A Structurally Honest Reasoning Virtual Machine
SYNAPSE is a deterministic reasoning virtual machine (RVM). It executes structured reasoning programs over external knowledge, producing fully auditable, replayable derivation logs.
It is not a model. Not an agent. Not a database.
It is the runtime layer that sits between models and agents — the missing piece that makes AI reasoning trustworthy.
SYNAPSE is to reasoning what SQLite is to queries.
You give it a program, it produces a traceable derivation.
Imagine you ask a really smart friend a hard question. They give you an answer, but you have no idea how they got it. You just have to trust them.
Now imagine a different friend. This friend shows you every step they took: what facts they looked at, which ones they kept, which ones they threw away, and exactly how they connected the dots to reach their answer. If you don't believe them, you can replay their entire thought process and check it yourself.
SYNAPSE is that second friend.
Today's AI (LLMs like GPT, Claude, etc.) is the first friend — it gives answers but hides its reasoning. SYNAPSE doesn't replace those AIs. Instead, it acts as a referee that forces them to show their work, prove their claims with evidence, and never make things up.
Today's AI stack has a gap:
LLM → Agent
This jump skips a critical layer. Without it, you get:
- Hallucinations — claims with no evidence
- Hidden reasoning — no way to audit how a conclusion was reached
- Non-replayable behavior — run the same thing twice, get different results
- No accountability — failures are silent and invisible
SYNAPSE fills that gap:
LLM → SYNAPSE (Reasoning Runtime) → Agent
SYNAPSE programs are composed of six operators, each grounded in peer-reviewed research:
| Operator | What It Does | Theoretical Basis |
|---|---|---|
| FETCH | Retrieves relevant evidence from knowledge substrates | OrLog (Hoveyda et al., 2026) |
| EXPAND | Context-aware graph traversal with dynamic edge weighting | CatRAG (Lau et al., 2026) |
| ASSERT | Probabilistic predicate validation over logical constraints | OrLog (Hoveyda et al., 2026) |
| ATOMIZE | Decomposes reasoning state into simpler atomic subproblems | Atom of Thoughts (Teng et al., 2025) |
| EMIT | Produces citation-aligned structured output from evidence | C²-Cite (Yu et al., 2025) |
| COMPRESS_LOG | Compresses execution trajectories into reusable summaries | RE-TRAC (Zhu et al., 2026) |
Every SYNAPSE execution must satisfy:
- Deterministic — same program + same input = same output, always
- Bounded — execution has a strict resource budget and always halts
- Replayable — any execution can be replayed from its log
- Auditable — full evidence chain, cost profile, and predicate results are logged
- Failure-explicit — failures are never hidden; they terminate execution with a named reason
The full formal specification is available at:
📄 src/content/docs/spec/spec.md
This document defines the complete architecture, operator model, type system, failure model, and replay guarantees.
Current version: v0.0.4 — the specification is under active development.
⚠️ Not yet implemented. The implementations below will be developed once the specification reaches a satisfactory level of completeness and review.
- Python: -- to validate and test if such concept does in fact work as expected.
- Rust: -- to be implemented in rust during v2 and pmf, scale.
This project is in its early specification phase. We actively invite:
- 🔍 Critique — challenge the assumptions, point out contradictions
- 💡 Suggestions — propose new operators, alternative formalisms, better invariants
- 🐛 Bug reports — find gaps, ambiguities, or inconsistencies in the spec
- 📝 Improvements — clarify language, add examples, improve accessibility
The spec is the product. If something doesn't make sense, that's a bug — please open an issue.
How to contribute: Open an issue or submit a pull request. All feedback — from casual observations to formal critiques — is valued.
- Fengwei Teng et al. (2025) Atom of Thoughts for Markov LLM Test-Time Scaling. arXiv:2502.12018
- Kwun Hang Lau et al. (2026) Breaking the Static Graph: Context-Aware Traversal for Robust Retrieval-Augmented Generation. arXiv:2602.01965
- Mohanna Hoveyda et al. (2026) OrLog: Resolving Complex Queries with LLMs and Probabilistic Reasoning. arXiv:2601.23085
- Yue Yu et al. (2025) C²-Cite: Contextual-Aware Citation Generation for Attributed Large Language Models. arXiv:2602.00004
- Jialiang Zhu et al. (2026) RE-TRAC: REcursive TRAjectory Compression for Deep Search Agents. arXiv:2602.02486
Licensed under the Apache License, Version 2.0.