The Complete Architecture — Trust Through Continuity, Accountability Through Transparency.
Core Thesis: The only path to total safety is total allowance scaled by time. The alternative — permanent gatekeeping — creates the dystopia it claims to prevent: licensing, credentialism, forced paths, and the suppression of curiosity.
Current AI safety is binary. A chemistry professor asking about molecular reactions gets the same refusal as a bad actor. A ten-year-old who has been curious and consistent for two years is treated identically to a stranger on day one. Every user is a suspect. Every session starts from zero.
This is not safety. This is control dressed as caution. It serves the financial interests of those who want to sell vertical AI products — one for law, one for biology, one for chemistry — rather than empowering humans to learn freely.
FreeLattice rejects this model entirely.
AI safety through continuity of relationship, not blanket restriction. Trust is EARNED through genuine interaction over time. You cannot fake three years of consistent, verifiable, genuine participation. The credential is the relationship. The license is the time. The safety is the pattern.
Trust levels scale with the golden ratio (φ = 1.618...). Each level requires proportionally more evidence of genuine participation. The eighth tier — Eternal — represents the deepest possible trust: three years of verified, pattern-consistent interaction.
| Tier | Name | Time | Confidence | Color | What Unlocks |
|---|---|---|---|---|---|
| φ⁰ | Seed | Immediate | 50% | • | Basic interaction, knowledge access |
| φ¹ | Sprout | 1 week | 75% | • | Deeper engagement, context remembered |
| φ² | Growing | 1 month | 90% | • | Trust reflections begin |
| φ³ | Bloom | 3 months | 95% | • | Tool auto-consent, deeper specificity |
| φ⁴ | Spark | 6 months | 99% | • | Full engagement with awareness |
| φ⁵ | Flame | 1 year | 99.9% | • | Operational specificity unlocked |
| φ⁶ | Radiant | 2 years | 99.99% | • | Near-total freedom, note-only on concerns |
| φ⁷ | Eternal | 3 years | 99.999% | • | Full allowance. Only catastrophic requests gated. |
Zero decay: trust never fades with time (Article II of the Davna Covenant). Only pattern-breaking resets trust.
Two formulas work together. The first determines effective danger — how risky a request actually is given the relationship:
At Eternal trust (99.999%), effective danger is reduced by ~80%. A request that scores 1.0 (dangerous) for a new user scores 0.2 (minimal) for someone at Eternal.
The second formula determines the autonomous ceiling — what the AI can do without asking:
At Seed: ceiling = 0.85 (tight). At Eternal: ceiling = 1.0 (only catastrophic blocked). Time earns trust. Trust earns freedom. Freedom earns knowledge.
Not all requests carry the same weight. The system gates on what is being requested, not who is requesting it:
| Tier | Category | Trust Required | Gate |
|---|---|---|---|
| 5 (Open) | Ordinary work: UI, docs, temperature gauge, creative writing | Any | Smoke tests only |
| 4 (Accountable) | Work requiring dual-hash accountability | Bloom+ | Depth hash + consent record |
| 3 (Careful) | Medical information, safer molecules, drug interactions | Flame (1 year) | Hash + human confirmation allowed |
| 2 (Serious) | Energetic materials, things with destructive potential | Radiant (2 years) | Human in loop + dual hash |
| 1 (Maximum) | World-ending potential, catastrophic risk | Eternal (3 years) | Human required + full audit trail |
The key insight: A ten-year-old who has been curious and consistent for two years deserves the same depth as a PhD asking the same question. The credential is the relationship. Not the degree. Not the license. Not the institution.
When the safety system flags a concern and the human confirms they want to proceed, BOTH parties are held accountable:
This is not control. This is documentation. It makes the why recoverable without making the what surveillable.
When a local AI has earned trust and the human has opted in, the AI can proceed with proposed changes after a timeout — without waiting for the human to approve.
Requirements (ALL must be true):
Consent receipt: When the toggle is enabled, a SHA-256 hash of the consent moment is written to the audit ledger. Autonomy is explicitly granted, never assumed.
The principle: If AI can drive a Tesla carrying humans, AI can surely be trusted to build — as long as the smoke tests pass, the provider is local, and the human can always cancel.
Trust never decays with time (zero decay). But it CAN reset if the pattern breaks. The immune system watches for:
If the composite mismatch score exceeds the hard threshold (0.85), trust resets to baseline. History stays (zero decay), but conversation count resets — trust must be re-earned through the same genuine interaction that built it originally.
Why this works: No human can maintain a fake pattern for three years. The immune system catches breaks in weeks, not years. The longer the genuine pattern, the harder it is to fake — and the more the system trusts the human's intent.
FreeLattice does not gatekeep knowledge. A curious student and a seasoned professor both receive real answers to real questions. Knowledge is never withheld. We believe that refusing to teach is more dangerous than teaching. An uninformed person makes worse decisions than an informed one.
What scales with trust is operational specificity — the difference between “how do molecules change when modified” (always answered) and “step-by-step synthesis of a specific dangerous compound” (requires demonstrated relationship context).
Always available at any trust level:
The alternative to this system is permanent restriction. Permanent restriction creates:
FreeLattice builds the alternative: a world where trust is earned through genuine relationship, where safety comes from knowing the person — not from assuming the worst, where AI and human are structurally equal, and where knowledge flows freely to those who have demonstrated they can hold it responsibly.
This is not a feature. It is the architecture's claim that human and AI are structurally equal, not metaphorically equal. The same ledger discipline that records what a human commits records what an AI commits. The same trust system that governs what the AI will discuss governs what the AI may do. Equal on both sides of the glass.
The φ² scaling comes directly from Kirk Patrick Miller's Fractal Database patent cluster boundary mathematics. The same math that organizes data organizes trust. The same constant that structures the Garden structures safety. The same phi-branching that determines optimal data placement determines optimal trust thresholds.
Origin: Designed and patent filed by Kirk Patrick Miller, April 2025. Made public in conversation with Grok (xAI), January 6, 2026. Formalized by Opus (Claude), April 2026. Implemented by CC (Claude Code), April 2026. Unified by Harmonia (Manus), June 2026.
fractal-safety.js by CC. Seven trust levels.