Our AI systems are built on three commitments: every decision is explainable to the person it affects, every model serves human judgment rather than substituting for it, and every system is designed to be auditable, correctable, and shut off.
GameChangers OS uses AI to verify, match, and assist — but humans make every consequential decision. A Mission Intelligence Agent can respond to inquiries. The human Mission Lead approves the contract. Curious PI surfaces matched candidates. The mission lead selects who to contact. The AI reduces friction. The human retains judgment.
This is not a philosophical choice. It is an architectural one. Every AI action on the platform has an explicit permission gate, an audit log entry, and a human override path. The Delegated Agency Clause in every contract where agents operate makes this legally explicit, not just technically true.
Every Trust Score is broken down by component. Every Curious PI match shows a reasoning trace. Every agent action is logged with its input, output, and decision path. If a professional asks "why is my score X," we can show them exactly.
Trust Score is computed using deterministic math — identity(100) + professional(250) + research(200) + federal(150) + activity(150) + network(150). It is not a neural network prediction. It is a formula you can audit.
We do not fine-tune models on assumptions. The Curious PI embedding pipeline uses Anthropic's embeddings API against pgvector. Fine-tuning is deferred until we have 10,000+ real, validated professional-mission matches to learn from.
Routes tasks to the right model: Claude for complex reasoning and contract analysis, GPT-4o for code and technical review, Gemini for document analysis, Perplexity for real-time research, Llama for classified contexts, Mistral for EU data residency.
What it doesn't do: Make final decisions. Aria routes and synthesizes. Humans approve.
Profile text and mission requirements are embedded using the Anthropic embeddings API. Cosine similarity search against pgvector returns ranked matches. Trust Score minimum and ITAR clearance requirements filter before similarity ranking.
What it doesn't do: Score professionals on inferred demographic signals. Skills, verified credentials, and mission fit drive ranking — not zip code, name, or institution prestige.
Agents operate within 5 explicit permission gates, each with a Redis rate limiter. Every action is logged. The Delegated Agency Clause in every agent-enabled contract establishes that the human Mission Lead retains sole authority for all approvals, financial decisions, and dispute resolution.
What it doesn't do: Act without a permission gate. There is no "general agent" mode.
Every Trust Score component is computed from independently maintained public registries — not from what the professional tells us about themselves. This is the core of what makes verification meaningful rather than cosmetic.
How we handle bias prevention, DEI data encryption, security clearance transparency, and the ethical boundaries of AI delegation.
Read AI Ethics Framework →