The Active Inference Language Model is conditional probability tables and a forward filter. Both are below, executing in this page over the exported matrices. Nothing here is a recording of a result.
This is a generated local demo: no server, no network after generation, and a
strict subset. It is ignored by version control and is not a distributed
publication artifact. The full local dashboard analyzes input across every
built level with bidirectional chat and speech —
python -m ailm.dashboard, then
http://127.0.0.1:8080/. Only its L3→L1g policy path generates
replies; the other level panels are input analyses, not a reply cascade.
Matrix explorer: execute the exported letter and word matrices locally. Switch to Analysis & chat above for input analysis and L3→L1g replies. These views share the same model builders; neither claims a joint reply cascade.
The emission is the identity, so observing a letter reveals the state exactly and the posterior is a point mass. The next prediction is then one column of the transition. Type anything; the bars are the model's belief about the next symbol, and the running figure is perplexity over what you typed.
Generation is policy selection. The lexicon's Stage 3 posterior supplies a
preference
C over the next letter given the prefix so far; the learned bigram
supplies the habit E. Because the transition is deterministic and
controllable, the policy-conditioned predictive is a point mass, the
information-gain term of the expected free energy is identically zero, and
selection reduces to σ(γC + ln E). Move γ and
select generate to draw a new sample at that precision.
The case level's admitted margin is an aggregate over three populations whose labels have different provenance. Partitioned by which oracle path assigned each label, it is carried entirely by the stratum a dependency rule generated, is not shown where English marks case morphologically, and reverses on the token-level adposition path.
| oracle path | tokens | margin (pts) | reading |
|---|