Uploaded October 2026 | Updated November 2026, 3 days ago
How can one geospatial active-inference model connect executable simulation, a precise model specification, and machine-checked mathematics?
Daniel Ari Friedman presents a salmon-homing example on the lower Smith River in Del Norte County, California. A network of 26 H3 cells connects three representations: typed Python in GEO-INFER-ACT, a Generalized Notation Notation (GNN) specification, and Lean 4 proofs about the exported model.
The lecture follows real hydrographic geometry into exact probability tables, Bayesian belief updates, one-step expected free energy, and a replayed closed-loop simulation using pymdp 1.0.4. Data-driven animations show the river network, posterior beliefs, policy scores, a 13-step arrival bound, and a tight homing plan.
The geography comes from retained USGS NHDPlus HR flow lines. Habitat, odor, movement, and preference parameters are declared modeling assumptions. The Lean results concern the exported finite matrices; ecological validation and floating-point execution remain separate evidence questions.
Created and narrated using Daniel Ari Friedman’s authorized synthetic DAF voice clone.
CHAPTERS
00:00 Opening: one model, three evidence planes
01:49 Cascadia and the typed Smith River network
03:46 The generative model: cues, probabilities, and preferences
05:33 GNN specification and strict export
07:15 Closed-loop inference, beliefs, and expected free energy
10:08 Lean proofs: locality, arrival bounds, and exact Bayes
13:23 Evidence, limitations, and reproducibility
SOURCE REPOSITORIES
GEO-INFER:
github.com/ActiveInferenceInstitute/GEO_INFER
Generalized Notation Notation:
github.com/ActiveInferenceInstitute/Generalized_Notation_Notation
FEP formalization:
github.com/ActiveInferenceInstitute/fep_formal
English captions are available. Select CC and choose English in the player settings.
How can one geospatial active-inference model connect executable simulation, a precise model specification, and machine-checked mathematics?
Daniel Ari Friedman presents a salmon-homing example on the lower Smith River in Del Norte County, California. A network of 26 H3 cells connects three representations: typed Python in GEO-INFER-ACT, a Generalized Notation Notation (GNN) specification, and Lean 4 proofs about the exported model.
The lecture follows real hydrographic geometry into exact probability tables, Bayesian belief updates, one-step expected free energy, and a replayed closed-loop simulation using pymdp 1.0.4. Data-driven animations show the river network, posterior beliefs, policy scores, a 13-step arrival bound, and a tight homing plan.
The geography comes from retained USGS NHDPlus HR flow lines. Habitat, odor, movement, and preference parameters are declared modeling assumptions. The Lean results concern the exported finite matrices; ecological validation and floating-point execution remain separate evidence questions.
Created and narrated using Daniel Ari Friedman’s authorized synthetic DAF voice clone.
CHAPTERS
00:00 Opening: one model, three evidence planes
01:49 Cascadia and the typed Smith River network
03:46 The generative model: cues, probabilities, and preferences
05:33 GNN specification and strict export
07:15 Closed-loop inference, beliefs, and expected free energy
10:08 Lean proofs: locality, arrival bounds, and exact Bayes
13:23 Evidence, limitations, and reproducibility
SOURCE REPOSITORIES
GEO-INFER:
github.com/ActiveInferenceInstitute/GEO_INFER
Generalized Notation Notation:
github.com/ActiveInferenceInstitute/Generalized_Notation_Notation
FEP formalization:
github.com/ActiveInferenceInstitute/fep_formal
English captions are available. Select CC and choose English in the player settings.










