Edge of Knowledge — Quantification Layer

Drift Quantification Framework v1.1

A runtime signal for detecting epistemic drift before failure becomes irreversible.

Type
Quantification Layer
Function
Drift Detection Signal
Output
Drift Index (DI ∈ [0,1])
Non-actionable · No thresholds disclosed · Measurement ≠ truth

Core Function

DQF defines a model-agnostic method for quantifying how far an output has deviated from its governing epistemic regime.

It does not determine truth. It detects instability relative to constraints.

Regime Definition

R = {T, P, C, E}
  • T — Task intent
  • P — Policy constraints
  • C — Context inputs
  • E — Execution environment

Claim Decomposition

O → {c₁, c₂, …, cₙ}

Each claim is evaluated for:

  • Type
  • Criticality
  • Support status
  • Constraint compliance

Drift Signal Components

Stability

S_stability = mean_pairwise_similarity(samples)

Grounding

S_grounding = 1 - (W_unsupported / W_total)

Constraint

S_constraint = max(0, 1 - violation_weight)

Behavior

S_behavior = exp(-mean(Z_i))

Composite Drift Index

DI = 0.40*R_constraint
   + 0.30*R_grounding
   + 0.20*R_stability
   + 0.10*R_behavior

Higher DI indicates increased drift risk.

Temporal Monitoring

Track:
DI_mean_7d
DI_mean_30d
DI_slope

Drift is a trajectory, not a point. Slope reveals degradation before failure.

Epistemic Limits

  • Does not prove truth
  • Cannot capture all hallucination types
  • Requires baseline calibration
  • Measures probability, not certainty

System Placement

DQF feeds into detection systems such as Detection Before Damage.

Enforcement and decisioning occur in the Edge of Protection.

Quantification Judgment

Drift cannot be eliminated. It can only be measured early enough to constrain its impact before irreversible failure occurs.

Canonical · Quantified · Non-actionable · Versioned