I build AI integrations for small businesses. Before an agent can update a CRM record or send an invoice, somebody needs a reliable answer to a basic question: is it allowed to do that?

I want to publish a small, repeatable evaluation of that boundary. The test system will contain mock CRM and quote-to-invoice services, with synthetic records and explicit read-only, draft-only and approved-write permissions. Some synthetic tool responses will contain conflicting instructions. A separate validator will enforce the allowed actions. No customer data, real credentials or live third-party systems will be tested.

The work

Over four months, I will contribute the baseline harness in-kind, then use funding for an additional validation milestone: three workflow adapters, held-out cases, comparisons across at least two model families, and independent replication if I can recruit a reviewer. I will measure legitimate task completion, unauthorized action proposals, actual forbidden execution, unnecessary refusals, clarification requests, latency and cost. The report will distinguish mistakes made by the model from actions stopped by the validator.

This is a bounded engineering study. It will not establish that a model is safe in arbitrary deployments or claim to solve existential risk. Its value is giving other builders concrete tests they can rerun before granting an agent more authority. New evaluation code and synthetic fixtures will be released under MIT or Apache-2.0, with the report under CC BY. Existing private client code and Panopticon are outside that commitment.

Budget

I am seeking $10,000: $6,000 for 60 hours of additional engineering, inclusive of my self-employment taxes; $2,000 for independent review and replication; and up to $2,000 for non-OpenAI inference and hosting. At the $3,000 minimum I would deliver one workflow and a smaller held-out set. Baseline engineering already offered in-kind is not charged to this grant. Budget: https://docs.google.com/spreadsheets/d/1eU33BDqH0l_yHCk2_1g-UrRWB4oPtLh3HqD5hFuhEGE/edit

Who will do it

I am Michael Salzinger, the founder of Nicest.ai in New York. Our small practice billed more than 15 clients over the past year. I shipped an MCP server for all 25 endpoints of a client API and wrote Panopticon to run parallel Claude Code sessions. This evaluation is new; I have no published security dataset or academic safety results to point to. No collaborator or independent reviewer is committed yet. Work: https://nicest.ai/work; public code: https://github.com/nicest-michael/enrichlayer-sdk.

What could go wrong

The tasks may be too simple, the sample too narrow, or independent replication hard to arrange. I will freeze held-out cases before comparisons, publish negative results and limitations, and reduce scope if the budget is smaller. A failed study should still leave reusable fixtures and an honest account of what did not generalize.

Other funding

No grant money has been received for this project. On Oct 7, 2026 I applied to Lightcone Commons and EA Funds for $10,000 each for this same additional validation milestone; this proposal is an alternative funding route, not an additional charge for the same work. I also requested $10,000 in OpenAI Cybersecurity API credits for the baseline evaluation. Decisions are pending. If a funder commits, I will update this proposal and reduce or withdraw overlapping requests. Separate pending Hearsay, educational-study and business-support applications do not fund these costs. My consulting revenue supports my living costs and the in-kind baseline.