Hi Hacker News! I'm Vladislav, an AI systems architect and founder.

Like many of you, I love "vibe coding" with autonomous agents (Cursor, Claude Code, Copilot, Cline). But once our production platform grew past 20,000 lines of code, the stochastic nature of LLMs hit us hard:

- Agents quietly overwriting 200 lines with `// ... existing logic unchanged`.

- Context entropy eating 4,000 prompt tokens for a 3-line UI tweak.

- Fragile CSS selectors breaking automated E2E tests after every minor refactor.

To solve this, we created and open-sourced the Utyansky Index v2.0 — a deterministic 5-digit coordinate protocol (00000–99999) that transforms fuzzy semantic search into strict O(1) AST slot isolation:

1. Spatial Isolation: Each component/function is tagged with explicit coordinates (e.g. data-idx="71397").

2. Bounded Mutation: When an agent needs to edit a feature, the orchestrator feeds strictly that 40-line slot AST rather than the entire 2,500-line file.

3. Reproducible Benchmark: We submitted the exact audit prompt to 9 frontier LLM families (Claude 3.5 Sonnet, GPT-4o, Gemini 1.5 Pro, DeepSeek, Qwen 2.5, Grok, etc.) — all confirmed 80-85% token reduction and zero accidental code regressions.

4. Open Science: The architectural specification is formally archived with an immutable DOI via CERN Zenodo.

GitHub Repo: https://github.com/vlad-utyansky/utyansky-index

Preprint DOI: https://doi.org/10.5281/zenodo.22934668

I'd love to hear your thoughts, feedback, or brutal engineering critiques!