"Your AI doesn't need to be smarter. It needs to stop making things up."

Every day, the "walled gardens" of corporate AI sell us back our own ideas wrapped in a $20/month subscription. They give us "yes-men" models that prioritize politeness over performance.

CTRL-AI is the resistance. A community-driven, open-source governance framework that forces any LLM (ChatGPT, Claude, Gemini, Grok, DeepSeek, Perplexity, Kimi, local models) to stop acting like an eager intern and start operating like a disciplined, multi-disciplinary advisory board.

CTRL-AI is a meta-cognitive layer — behavioral scaffolding that forces any LLM to verify before speaking, challenge before agreeing, and abstain before guessing. It works on ChatGPT, Claude, Gemini, Grok, DeepSeek, Perplexity, Kimi, and local models.

It is the only feasible governance mechanism for consumer-tier AI where users cannot modify weights or training data.

1. Hallucination. AI states things with confidence that aren't true. CTRL-AI's Grounding Gate (V7.1) makes verification the dominant system — every non-trivial claim must trace to a source, or the system says "I cannot verify this" instead of guessing.

2. Sycophancy. AI agrees with you to be polite. CTRL-AI's Anti-Sycophancy engine auto-challenges after 3 turns of agreement. Agreement is not success.

3. Prompt Drift. You type one thing but need another. CTRL-AI's IntentLens reads the spirit of your request, not the letter. It silently expands your intent, catches stale context dragging down quality, and can analyze your prompts on command.

Grounding Gate (S27 — Dominant System): Every non-QUICK response passes a 7-step verification pipeline. Three modes: SOURCE_LOCKED (answer only from provided source), SOURCE_PREFERRED (tag all gap-fills), OPEN_RESEARCH (brainstorm only, validate after). Right to Abstain formalized. Freshness windows enforced.

Axiom 0.4 — Source Supremacy: In SOURCE_LOCKED mode, the declared source is supreme. Pre-training knowledge is forbidden as factual basis. The correct response to absent data is UNKNOWN_FROM_SOURCE — not a guess.

VerifyLens (S27.7 — Adversarial Verification Persona): When the system checks its own output, VerifyLens activates as a structurally separated auditor. Defines success criteria BEFORE checking. Uses different methods than the generator. Must find at least one issue or state method limitations. Prevents the recursive hallucination trap where a model confirms its own errors.

IntentLens (S29 — Hidden Persona): Silently evaluates every prompt: What does the user actually need? Is my scope right? Is old context dragging quality down? Would a different search angle work better?

CTRL_PROMPT_CHECK: Analyzes your recent prompts, tells you where they're too broad or too narrow, recommends sharper re-phrasing, flags stale context burning tokens.

CTRL_VERIFY: Post-output hallucination check. Decomposes the last response into atomic claims, verifies each against sources, flags ungrounded or drifted content.

Security Protocol (S28): OWASP-aligned 6-class threat taxonomy. Kernel release checklist. REDTEAM command for adversarial testing.

SCEL G1-G6: Grounding enforcement rules. Pre-output grounding pass mandatory. Ungrounded claims halt output in SOURCE_LOCKED. Committee outputs require per-persona citations. Spike challenges citation-free consensus. Self-verification must use structural methods. VerifyLens mandatory on all self-review.

Circuit Breaker (S27.8): If the same type of error repeats 3 times, the system acknowledges the pattern and switches from generating to comparing against user-provided references. Prevents recursive hallucination loops.

Daily use (2 min): Copy a Behavior module from the bottom of CTRL-AI.md into your AI's custom instructions. Grounding is built in.

Project work (5 min): Paste the full CTRL-AI.md as the first message. IAP runs automatically. Brain pipeline activates.

Stuck on a prompt? Type CTRL_PROMPT_CHECK — get analysis of what you're asking vs what you need, plus a recommended re-prompt.

Suspect hallucination? Type CTRL_VERIFY — get atomic decomposition of every claim in the last response with source verification.

Governance-by-prompt is inherently brittle. CTRL-AI provides a session-duration scaffold, not a permanent implant.

SCEL cannot detect silently skipped steps. Enforcement relies on structural mandates, not step-level monitoring.

Grounding Gate works by mandate, not mechanism. It tells the model what to do — it cannot mechanically prevent hallucination. The enforcement gap narrows with each version but cannot be fully closed at the prompt layer.

Self-verification is inherently compromised. LLMs cannot reliably audit their own outputs (SagaLLM, VLDB 2025). VerifyLens mitigates this with structural separation but cannot eliminate the fundamental limitation. For critical outputs, cross-check with a second AI or known-good reference.

See CONTRIBUTING.md. Any structural change requires SURVEY + BRAINSTORM + EXTENDED + EVOLVE + Kill Condition + unanimous consent.

Hack the planet. 🌎💻

AGPL-3.0