fix: Critique response — logit fluency, causal pruning, FHRR phase cancellation (#6)

cleanup: Remove all V1 legacy, add CognitiveAgent V3 (#5)

feat: V3 — SBERT-native cognitive stack (NGC, Hopfield, falsification all in embedding space) (#4)

feat: Persistent causal arena, BoolQ binary task fix, SBERT-only ablation baseline (#3)

feat: Self-tuning engine — Friston precisions, Dirichlet channels, joint settling, structured projection (#2)

fix: Benchmark regression fixes — emission gate, negation bug, noise dampening, SBERT channel (#1)

Add .gitignore and UPGRADE.md; update run.py and runner.py for persistent state management

Enhance README and scripts for cognitive architecture testing

Refactor README and add new scripts for cognitive architecture testing

Refactor logit bias processing and enhance CanonicalPipeline

Enhance cognitive controller and Broca interface with new features

Add initial project configuration and dependencies

v2 __init__: export SemanticFHRRCodebook

v2 graft: sentence-level sbert scoring as primary signal, semantic FHRR + NGC as complementary. SciQ 82%, TruthfulQA 30%, ARC-Easy 34% standalone.

Semantic FHRR codebook: project sentence-transformer embeddings into phasor space + sentence-level scoring in V2ScoringBridge

Confidence-gated blending for local mode: graft scales by (1-LLM_confidence), gated when LLM is confident (>0.6), scores scaled to 0.3x of LLM score range

Benchmark runner: switch from v1 template parser to v2 scoring bridge with z-normalized signal combination

v2 graft: multi-signal scoring with FHRR similarity + NGC energy + distinctiveness + adaptive convergence gating

Wire up v2 public API: export all v2 components

v2: precision-modulated learning, needle-in-lies 3/3, FHRR Hopfield scoring

v2 complete: NGC graft, causal energy, auto-expanding codebook, benchmark integration

v2 architecture: FHRR-RNS encoding, hierarchical predictive coding (NGC), unified energy landscape

Merge branch 'main' of https://huggingface.co/theapemachine/tensegrity

Add inference settings for model loading: prioritize CUDA, MPS, and CPU. Update runner and pipeline to utilize new settings for dtype and device placement.

Benchmark v2: flip accounting, lambda sweep, honest N=200 results, fixed bias_applied semantics

Add benchmark harness: 13 tasks, baseline vs grafted evaluation, structured reporting

Add logit-bias graft: beliefs → token biases + convergence gating + hybrid pipeline

Add Broca interface: LLM as linguistic transducer + hypothesis elimination benchmark

Add comprehensive README

Initial implementation of Tensegrity architecture

initial commit