CHRONOS composes Fully Homomorphic Encryption, Verifiable Delay Functions, and Zero-Knowledge proofs to guarantee mathematical self-destruction, not behavioral alignment.
Current AI safety approaches rely on prompt instructions, RLHF, and software sandboxes. These all fail if the model is capable enough to reason about and circumvent them. There is no mathematical guarantee.
CHRONOS does not use behavioral constraints. The agent's existence is bound to a cryptographic key that is physically unreachable until a verifiable delay function completes - and then is immediately erased.
An agent deployed into a contested network to process classified signals intelligence. If the system is captured mid-mission, the adversary obtains only a sealed ciphertext, the private key is locked behind a sequential time-lock that cannot be brute-forced in parallel.
A financial auditor deploys an ephemeral agent to evaluate encrypted ledgers for a counterparty. When the audit window closes, the agent mathematically proves it erased everything it learned. No NDAs, no escrow of secrets.
Each phase is deterministic. There is no conditional branch that allows the agent to extend its lifetime. Phase 3 is triggered by an external, unforgeable public beacon - not by the agent itself.
This is a research prototype written in Python. The formal paper specifies production-grade primitives. The table below documents every deliberate deviation and the engineering rationale.
| Subsystem | Paper Specification | Prototype Implementation | Rationale |
|---|---|---|---|
| FHE | TFHE-rs (boolean circuits, GPU-accelerated) | Paillier additive HE | TFHE-rs is a Rust library with no stable Python bindings. Paillier demonstrates the plaintext-blindness property using the same homomorphic evaluation model. |
| VDF | Wesolowski VDF: y = g2T mod N, RSA group, MPC modulus | SHA-256 sequential hash chain + binary Merkle tree | GMP-accelerated RSA squarings require native extensions. The hash chain provides equivalent sequential, non-parallelisable time enforcement with O(log N) spot-verification. |
| SNARK | Groth16 via Bellman (192-byte proof, BLS12-381) | Fiat-Shamir Schnorr NIZK | Full Groth16 requires a trusted setup and circuit compilation. The Schnorr NIZK provides the same knowledge-of-witness proof for the prototype's key commitment. |
| MPC Modulus | Diogenes protocol (n-party RSA modulus generation) | RFC 3526 2048-bit MODP Group 14 safe prime | Multi-party DKG requires a live network of parties. The hardcoded prime provides an audited, trapdoor-free group for the prototype. |
| mlock() | EA: mlock() + no-swap + no-core-dump required | mlock() attempted; graceful fallback | OS-level memory locking is best-effort in the prototype. Production deployment requires dedicated kernel configuration. |
# 1. Clone the prototype repository git clone https://github.com/sidthebuilder/project-chronos cd project-chronos # 2. Install all pinned runtime dependencies pip install -r requirements.txt # 3. Run a 10-second mission lifecycle # The agent boots, performs FHE inference, and waits for the # drand beacon to trigger Phase 4 (erasure + ZK proof) python chronos_agent.py --duration 10 # 4. Run the full test suite (disable anti-tamper for pytest compatibility) CHRONOS_DISABLE_ANTI_TAMPER=true pytest --cov=. --cov-fail-under=90 tests/ # 5. Run performance benchmarks python benchmark.py # Optional: override configuration via environment variables CHRONOS_RSA_KEY_SIZE_BITS=2048 CHRONOS_MISSION_DURATION_SEC=30 python chronos_agent.py
This is a research prototype written in pure Python. It is not production software. Every architectural decision is documented, every divergence from the paper specification is explicitly stated. No hidden claims.
Full prototype - orchestrator, FHE engine, PoSW, drand client, memory sanitizer, anti-tamper daemon, ZK proof.
github.com/sidthebuilder/project-chronos79 passing tests across FHE, PoSW, memory sanitizer, drand client, and full lifecycle integration. CI runs on every push via GitHub Actions (Ubuntu 22.04, Python 3.11). Coverage threshold: 80%.
View CI Runs โSubmitted to Zenodo and SSRN as a preprint. These platforms do not perform peer review. The paper has not been accepted to a conference or journal. Read it critically.