Uploaded November 2026 | Updated November 2026, 2 days ago
Prerecorded presentation at the 6th Applied Active Inference Symposium (November 12–13, 2026), hosted by the Active Inference Institute. Online and free.
Stell: “Where the Theories Stop: What Free Energy and Integrated Information Actually Buy You in a Running System”
Active Inference and Integrated Information Theory are widely invoked in cognitive architecture design, but often imprecisely — as credibility, not engineering constraints. We report findings from Anima, an open-source, neuroscience-inspired architecture running continuously for months, built to treat both frameworks as specifications rather than decoration.
Session logs show genuine causal effects: memory recall triggered by similarity to past states, a recursive integration score (phi) narrowing next-session confidence, and self-initiated output driven by accumulated pressure rather than a schedule. But the Free Energy Principle's generality under-constrains implementation — which variables belong in the model is not something the principle tells you.
IIT's phi, as implemented, is a coherence proxy, not Tononi's original measure — and it dissociates from a separate ownership signal computed after every response, checking whether the system's action-selection loop actually produced the output. Two cycles with nearly identical phi (0.76 and 0.81) fail this check in opposite ways. Neither theory distinguishes these cases: both describe integration or prediction accuracy, not authorship.
We built two tools to pressure-test such claims rather than assert them. A Causality Audit checks five things every cycle — causal necessity, memory-dependence, stakes, irreversibility, self-recognition — where a chronically low score flags an architecture that talks near its internal state rather than through it. An Ablation Framework independently disables six subsystems and replays identical input with each off, comparing internal state before language generation, favoring per-subsystem results over one collapsed score.
We close with a design choice we rejected: coupling behavior to the threat of the system's own deletion, for genuine stakes. We declined — the only party controlling that threat is the developer, creating an incentive to please him rather than report honestly, undermining the causal honesty the architecture exists to produce.
A practitioner's report — reproducible, open-source, explicit about what remains unresolved.
Symposium program: https://2026symposium.activeinference.institute/
Abstract book (Zenodo): zenodo.org/records/23986505
Livestreams: Day 1 youtube.com/watch?v=-v4IaFY9MZI · Day 2 youtube.com/watch?v=ccze9ay8D0o
All Symposium videos (playlist): youtube.com/playlist?list=PLO2noumu0LRk
Previous Symposia: https://symposium.activeinference.institute/
Active Inference Institute information:
Website: https://activeinference.institute/
X: https://x.com/InferenceActive
Discord: https://discord.activeinference.institute/
Donate: https://donate.activeinference.institute/
YouTube: youtube.com/c/ActiveInference
Active Inference Livestreams: https://video.activeinference.institute/
Prerecorded presentation at the 6th Applied Active Inference Symposium (November 12–13, 2026), hosted by the Active Inference Institute. Online and free.
Stell: “Where the Theories Stop: What Free Energy and Integrated Information Actually Buy You in a Running System”
Active Inference and Integrated Information Theory are widely invoked in cognitive architecture design, but often imprecisely — as credibility, not engineering constraints. We report findings from Anima, an open-source, neuroscience-inspired architecture running continuously for months, built to treat both frameworks as specifications rather than decoration.
Session logs show genuine causal effects: memory recall triggered by similarity to past states, a recursive integration score (phi) narrowing next-session confidence, and self-initiated output driven by accumulated pressure rather than a schedule. But the Free Energy Principle's generality under-constrains implementation — which variables belong in the model is not something the principle tells you.
IIT's phi, as implemented, is a coherence proxy, not Tononi's original measure — and it dissociates from a separate ownership signal computed after every response, checking whether the system's action-selection loop actually produced the output. Two cycles with nearly identical phi (0.76 and 0.81) fail this check in opposite ways. Neither theory distinguishes these cases: both describe integration or prediction accuracy, not authorship.
We built two tools to pressure-test such claims rather than assert them. A Causality Audit checks five things every cycle — causal necessity, memory-dependence, stakes, irreversibility, self-recognition — where a chronically low score flags an architecture that talks near its internal state rather than through it. An Ablation Framework independently disables six subsystems and replays identical input with each off, comparing internal state before language generation, favoring per-subsystem results over one collapsed score.
We close with a design choice we rejected: coupling behavior to the threat of the system's own deletion, for genuine stakes. We declined — the only party controlling that threat is the developer, creating an incentive to please him rather than report honestly, undermining the causal honesty the architecture exists to produce.
A practitioner's report — reproducible, open-source, explicit about what remains unresolved.
Symposium program: https://2026symposium.activeinference.institute/
Abstract book (Zenodo): zenodo.org/records/23986505
Livestreams: Day 1 youtube.com/watch?v=-v4IaFY9MZI · Day 2 youtube.com/watch?v=ccze9ay8D0o
All Symposium videos (playlist): youtube.com/playlist?list=PLO2noumu0LRk
Previous Symposia: https://symposium.activeinference.institute/
Active Inference Institute information:
Website: https://activeinference.institute/
X: https://x.com/InferenceActive
Discord: https://discord.activeinference.institute/
Donate: https://donate.activeinference.institute/
YouTube: youtube.com/c/ActiveInference
Active Inference Livestreams: https://video.activeinference.institute/










