Project summary

I measure how often an open-weight language model obeys a command hidden inside the data it was asked to read (an email, a PDF, a web page), and I lower that number with a fine-tuning recipe. I want to release both as an open benchmark with a public leaderboard, so anyone deploying an agent can pick a safer model and check that a fix worked.

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

Three months. Month 1: clean release of the measurement harness and the leaderboard on GitHub, under Apache-2.0, reproducible on a single free GPU (Colab T4 or Modal L4). Month 2: run the second batch, eight open models in three sizes each, and the correction recipe on four of them, with the weights published. Month 3: a write-up for an AI safety venue or arXiv, a public post, and a short guide for teams that deploy agents.

How will you know it worked?

The harness and the leaderboard are public and someone outside my company reproduces a row. At least four models have a before and after obey-rate with the weights downloadable. The recipe lowers the obey-rate without lowering task accuracy on the held-out tasks, with confidence intervals on every number.

How will this funding be used?

15,000 USD: three months of my own time, full time, at 5,000 a month. 3,000 USD: GPU time for the second batch and the four fine-tuning runs. Minimum 6,000 USD pays for the first month and the release.

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

Just me. Giuseppe Canale, CISSP, more than 25 years in software and security, co-founder and CTO of Silicon Psyche Labs in Italy, where I built the reader that measures model behaviour from output alone, benchmarked open prompt-injection defences on BIPIA, published a first injectability leaderboard and trained a first correction on a small model with the weights saved. I wrote the Cybersecurity Psychology Framework (cpf3.org) and a published machine learning book. Code: github.com/SiliconPsycheLabs/PSA-core and github.com/xbeat.

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

The obey-rate may be unstable across prompt formats and languages, which would make the leaderboard noisy; the mitigation is five languages and several document formats in the test set, with intervals shown. The correction may cost task accuracy; if it does, the benchmark still ships and the recipe is reported as a negative result. If I run out of money before the end, the harness and the first batch are released anyway in month 1.

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

Nothing for this work. The company applied to the EIC Accelerator in November 2026 for a broader product. Small individual grants (BlueDot, GTR, NLnet) were requested this week for the same open release; any overlap will be disclosed here and the goal reduced.