Skip to main content

Murali v0.1.7: Deterministic Composition and Semantic AI State

· 2 min read
Ravi Shankar
Ravi Shankar

Murali v0.1.7 strengthens the engine underneath technical storytelling and adds the first deep semantic layer for AI lessons.

One Timeline, Local-Time Clips​

A scene now has one runtime timeline. Reusable Clip sections begin at local time zero and can be appended, overlaid, or placed explicitly. Composition converts those local times into one deterministic schedule before playback.

Seeking is stricter too. Reversible operations reconstruct from baseline state, while one-shot callbacks, frame-dependent updaters, and history-dependent paths report when arbitrary seeking is not meaningful.

Semantic AI Workflows​

TensorSnapshot now carries shape, named axes, stable element IDs, and validated values. Murali can compute and animate matrix multiplication, masking, softmax, reshape, split/merge, broadcasting, sampling, higher-rank slices, and explicit 2D projections without flattening semantic axes.

The new versioned JSON trace boundary imports tokens, tensors, model metadata, and typed events. The self_attention_lesson example uses that boundary to compute Q, K, V, causal attention, residual output, logits, probabilities, and a deterministic sample from one coherent input.

A More Reliable Engine​

This release also adds stable render layers, stronger 3D depth behavior, automatic dirty-state tracking, dynamic mesh capacity, u32 mesh indices, transformed layout bounds, centralized validation, resize-correct Typst caching, and broader regression coverage.

The new roadmap contains only future work. Completed review items now live where they belong: in the documentation and release history.