theapemachine

/

tensegrity

like 0

Critique Response: Logit graft fluency, causal arena pruning, FHRR phase cancellation

Three fixes from the audio review critique.

1. Logit graft: dynamic temperature scaling replaces hard -inf (logit_bias.py)

Before: Eliminated hypothesis tokens were suppressed to -inf, which "violently collides with the model's syntactic expectations" — the LLM can't use structurally necessary tokens (pronouns, conjunctions) even when grammar demands them.

After: Suppressed tokens get -max_bias (default -8.0) instead of -inf. This makes them very unlikely but not impossible. If the LLM's autoregressive prior strongly demands a suppressed token for grammatical coherence, it can still use it — the cognitive constraint shapes the distribution without breaking fluency.

Same fix applied to StaticLogitBiasBuilder for the remote API path.

2. Causal arena: structural pruning + DAG merging (arena.py)

Before: Every proposed SCM was registered as a separate competing model, even if structurally identical to an existing one. This risked combinatorial explosion of the 250,000 counterfactual world limit.

After:

• register_model() now checks for structural equivalence (same variables, same edges) before registration. Duplicate DAGs are merged by averaging their Dirichlet pseudocounts instead of adding a separate model.
• compete() adds an energy threshold filter: models whose single-step log-likelihood is >20 nats worse than the best are flagged for fast elimination, avoiding wasted counterfactual computation.

3. FHRR: sparse block coding + hierarchical temporal bundling (fhrr.py)

Before: bundle() added all complex phasors and normalized. With dense SBERT-grounded embeddings, bundling >20 vectors caused phase wrapping — the "superposition catastrophe" where constituent meanings wash out into noise.

After:

• bundle() accepts optional top_k parameter: when set, only the top_k dimensions with largest magnitude are preserved per vector before addition. This induces sparsity that prevents phase wrapping.
• encode_sequence() uses hierarchical temporal bundling: tokens are bundled within local windows (default 16), then window summaries are bundled with sparse top_k. Short sequences still use direct bundling. This preserves high-resolution semantic detail within recent context while summarizing distant tokens.