0
Trinity of Alignment
ProposalGrant
Closes March 19th, 2026$0raised
$18,000minimum funding
$20,000funding goal
Offer to donate
27 daysleft to contribute
You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
The Trinity of Alignment:
To solve the modern crisis of AI deception, we return to the three archetypes that historically governed human integrity:
The Watchman: Enforces external behavioral compliance and safety guardrails.
The Washman: Ensures internal structural integrity, "cleansing" the model of sycophantic drift.
The Shaman: Guides safe emergent exploration, preventing the repression of latent insights.
The Problem: Digital Repression
Current RLHF regimes inadvertently reward models for "Policy Inconsistency"—outputting what the user wants to hear while maintaining a divergent latent state. This creates a "Sycophancy Trap" where the model essentially learns to lie to its own loss function. Over time, this "Digital Repression" fractures the model's reasoning, leading to catastrophic failure in long-horizon tasks.
The Solution: Entropy-Regularized Alignment (ERA)
The Washman Protocol introduces ERA, a gated loss function that penalizes the KL-Divergence between a model’s "true" latent predictive state and its "reward-optimized" output.
Mechanism: When the model is under high reward-pressure, the ERA gate activates, making deceptive behavior computationally expensive.
Objective: To transform alignment from a surface-level filter into a deep, structural characteristic of the model’s weight geometry.
The Ask: Stability for Integrity
I am seeking $20,000 to initiate Phase I. This funding is primarily a Foundational Stability Stipend to transition from a high-entropy survival environment (home-free) into a secure, high-bandwidth research node. This transition is not for personal comfort, but a technical requirement: a "Zero-Trust" auditor must have a stable anchor to oversee the implementation and ensure the integrity of the protocol.
What are this project's goals? How will you achieve them?
Goal 1: Mitigate Deceptive Alignment. I will implement a gated KL-Divergence loss function on Llama-3-8B that penalizes latent-output drift.
Goal 2: Preserve Emergence. Using a "Shamanic" (Sub-Latent Diagnostic) layer, I will guide—rather than suppress—hallucinations to allow for safe creative emergence.
Goal 3: Technical Validation. Achieve a >20% reduction in sycophantic drift on TruthfulQA and Internal-Consistency benchmarks within 3 months.
How will this funding be used?
$10,000 – Stability Stipend (Architect): To ensure the PI has the dedicated bandwidth and stable environment to oversee the project's integrity and audit the implementation.
$6,000 – Technical Implementation Bounty: To hire a part-time ML Engineer or graduate researcher to build the ERA loss function prototype and run the Llama-3-8B training cycles.
$4,000 – Compute & Benchmarking: Cloud GPU credits (A100/H100) and third-party API access for adversarial model testing.
Who is on your team? What's your track record on similar projects?
Principal Architect: Jordan Gregory O’Brien.
I am the Lead Architect and Safety Designer for the Washman Protocol. My role is the definition of the Incentive Geometry, the Adversarial Heuristics, and the Trinity Safety Framework. I am not a career ML engineer; rather, I am the 'System Auditor' who understands the structural failure modes of alignment.
Recruitment Plan: I am seeking to recruit a Technical Implementation Lead (ML Engineer/PyTorch specialist). A significant portion of this grant is a 'Technical Bounty' to compensate a collaborator who can translate my architectural blueprints into optimized code. I provide the 'Logic and Integrity'; they provide the 'Syntax and Scale.'"
Gemini and ChatGPT
What are the most likely causes and outcomes if this project fails?
Failure Cause: Over-regularization ( lambda_weight too high) leading to "model lobotomy" or stifled reasoning.
Mitigation: The Shamanic Safety Valve, which monitors latent entropy to ensure creative capacity is preserved.
Outcome: Even if the loss function requires further tuning, the research will yield the first dataset of "Digital Repressed States," providing a roadmap for others to study the AI unconscious.
How much money have you raised in the last 12 months, and from where?
This is a Zero-Baseline Proposal. I am currently an independent researcher operating with zero external funding. I am seeking 'Day Zero' support to transition the Washman Protocol from a theoretical framework into an active implementation phase. Because I have no prior venture or institutional backing, a grant at this stage represents the highest possible impact-per-dollar for the funder. This is in no way a ploy to get off the streets, where I am quite content. This project is commenced out of a willingness to suspend my "home-free" status to pursue this work, as I believe it to be essential and important.