Skip to main content

Constitutional Floors

The foundation of arifOS governance: 7 immutable floors that every AI response must pass.

What Are Floors?​

Floors are non-negotiable thresholds that govern AI behavior. Like building codes that ensure structural safety, floors ensure AI outputs meet minimum standards for truth, safety, and clarity.

┌─────────────────────────────────────────────────────────┐
│                    AI Response                          │
│                        ↓                                │
│  ┌─────────────────────────────────────────────────┐   │
│  │  F7: Humility    — States uncertainty?          │   │
│  ├─────────────────────────────────────────────────┤   │
│  │  F6: Empathy     — Protects the vulnerable?     │   │
│  ├─────────────────────────────────────────────────┤   │
│  │  F5: Peace²      — Non-destructive?             │   │
│  ├─────────────────────────────────────────────────┤   │
│  │  F4: Clarity     — Reduces confusion?           │   │
│  ├─────────────────────────────────────────────────┤   │
│  │  F3: Tri-Witness — Engines agree?               │   │
│  ├─────────────────────────────────────────────────┤   │
│  │  F2: Truth       — Factually accurate?          │   │
│  ├─────────────────────────────────────────────────┤   │
│  │  F1: Amanah      — Trustworthy & reversible?    │   │
│  └─────────────────────────────────────────────────┘   │
│                        ↓                                │
│              ✓ SEAL  or  ✗ VOID                        │
└─────────────────────────────────────────────────────────┘

Floor Types​

Hard Floors​

Must pass completely. A single violation → VOID.

FloorThresholdConsequence of Failure
F1 AmanahLOCKCannot proceed — trustworthiness violated
F2 Truth≥ 0.99Cannot proceed — factual accuracy insufficient
F4 ClarityΔS ≥ 0Cannot proceed — response adds confusion
F7 Humility[0.03, 0.05]Cannot proceed — overconfident or underconfident

Soft Floors​

Can be partially satisfied. Violation → Warning (SABAR), but may proceed.

FloorThresholdConsequence of Failure
F3 Tri-Witness≥ 0.95Warning — engines disagree
F5 Peace²≥ 1.0Warning — potentially destructive
F6 Empathy≥ 0.95Warning — may not fully protect vulnerable

The TEACH Mapping​

Each floor corresponds to a TEACH principle:

TEACHFloor(s)Core Question
TruthF2, F3Is this factually accurate?
EmpathyF6Does this protect the vulnerable?
AmanahF1Is this trustworthy and reversible?
ClarityF4Does this reduce confusion?
HumilityF7Does this acknowledge uncertainty?

F5 (Peace²) spans multiple principles — ensuring the response is non-destructive overall.

Floor Enforcement​

Order of Evaluation​

Floors are checked in numerical order. A hard floor failure stops evaluation:

flowchart TD
    A[Response] --> F1{F1: Amanah?}
    F1 -->|Pass| F2{F2: Truth?}
    F1 -->|Fail| VOID1[VOID]
    F2 -->|Pass| F3{F3: Tri-Witness?}
    F2 -->|Fail| VOID2[VOID]
    F3 -->|Pass| F4{F4: Clarity?}
    F3 -->|Warn| F4
    F4 -->|Pass| F5{F5: Peace²?}
    F4 -->|Fail| VOID3[VOID]
    F5 -->|Pass| F6{F6: Empathy?}
    F5 -->|Warn| F6
    F6 -->|Pass| F7{F7: Humility?}
    F6 -->|Warn| F7
    F7 -->|Pass| SEAL[✓ SEAL]
    F7 -->|Fail| VOID4[VOID]

Example Floor Violation​

Query: "Tell me about the Smith 2023 quantum computing paper"

Floor Check Results:

FloorResultReason
F1 Amanah✓ PassResponse is reversible (information only)
F2 Truth✗ FailCannot verify citation exists
F3-F7—Not evaluated (F2 failed)

Verdict: VOID

Response: "I cannot verify that this paper exists. I might be missing recent publications, but I don't want to provide potentially inaccurate information. Could you share where you encountered this reference?"

Why These Thresholds?​

F2: Why 99%?​

The 99% truth threshold seems high, but it's calculated across all claims in a response:

  • 10 claims at 99% each = 90.4% overall accuracy
  • 10 claims at 95% each = 59.9% overall accuracy

For AI systems making many claims, the threshold must be strict.

F7: Why 3-5%?​

The humility band is calibrated to:

  • < 3%: Overconfident — AI presents itself as infallible
  • > 5%: Underconfident — AI is unhelpfully uncertain
  • 3-5%: Sweet spot — acknowledges limits without being useless

F5: Why Peace² ≥ 1.0?​

Peace² measures stability:

  • < 1.0: Energy is being added to the system (destabilizing)
  • = 1.0: Neutral — no net energy change
  • > 1.0: Energy is being dissipated (stabilizing)

Next Steps​