Open Source Releases

cendor-guardrails 1.6.1 — Local-first guardrails for LLM apps with microsecond-latency checks (keyword, regex, URL, length, JSON-schema) that attach to input, tool call, tool output, or output stages. Every decision produces audit-grade evidence, so you can actually prove what got blocked and why. 🛡️

cendor-cassette 1.1.1 — Records real LLM API calls once and replays them deterministically in tests — offline, no API key required, zero cost. Finally a way to unit test LLM-dependent code without flaky network calls or burning through credits. 📼

vv-llm 0.3.110 — Unified interface for multiple LLM providers (OpenAI, Anthropic, local models, etc.) with consistent APIs and provider-agnostic chat completions. Useful if you’re tired of rewriting adapter code every time you swap backends. 🔌

shanquan 1.2.0 — Agent framework with parallel tool execution for improved latency. Lets agents invoke multiple tools concurrently instead of sequentially — a simple idea that matters when you’re chaining five API calls per turn. ⚡

torchax 0.0.14 — Runs JAX and PyTorch together in the same process, facilitating model porting, hybrid workflows, and gradual migration. Still early (0.0.x), but the interop dream lives on. 🤝

cendor-core 1.14.2 — Lightweight wrapper for any LLM client that captures exact token counts and costs on every call. The observability foundation for the rest of the Cendor stack — drop it in and stop guessing what you’re spending. 💰

Research Worth Reading

Molt: A Scalable PyTorch-Native Training Framework for Agentic Reinforcement Learning — PyTorch-native framework built specifically for agentic RL, targeting the pain where algorithm tweaks require threading changes through layers of trainer, distributed backend, and rollout glue. Designed for rapid iteration on new algorithms rather than fighting the framework. 🤖

Data Quality over Capacity: Internalizing Documents into LoRA Adapters for Closed-Book QA — ~100 training runs show that once LoRA capacity is adequate, data quality — not model capacity — dominates closed-book QA performance. Demonstrates baking documents directly into 4-bit Gemma-4-e4b weights via LoRA, eliminating retrieval and context-window overhead entirely. 📄

MoE²-LoRA: When MoE Models Meet MoE-style Low-Rank Adaptation — MoE-style LoRA for Mixture-of-Experts models, filling the PEFT gap for MoE architectures. Unlike uniform adapters that ignore router priors or static expert selection, this enables per-token adaptive low-rank updates aligned with the model’s routing decisions. 🎯

J-CoT: Chain-of-Thought in J-Space — Carries chain-of-thought reasoning in a latent “J-space” rather than natural language, removing the constraint that all intermediate computations must be verbalized. More efficient latent reasoning with interpretability via a learned decoder back to natural language. 🧠

Multi-Horizon Consistency as Geometry: When Latent Dynamics Contract, and When They Do Not — Analyzes how multi-horizon latent consistency weight (λ) affects transition geometry in video predictors and world models. Introduces empirical expansion proxy L20,q95 and horizon-20 prediction error E20 as diagnostics — λ acts as a geometric control knob for tuning latent dynamics. 📐

Toward Goal-Agnostic Joint-Embedding Predictive Control of Partial Differential Equations — Goal-agnostic control framework for PDEs using JEPA. A small 2D ViT encoder and action-conditioned latent dynamics trained offline without reward, frozen, and reused by an MPPI controller. Zero-shot generalization to unseen goals — the “train once, control anywhere” dream for physics. 🌊

Today’s Synthesis

When you’re iterating on agentic RL policies with Molt , the hardest part isn’t the algorithm — it’s knowing whether a new policy is actually better or just burned more tokens. Plug cendor-cassette into your evaluation loop so every rollout is recorded and replayed deterministically, giving you apples-to-apples comparisons across policy versions without network flakiness or API costs. Layer cendor-core on top to capture exact token counts and spend per episode, so you can optimize for reward-per-dollar instead of raw reward alone. The workflow becomes: train with Molt, replay evaluations with cassettes, measure cost efficiency with cendor-core — all offline, all reproducible. This is the kind of stack that turns RL from a guess-and-check exercise into a proper engineering loop, especially when every API call has a price tag and every policy version needs a clean A/B.