitoring.