You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
Syzygy Rosetta is pre-output governance middleware for AI systems. Current safety tools operate reactively at the output layer. Rosetta intercepts at the interaction layer, before generation — using a hybrid architecture: deterministic policy rules form the constitutional enforcement floor, while semantic and heuristic evaluation handles contextual nuance. No probabilistic drift can override hard invariants. Built for multi-model enterprise deployments where interaction failure, not output failure, is the primary liability.
Goal: Complete high-load latency stress-testing and adversarial jailbreak testing of Syzygy Rosetta v1.1’s hybrid architecture — specifically the deterministic policy layer and its interaction with semantic/heuristic backends. We will run structured compute trials across multi-model API configurations and commission a security researcher to attempt systematic bypass of both the deterministic rules and the coherence evaluation logic. Deliverable: published stress-test report and updated invariant set.
~$20K compute costs for high-load latency testing across multi-model configurations; ~$15K security researcher contract (adversarial jailbreak scenarios targeting both deterministic and semantic layers); ~$10K lead developer time for iteration and documentation. All spending reported in weekly public updates.
Sarasha Elion (founder, Trivian Institute 501c3) — 14+ months cross-platform AI governance research, MFA, 20+ years contemplative/somatic practice informing relational AI design. Faiyaz (lead developer) — architecture and implementation of the full Rosetta stack including FastAPI middleware, policy pack inheritance system, and hybrid deterministic/semantic evaluation pipeline. Meili Liang (ML Risk Engineer) — validated Rosetta across 21 structured attack scenarios.
Primary risk: latency-safety tradeoff — the deterministic policy layer plus semantic evaluation introduces overhead that may be impractical at enterprise scale. Secondary risk: adversarial testing reveals gaps where probabilistic backends can be manipulated to circumvent deterministic floors. Outcome if failed: findings published openly as negative results; the governance gap remains documented for future builders.
$0 external funding to date. Self-funded research phase through Trivian Institute. This Manifund application is the first external capital raise, intended as bridge funding while NSF and foundation applications are in review.
There are no bids on this project.