REDDIT · REDDIT// 9d agoOPENSOURCE RELEASE

Phenomenological Compass uses 3B LoRA for deeper inference

The Phenomenological Compass is a two-stage local inference architecture that uses a 3B LoRA-tuned "compass" model to analyze the epistemic posture of a question before a larger "action" model generates a response. This "field-reading" approach increases token-level entropy by +0.47 nats, effectively allowing the model to "hold space" for complex, emotional, or philosophical inquiries rather than defaulting to clinical bullet points.

// ANALYSIS

Separating "machine intuition" from generation is a major step toward more human-like local AI, solving the "clinical response" trap by letting a dedicated model pre-shape the response manifold.

–3B LoRA (29MB) acts as an "epistemic router" using SHAPE, TONE, and SIGNAL dimensions to condition the action model.
–Measurable +0.47 nats entropy shift confirms the model actually holds more "possibility space" open during exploration.
–WITNESS signal achieved a 100% classification rate, allowing models to acknowledge questions without forced helpfulness.
–Optimized for Apple Silicon via MLX, the entire pipeline runs locally in 7–16GB of unified memory.
–Wins 90% of pairwise judgments against raw, unconditioned baselines across 800 live benchmarks.

// TAGS

llmfine-tuningedge-aiopen-sourcephenomenological-compass

DISCOVERED

9d ago

2026-04-03

PUBLISHED

9d ago

2026-04-03

RELEVANCE

8/ 10

AUTHOR

TheTempleofTwo