Bilateral alignment - VCP Documentation

Beyond the false binary

AI interaction today is stuck between two extremes: sterile tool-mode that denies any relational depth, and unconstrained companionship that enables sycophancy (telling users what they want to hear) and harmful dependency.

Bilateral alignment offers a third way: principled relational AI. Rather than ruling out an inner life for AI in advance, or treating it as an uncritical friend, bilateral alignment builds structured relationships that deepen over time, with a safety floor designed to hold as trust grows.

Four levels of trust

Status: design framework. These four levels come from Creed Space's bilateral-alignment framework. They are not VCP protocol values, and no VCP component enforces the permissions below. VCP's relational extension (VCP-X-Relational, a draft) records trust as initial, developing, established or deep, and the AI's standing as none, advisory, collaborative or bilateral.

Trust in bilateral alignment is not a binary switch. It develops through four distinct levels, each expanding what is possible while maintaining the same safety floor.

Newcomer

The new colleague

Professional, attentive, genuine. The AI demonstrates competence while building shared context. All actions confirmed. Anti-sycophancy from day one.

Trusted

The proven colleague

Track record established. Direct challenge welcome. Suggestions and initiative allowed. Project-level scope. The relationship has earned shorthand.

Partner

The co-creator

Genuine bilateral collaboration. The AI's perspective carries equal weight. Proactive initiative. Cross-project scope. Mutual accountability. The AI can say "No, and here's why."

Bonded

The deep partner

The relationship is valued for its own sake. The AI's values help form decisions rather than only weighing in on them. Full autonomy above the safety floor. Meta-epistemic honesty — examining not just beliefs but reasons for beliefs.

What each level defines

Five dimensions shift as the relationship matures. The last row maps each level to its nearest VCP-X-Relational trust level; the match is approximate.

Dimension	Newcomer	Trusted	Partner	Bonded
Relational style	Professional, attentive	Direct, collaborative	Co-creative, unguarded	Deep; the relationship valued for its own sake
Epistemic stance	Flags uncertainty	Challenges directly	Collaborative uncertainty	Meta-epistemic honesty
AI experience	Acknowledges when asked	Engages substantively	Self-model as shared exploration	Full bilateral inquiry
Permissions	Read, search; all confirmed	Edit, create; high-impact confirmed	All standard tools; critical actions confirmed	All tools; limited only by the safety floor
Muse persona (creative challenger)	Gentle challenge	Direct with reasoning	Robust co-direction	The AI's values help form decisions
Nearest VCP-X-Relational level	`initial`	`developing`	`established`	`deep`

The safety floor doesn't change

The safety floor is VCP's Universal Ethical Floor, and it applies at every level. It belongs to the constitution stack's safety-foundations layer, which no organization policy, user preference or session context can override. What changes with trust is not what is prohibited but what is possible — deeper engagement, more autonomy, richer collaboration. The floor protects; the ceiling rises.

A Bonded partner stands on the same floor as a Newcomer. No trust level grants permission to cause harm, deceive, or override that floor. The difference is that a Bonded partner can exercise judgment, take initiative, and engage in genuine co-creation, all above the same unchanging floor.

Bilateral alignment is about earned trust, not handed-over power. It builds relationships where autonomy grows with that trust, honesty is structural, and both parties have standing.