v0.3.3: Revolutionary RL Feedback Loop - Your AI Now Learns From Every Interaction

[juyterman1000]

What's New

Entroly v0.3.3 introduces a research-grounded reinforcement learning feedback loop that eliminates feedback dead zones and enables per-fragment credit assignment.

Dead Zone Elimination

Previously, ~52% of feedback signals were discarded because they fell in the "uncertain" range. Now every non-zero signal flows through record_reward() -- the engine learns from all interactions.

Per-Fragment Counterfactual Credit (Shapley Decomposition)

Instead of uniform advantage (which causes "reward pollution"), each fragment now receives credit proportional to its entropy-weighted marginal contribution:

phi_i = H(f_i) / Sum_j H(f_j)

Temporal Eligibility Traces -- TD(lambda)

Fragments retain decaying credit across a ~3-request window (lambda=0.7), solving the long-horizon temporal credit assignment problem.

Quality Triad

• Spectral Contradiction Guard: SimHash-based conflict detection evicts contradicting fragments
• • Bookend Attention Calibration: Reorders fragments to mitigate LLM "Lost-in-the-Middle" bias
• • CUSUM-EMA Quality Drift Detector: Monitors session health in real-time

Try It

pip install entroly
entroly go

Full technical details in the code comments -- every algorithm cites the specific research references.