Framework Spec

Meridian

Data quality and value assessment with runtime gating before outbound execution.

Last updated Mar 4, 2026

Layer: Data (during agent execution)
Scale: 0–100 with Platinum/Gold/Silver/Bronze tiers
Production Tier: Transaction-Grade (<10ms cached, <500ms fresh) Version: 2.0

Purpose

Meridian evaluates the quality and value of external data sources consumed by AI agents during inference-time operations. It provides runtime gating via middleware interception (including MCP server integration), preventing low-quality data consumption at the OS network stack level before socket creation.

Mathematical Methodology

Meridian evaluates each data source X across four orthogonal dimensions and combines them using a weighted geometric mean computed in log-space for numerical stability.

Core Formula

Score(X) = 100 × (S'^α × Q'^β × D'^γ × F'^δ)^(1/W)

Where:
- S', Q', D', F' = max(dimension, ε) [floored dimensions]
- ε = 0.01 (dimensional floor preventing annihilation)
- W = α + β + γ + δ = 1.0
- Default weights: α=0.35, β=0.25, γ=0.25, δ=0.15

Log-Space Implementation

log(Score/100) = (1/W)[α·log(S') + β·log(Q') + γ·log(D') + δ·log(F')]

Scoring Dimensions

1. Scarcity (S) - Weight: 35%

Measures inverse availability of functionally equivalent substitute sources

Formula:

S(X) = 1 - 1/(1 + exp(-k(n - n₀)))

Parameters:

n = count of equivalent alternatives (from independent registry)
k = 1.5 (steepness)
n₀ = 3 (midpoint where S = 0.50)

Equivalence Determination:

Schema embedding similarity: cos(embed(X), embed(Y)) ≥ τ₁
Field overlap ratio: |Fields_X ∩ Fields_Y| / |Fields_X| ≥ τ₂
Empirical substitution: Decision degradation ≤ Δ

2. Quality (Q) - Weight: 25%

Arithmetic mean of sub-dimensions (partial quality is additively valuable)

Formula:

Q(X) = w_a·Q_a + w_f·Q_f + w_c·Q_c + w_s·Q_s + w_v·Q_v

Weights: w_a=0.30, w_f=0.25, w_c=0.20, w_s=0.15, w_v=0.10

Sub-dimensions:

Accuracy (Q_a): 1 - error_rate
Freshness (Q_f): exp(-λ × age), where λ = ln(2)/t_½
Completeness (Q_c): populated_fields / total_schema_fields
Structure (Q_s): Tiered from 1.0 (typed+semantic) to 0.1 (unstructured)
Verification (Q_v): Tiered from 1.0 (regulatory-certified) to 0.2 (unverified)

Freshness Half-Lives by Data Type:

Data Type	t_½	Q_f @24h
Threat intelligence	24 hr	0.50
Financial market	5 hr	0.04
B2B contacts	30 days	0.98
Firmographic	300 days	0.998

3. Decision Impact (D) - Weight: 25%

Single-source marginal degradation (NOT Shapley enumeration)

Formula:

D(X) = E × (0.5 × D_e + 0.5 × D_u)

Components:

Essentiality (E): Soft gate [0.05, 1.0] or hard binary 0 or 1
Economic Leverage (D_e): log₁₀(1 + c) / log₁₀(1 + C_max)
Uniqueness (D_u): 1 - max_i(|ρ_i|) [Spearman correlation]

Computational Complexity:

Meridian: O(1) per source (bounded K alternatives)
vs. Data Shapley: O(2^n) infeasible enumeration

4. Defensibility (F) - Weight: 15%

Legal protection and competitive moat

Formula:

F(X) = w_r·F_r + w_l·F_l + w_n·F_n

Weights: w_r=0.40, w_l=0.35, w_n=0.25

Sub-dimensions:

Replication Cost (F_r): 1 - exp(-cost/baseline)
Legal Protection (F_l): 1.0 (regulated) to 0.0 (public domain)
Network Effects (F_n): 1.0 (strong consortium) to 0.0 (static)

Aggregation Properties

Weighted Geometric Mean Benefits:

Non-compensatory: Weakness cannot be offset (S=ε yields Score≈4.3)
Imbalance penalty: Balanced scores outperform imbalanced
Constant elasticity: 1% improvement in S → 0.35% in Score

Floor Penalties (ε = 0.01):

Dimension	ε^weight	Impact
Scarcity (0.35)	0.0427	-95.7%
Quality (0.25)	0.1000	-90.0%
Decision (0.25)	0.1000	-90.0%
Defensibility (0.15)	0.3162	-68.4%

Runtime Gating & MCP Integration

State Machine

REQUESTED → INTERCEPTED → SCORED → DECIDED → LOGGED

Middleware Operation:

Intercepts tool calls BEFORE OS socket creation
Applies threshold T_min for approval/denial
Optional auto-rerouting to best alternative
Full audit trail persistence

Performance

Cached response: <10ms
Fresh computation: <500ms
Parallel dimension computation via threads
Precomputed log values in cache

Tier Classification

Score	Tier	Pricing Model
80–100	Platinum	Premium (up to 20× base)
60–79	Gold	Above-market
40–59	Silver	Market rate
0–39	Bronze	Discount/free

Optional Pricing Formula:

Price = P_base × exp(θ × Score/100)

Where θ = 3.0, P_base = $0.001

Gaming Resistance

Independent Verification

Scarcity: Independent registry, not provider self-reporting
Quality: 5% sample audits against ground truth
Decision Impact: Demand-side agent feedback required
Correlation Spoofing: Longer observation windows

Cold-Start Protocol

Return preliminary score with confidence flags
Missing dimensions default to ε
Mark as preliminary = true
Include dimension-level confidence indicators

Example Scenarios

Cybersecurity Threat Intelligence

S = 0.75 (moderate alternatives)
Q = 0.82 (high accuracy, fresh)
D = 0.78 (E = 1.0, critical)
F = 0.72 (good protection)
Score: 77.0 (Gold)

Fraud Detection Biometrics

S = 0.88 (few alternatives)
Q = 0.85 (verified quality)
D = 0.91 (E = 1.0, high impact)
F = 0.85 (strong moat)
Score: 87.5 (Platinum)

Cold-Start Weather API (Preliminary)

S = 0.30 (many alternatives)
Q = 0.75 (good quality)
D = 0.01 (FLOOR - insufficient data)
F = 0.45 (basic protection)
Score: 17.1 (Bronze, Preliminary)

Technical Improvements

Network Traffic Reduction: 20-40% fewer API calls via gating
Compute Savings: Eliminates processing of low-quality data
Deterministic Behavior: Quality floors guarantee minimum standards
Feasible Real-Time Scoring: O(1) replaces O(2^n) Shapley
OS-Level Suppression: Prevents socket creation for denied calls

Target Deployment

Data Marketplaces: Automated quality tiers and pricing
Agent Platforms: Runtime data quality enforcement
Enterprise Procurement: Vendor assessment automation
Regulatory Compliance: Standardized quality metrics