You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$36,000 left to reach the funding goal.
Coding agents now run commands, installs and deploys on real machines, and the agent's own report is usually the only record of what happened. CoreScout is an open-source, local-first Rust service and MCP server I built that watches those actions. On my machine it has logged 4,166 agent actions so far, and it caught a rebuild chain that failed 5 out of 5 times while nothing ever checked the result. This project turns it into a general audit layer and publishes the data.
1. Capture on any agent host, not just mine: a hook/MCP adapter that records each agent tool call, the outcome the agent reported, and the outcome observed independently (exit code, file state, whether the service is actually reachable). Claude Code hooks already work; add Codex CLI and one open-source harness.
2. An unverified-claim detector: flag actions reported as done that nothing ever checked.
3. Randomized checks: on a small random share of occasions the layer re-checks or holds the action, so claims about what works are causal, not just correlations. CoreScout already keeps 'seen together' and 'verified by trial' as separate labels.
4. Publish a scrubbed public dataset of claimed vs observed agent outcomes on Hugging Face, with a short write-up of base rates.
It runs on machines other than mine; at least 3 outside users contribute opt-in data; the dataset and write-up are public; at least one agent-harness project adopts the claim-checking hook or cites the numbers.
Minimum $8,000: 2 months of my time to ship the cross-host adapter and the unverified-claim detector. Full $36,000: 6 months ($5,000/month stipend), about $700/month of model API and compute for running agents under audit, and a small contingency. Adds randomized checks, the dataset and the write-up.
Just me (Matthew Busel, Tensorust LLC, New York). I have shipped alone: CoreScout (github.com/Mattbusel/corescout, 33 crates on crates.io), tokio-prompt-orchestrator (Rust LLM gateway, 60 GitHub stars), a Rust agent swarm harness with a shared build/test verification engine and a default safety guard that blocks agents from git push and credential reads, the SWE-Graph Lite dataset on Hugging Face, a Steam game and several iOS apps. 70 crates on crates.io in total, 9,267 downloads (much of it release-day mirror traffic, to be honest about it).
Most likely failure: nobody outside my machine installs it, so the dataset stays one machine deep. Mitigation: ship as a one-line Claude Code hook and post the first numbers early to recruit contributors. Second risk: the randomized checks are too rare to reach significance in 6 months; then I publish the descriptive base rates and the method, which is still useful.
$0 in grants or donations. I applied to the EA Funds Transformative AI Fund for this same project on Nov 3, 2026 (pending) and will update here if that comes through. Separate pending applications for other projects: NSF SBIR pitch, Tarbell, GitHub Secure Open Source Fund.