Research Contributors — Computational Validation of Regulatory States in Conflict Narratives

AI Safety

Emotional gradients for safer models, safer decisions, and fewer false positives

Most AI safety systems still rely on a binary collapse: safe vs unsafe, healthy vs risky, acceptable vs harmful. Human emotional reality does not operate in binaries. It operates on gradients, and those gradients change with nervous system state.

TEG-Blue is a structured, trauma-informed framework that makes those gradients visible, operational, and computationally legible, so AI systems can reason about emotional content with more accuracy and less harm.

Why binary safety fails in real life language

A single sentence can represent multiple realities.

Example:

“I can’t do this anymore.”

A binary classifier tends to treat that as one thing. A gradient approach can distinguish at least four different patterns:

• Connection: a boundary, a clean exit, a moment of truth, a repair attempt
• Protection: overwhelm, shutdown, grief, temporary distress needing support
• Control: coercive framing, threat of abandonment, emotional leverage
• Domination: intimidation, punishment language, containment of the other

Without gradients, systems either:

• over-flag healthy boundaries as “risk”
• miss manipulation that hides behind soft language
• treat trauma signals as toxicity
• turn complex emotional language into policy errors

What TEG-Blue contributes to AI safety

TEG-Blue provides a gradient-based measurement layer and a theoretical architecture for why these patterns exist and how they shift under stress.

1) A measurement model that is not personality-based

The Four-Mode Gradient maps observable regulatory patterns: Connection, Protection, Control, Domination. These are states, not types.

2) Structured schemas for machine use

A proposed schema layer translates emotional pattern logic into formats AI systems can consume, including gradient classifications and consistent terminology.

3) Safety that does not punish the user for being human

The goal is fewer false positives, fewer harmful interventions, and better separation of:

• distress vs threat
• boundary vs hostility
• trauma language vs manipulation language
• disagreement vs abuse

Where this applies

• Content moderation: reduce false positives for vulnerable users, detect coercive patterns more accurately
• Mental health adjacent products: avoid collapsing grief, shutdown, and boundaries into “risk”
• Assistant behavior and refusals: safer decisions when the user is dysregulated without becoming punitive or patronizing
• Misuse prevention: detect patterns of coercion, grooming, and domination that often remain “polite” on the surface
• Evaluation: test models on gradient sensitivity instead of binary pass fail

What we are building next

We are developing:

• annotation protocols for the Four-Mode Gradient in natural language
• reliability and validity studies
• evaluation sets focused on ambiguity and high-risk misclassification cases
• machine-readable schema drafts and misuse checks

Collaborate

We welcome collaboration from AI safety researchers, NLP researchers, alignment teams, and evaluators who want to work on bounded tasks.

Useful profiles:

• NLP and evaluation researchers
• AI safety and policy folks working on real-world harms
• researchers in emotion regulation, trauma, and interpersonal dynamics
• red-teamers focused on coercion and manipulation patterns

If you want to contribute, start here: teg-blue.org/start-here, then see teg-blue.org/collaborate.

Research principle

We do not treat emotions as noise.

We treat them as data about safety, threat, belonging, and meaning, and we build tools that can handle that data without flattening it into binaries.