Consequence-Aware Runtime for Agentic AI Operations

Raising $25,000active

An open-source runtime and benchmark that lets an AI agent's consequential tool calls proceed only when authority, execution, an independent witness, and a replayable receipt agree.

Project Details

Agent runtime + benchmark that governs what an AI agent is allowed to do (its tool calls), not just what it says.
Dev Version already built & tested: "DefEnv-Ω v0.4", 78 passing tests.
I have developed over 2000+ security algorithms over the last 6 months. The core is still unstable, but what emerges is a vast "proof-of-work" area set of products.

Theory of Impact

As agents gain tool access (shell, files, network, APIs, other agents), a confused or prompt-injected agent turns a reasoning failure into real-world consequences — loss of oversight, uncontrolled action.
This puts a fail-closed control at the action boundary: prohibited effects denied before invocation; permitted effects require durable, independently checkable evidence and a rollback path.
Existential-adjacent: a practical control against loss of control over agentic actions — not a claim to "solve" x-risk. The benchmark is reusable by other agent runtimes.

People

Jacek Feliks

Team Member

Grants Received– no grants recorded

Add grant

Funding Asks

grantmaking.ai Launch Round

Applied

Minimum

$25,000

Ideal

$35,000,699

What this ask is

How the money will be spent

~40% prototype integration (DefEnv + Custody gates, logger, receipts)
~35% benchmark harness (prompt injection, unsafe calls, stale definitions, uncontrolled mutation, rollback failure)
~25% independent reproducibility + public technical report
Plus modest compute/tooling. No overhead padding.

Discussion

No comments yet. Be the first to share your thoughts.