Project summary

I propose a 12-week, part-time research pilot on keeping AI safety evidence valid after changes to models, prompts, tools and operating assumptions. Using controlled, sandboxed agent tasks, I will compare static evidence reuse, full re-evaluation and dependency-aware selective re-evaluation. The study will measure missed invalidations, unnecessary re-testing, reproducibility and compute cost. Outputs will be an executable benchmark, an evidence-dependency schema and a report including negative results.

What are this project's goals? How will you achieve them?

Weeks 1–2: define safety claims, change categories and independent invalidation labels before benchmarking. Weeks 3–5: implement a sandboxed harness and the three comparison methods. Weeks 6–9: run model, prompt and tool changes alongside unchanged controls; record run provenance, missed invalidations, unnecessary invalidations, re-test cost and replay success. Weeks 10–12: seek independent technical review, analyse failures and document applicability limits. No reviewer is currently appointed.

How will you know it worked?

Success is a reproducible comparison with explicit uncertainty and evidence about whether selective re-evaluation reduces cost without increasing missed invalidations on the tested tasks. A negative result is useful and will be reported. Small-benchmark success would not establish general safety, certify deployment or demonstrate a reduction in catastrophic risk. The intended benefit is to help evaluators avoid relying on obsolete assurances when AI systems change; adoption and external validity remain uncertain.

How will this funding be used?

The full pilot requires US$13,200: researcher time $9,000 (180 hours at $50 gross/hour, including personal tax); compute/model access $1,500; independent technical review $1,000; reproducibility tools/storage $500; contingency $1,200. These are prospective estimates, not incurred expenses or supplier quotations. No travel or degree tuition is requested. The proposed schedule is 11 January–4 April 2027, 15 hours/week, subject to funding and payment confirmation. Work will be remote from Bengaluru, India; payment remains subject to Manifund's eligibility and administrative review.

Who is on your team? What's your track record on similar projects?

I am Srinivasa Dinakar Jainti, an independent researcher with over 18 years in semiconductor front-end design and verification, including SystemVerilog/UVM, assertions, PCIe/AXI, formal verification and emulation. I hold an MS in VLSI-CAD and am pursuing a DBA with a generative-AI specialization. My verification background supports test design and failure analysis; transfer to AI safety remains a research hypothesis.

Public AI-assisted AgenTrust contributions include weight-custody-manifest #248 (model-based lifecycle tests over 7,776 length-five operation sequences and targeted mutation checks), trace-registry #167 (an offline verifier corpus with 28 accepted/refused cases across five operations), and cmcp #823 (paired deterministic tests separating blocked gateway admission from delegated worker completion). These are examples of bounded test engineering, not proof of comprehensive AI safety.

Portfolio: https://github.com/dinakarjs

https://github.com/agentrust-io/weight-custody-manifest/pull/248

https://github.com/agentrust-io/trace-registry/pull/167

https://github.com/agentrust-io/cmcp/pull/823

What are the most likely causes and outcomes if this project fails?

The benchmark may be too narrow, invalidation labels may be unreliable, selective re-testing may miss relevant dependencies, or its bookkeeping cost may exceed savings. Model changes and API variability may limit replay. External review may be unavailable. I will document these limitations, report negative results and avoid claims beyond the measured tasks. The project uses synthetic or permissioned public materials, with no employer/customer confidential information or private repositories. No lobbying is planned.

How much money have you raised in the last 12 months, and from where?

No grant funding has been received for this pilot. Applications requesting $13,200 for the same pilot were submitted to EA Funds and Iliad on 2 November 2026; neither is treated here as an award. This is an alternative funding route for the same work, not an additional budget. Any overlapping award will be disclosed and the funding request adjusted to avoid charging the same costs twice.

Drafting disclosure

This proposal was prepared with AI drafting assistance. I am responsible for the final factual claims, research decisions and delivery.