Skip to main content
The biggest risk with traditional LLM compression tools is syntax corruption: algorithms accidentally remove quotes, commas, brackets, or code tokens, causing downstream parsers or compilers to crash. Curtly solves this fundamentally through the Safe Vault™: an in-memory Abstract Syntax Tree (AST) boundary scanner that extracts and locks all critical technical tokens into an immutable vault before any natural language compression takes place.

​
The 5-Step Vault Lifecycle

  1. AST Boundary Scan: Identifies code fences, inline backticks, JSON keys and values, SQL queries, URLs, and template variables.
  2. Atomic Placeholder Replacement: Swaps each isolated entity with an atomic memory token (e.g. __CTLY_VAULT_0__).
  3. Natural Language Pruning: The compression engine prunes surrounding text without touching the protected tokens.
  4. Placeholder Matching: Verifies that every single placeholder remains intact and in the correct sequence.
  5. Byte-Level Rehydration: Rehydrates the original exact content back into position with 100% bit-for-bit fidelity.

​
Protected Token Categories

Multi-line code blocks (Python, TypeScript, Go, Rust, C++, Bash) including indentation, whitespace, language tags, and inline backtick code snippets are 100% preserved.
Valid JSON payloads, JSON Schema definitions, OpenAPI specs, and nested structures are protected from delimiter stripping or quote alteration.
SQL queries containing SELECT, JOIN, WITH, and WHERE clauses are isolated from grammatical restructuring.
Dynamic placeholders such as mustache variables or Jinja2 markers are protected from token splitting.
HTTP/HTTPS URLs, query parameters, endpoints, and complex regex patterns remain unmodified.