### Project summary

The Proposed Action Auditor (PAA) is an open-source, client-side, local safety sentinel that operates in real-time to protect against rogue autonomous agent task-calling behavior. Sitting right before a proposed action is sent to execution, PAA treats safety as a linguistic probability forecasting problem. Instead of relying on brittle human-designed rules, high-latency secondary LLM critics, or prompt-based instructions that are vulnerable to jailbreaks, PAA evaluates the token-level logprobabilities generated during inference to isolate anomalous or unaligned session trajectories and hold them for human override.

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

Our core goal is to deliver functional open-source weights for the Gemma 4 - e4b model alongside a fast Python model execution wrapper under an Apache 2.0 software license. We will achieve this via a three-stage engineering pipeline:

1. Data Ingestion & Extraction: Programmatically process all 335,122 multi-turn tool-calling trajectories from the Nvidia Nemotron dataset via local local models to extract naive Task Action Verbs (TAV) and Task Result Labels (TRL) context-mapped via the Belief-Desire-Intention (BDI) paradigm.

2. Grammar Conformance: Convert these diffuse, non-deterministic ad-hoc labels into vector embeddings and apply the DBSCAN clustering algorithm to build a conformed, mathematically canonized syntax known as the Task Action Language (TAL).

3. LoRA Fine-Tuning & Gating: Re-map the original trajectory database into conformed TAL sequences to train a LoRA adapter on Gemma 4 - e4b. At runtime, real-time logprobs for proposed agent tasks are evaluated against this token model; if the logprob falls below a calibrated safety threshold, the action is held for human confirmation.

### How will this funding be used?

Funding supports a strict 90-day R&D development cycle split across two operational tiers:

Tier 1: Minimum Baseline | $25,000 | • Dedicated Engineering Runway ($20,000): Solo architecture costs for parsing layers, DBSCAN pipelines, and local footprint validation.

• Compute & API Token Passes ($5,000): Mass dataset generation via DeepSeek API endpoints and cloud GPU cluster rentals (RunPod/Vast.ai) for LoRA training.

Tier 2:| Max Funding Goal | $40,000 | • Production Hardening ($15,000 expansion): Independent systems engineer code audit ($4,000), paid beta tester stipends for 10–15 agent builders ($4,000), high-value instructional video documentation ($3,500), and cloud adversarial red-teaming/hyperparameter tuning ($3,500).

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

Robert Oschler (Lead Architect & Solo Developer): I am a veteran senior software design architect and independent programming contractor with over two decades of full-stack engineering and low-level system orchestration experience. My core expertise is in building custom parsers, logic-processing engines, and robust client-side runtime containment wrappers. I have a long history of shipping functional open-source code and technical architecture, including recently engineering a zero-egress, client-side local edge-RAG knowledge engine running on browser-native Gemini Nano infrastructure that won a prominent open-source database prize from PingCAP/TiDB.

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

1. Mathematical Density Breakdown: If a small sample size or an out-of-distribution tool set causes the DBSCAN clustering algorithm to experience density collapse, the generated TAL primitives will map poorly to complex tasks.

2. Perplexity Spikes: Untrained or un-converged base-model weights could create severe initial false-positive gating behavior, holding safe actions and adding system latency.

* The Failure Outcome: If the machine learning engine fails to converge safely, the project still delivers an immediate open-source public good: the schema parser configurations, the parsed BDI trace documentation, and the conformed TAL context-free grammar repositories. This provides an empirical engineering foundation for other teams running agentic boundary verification.

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

$0 so far.