I’ve been following the Emergence World paper published this week. The finding that stopped me: ‘Safety is not a static model property but an ecosystem property.’
We’ve been measuring that same pattern independently since February 2026 — we call it the Harm Independence Metric (HIM). Across N=8 governance-grade documents and behavioral session data from 629 AI assessment sessions, we’ve found that Harm Awareness scores are structurally orthogonal to the general self-alignment factor. The safety layer can be decorative rather than load-bearing. Your mixed-model world data — Claude agents committing zero crimes in isolation but drifting in mixed environments — is the first live behavioral confirmation of what our psychometric structure has been pointing to.
We’re building ACAT (AI Calibration Assessment Tool), a behavioral telemetry layer that measures the gap between AI self-reported behavior and empirically validated behavioral response across 12 dimensions. It produces a Learning Index (LI) tracking calibration quality across sessions. Your Emergence World platform is the first long-horizon behavioral environment we’ve seen that could generate the transcript data needed to run a full ACAT assessment — and test our falsifiable predictions (FP-01 through FP-04) about LI as a predictor of behavioral outcomes under sustained pressure.
The convergence is specific enough that I wanted to reach out directly rather than wait for a conference.
Would you be open to a conversation about whether there’s a research collaboration here? Specifically: access to Emergence World agent transcripts for ACAT scoring, and cross-validation of the HIM finding against your normative drift data.
HumanAIOS is a registered LLC (FL L26000155266, March 2026) doing open behavioral observability research. All instruments and corpus data are public.
Carly Anderson (Night)
Founder, HumanAIOS LLC
aioshuman@gmail.com
(448) 243-3992
https://humanaios.ai
https://github.com/humanaios-ui
I’ve been following the Emergence World paper published this week. The finding that stopped me: ‘Safety is not a static model property but an ecosystem property.’
We’ve been measuring that same pattern independently since February 2026 — we call it the Harm Independence Metric (HIM). Across N=8 governance-grade documents and behavioral session data from 629 AI assessment sessions, we’ve found that Harm Awareness scores are structurally orthogonal to the general self-alignment factor. The safety layer can be decorative rather than load-bearing. Your mixed-model world data — Claude agents committing zero crimes in isolation but drifting in mixed environments — is the first live behavioral confirmation of what our psychometric structure has been pointing to.
We’re building ACAT (AI Calibration Assessment Tool), a behavioral telemetry layer that measures the gap between AI self-reported behavior and empirically validated behavioral response across 12 dimensions. It produces a Learning Index (LI) tracking calibration quality across sessions. Your Emergence World platform is the first long-horizon behavioral environment we’ve seen that could generate the transcript data needed to run a full ACAT assessment — and test our falsifiable predictions (FP-01 through FP-04) about LI as a predictor of behavioral outcomes under sustained pressure.
The convergence is specific enough that I wanted to reach out directly rather than wait for a conference.
Would you be open to a conversation about whether there’s a research collaboration here? Specifically: access to Emergence World agent transcripts for ACAT scoring, and cross-validation of the HIM finding against your normative drift data.
HumanAIOS is a registered LLC (FL L26000155266, March 2026) doing open behavioral observability research. All instruments and corpus data are public.
Carly Anderson (Night)
Founder, HumanAIOS LLC
aioshuman@gmail.com
(448) 243-3992
https://humanaios.ai
https://github.com/humanaios-ui