Annotations should be earned, not asserted.
CascadeAnnote treats every label as a falsifiable claim — produced by a transparent reasoning chain and anchored to a content-addressed receipt that anyone can audit.
One model, one shot, one prayer is not enough.
The default annotation workflow is to pipe text into a single model and trust the output. That approach is fragile: it provides no evidence, no calibration, and no recovery path when the model is uncertain. CascadeAnnote refuses to ship that as production behaviour.
Instead, every label is the output of a four-stage pipeline. Layer 1 retrieves the most relevant labeled exemplars. Layer 2 composes a structured chain-of-thought prompt around them. Layer 3 runs self-consistency voting across diverse temperatures. Layer 4 detects low confidence and engages a deeper, cooler re-vote with a wider evidence window.
Every layer emits a typed trace. Every annotation produces a content-addressed receipt that can be anchored to 0G Storage and 0G Chain. Verification is a side effect of how the system runs — not a feature you bolt on later.
A focused production stack.
Runtime
Next.js 15 + React 19
Single deployment, App Router, edge-friendly API routes.
Engine
Pure TypeScript
TF-IDF retriever, structured CoT prompt, ICL classifier, voter, fallback.
Data
0G Storage + Chain
Receipts content-addressed and anchored on-chain when configured.
Where this is heading.
- 01
Bring-your-own embeddings
Plug in OpenAI, Cohere, or local embedding models for Layer 1.
- 02
Active learning loop
Surface low-confidence annotations as next-to-label suggestions.
- 03
Multi-agent voting
Run multiple agent identities in parallel and aggregate the cascade.
- 04
Sealed inference adapter
First-class 0G Sealed Inference path for confidential annotation.