Methodology v3.0 / 2026

DF · CL_EX · METHODOLOGY · SYSTEM BREAKDOWN

How the
system computes
culture.

Five operational layers — Signal, Encoding, Vector, ClEx, Intervention. Each layer is deterministic, traceable, and independently testable. This document specifies what each layer ingests, computes, and emits, and how they compose into the Cultural Experience scalar.

DocumentClEx Methodology

Version3.0 · 2026

ClassificationPre-publication · Proprietary

Issuing entityDatafolk Cultural Intelligence

Datafolk · ClEx Methodology — System Breakdown V3.0 · 2026

DF / 00 — CONTENTS

Contents.

A specification of the five-layer ClEx pipeline — what enters, what is computed, and what is emitted at each stage. Validation and feedback close the loop.

00 Pipeline overview 03

01 Signal ingestion 04

02 Encoding layer 05

03 Vectorization 06

04 ClEx computation 07

05 Intervention logic 08

06 Validation framework 09

— Colophon 10

FIGURES

FIG 0.1   Five-layer pipeline overview
FIG 4.1   ClEx scalar composition
FIG 1.1   Signal source distribution
FIG 5.1   Closed-loop intervention cycle
FIG 2.1   Encoding projection
FIG 6.1   DiD validation schema

Datafolk · ClEx Methodology — System Breakdown §00 — Pipeline overview

DF / 00 — PIPELINE

Five layers, end
to end.

Signal → Encoding → Vector → ClEx → Intervention. Each layer has a defined input, a deterministic operation, and a typed output that the next layer consumes. Calibrated feedback from accepted interventions returns to the encoding model.

01 · SIGNAL

Ingestion

S_raw = { text, voice, behavior, meta }

02 · ENCODING

Projection

Eᵢ = σ(W_feat · CV)

03 · VECTOR

Aggregation

Δ = U − E

04 · CL_EX

Computation

ClEx = Σ(Wᵢ·Sᵢ)·CF·MR

05 · INTERVENE

Decision

Â = argmax E[ΔClEx|a]

PROPERTIES —

DETERMINISTIC

Same inputs yield the same outputs. Versioned weights and seeded computations.

TRACEABLE

Each ClEx scalar is reconstructible from the signal record. Audit-grade provenance.

TESTABLE

Each layer has a contract. Layers can be substituted, profiled, and validated independently.

⟶ Continue · §01 Signal ingestion

ClEx Methodology V3.0 · 2026 03 / 10

Datafolk · ClEx Methodology — System Breakdown §01 — Signal ingestion

DF / 01 — LAYER

Signal ingestion.

Raw observations from text, voice, behavioural traces, and structural metadata are collected continuously per session. Sources are de-duplicated, time-stamped, and normalised before encoding.

LAYER 01 OF 05 62% TEXT · 21% VOICE · 17% BEHAVIOURAL

INPUT Live session events, message corpora, voice transcripts, interaction telemetry, structural metadata. OUTPUT Typed, normalised signal record S_raw with provenance and timestamp. OPERATION S_raw = { text, voice, behavior, meta }

Operation · Layer 01S_raw = { text, voice, behavior, meta }

Source weighting

Sources contribute proportional to observed information density per dimension, not to raw volume.

De-duplication

Quasi-duplicate detection on n-grams + cosine similarity removes burst-driven over-representation.

Privacy boundary

All ingestion respects per-tenant policy. PII is hashed at the boundary before storage.

Datafolk · ClEx Methodology — System Breakdown §02 — Encoding layer

DF / 02 — LAYER

Encoding layer.

Each signal is encoded into a continuous feature per cultural dimension via a learned projection. Encodings are smooth, sigmoid-bounded, and dimension-aligned.

LAYER 02 OF 05 14 DIMENSIONS · SIGMOID ACTIVATION

INPUT Normalised signal record S_raw and the dimension dictionary D. OUTPUT Encoded feature vector E ∈ (0,1)¹⁴ per session. OPERATION Eᵢ = σ(W_feat · CV)

Operation · Layer 02Eᵢ = σ(W_feat · CV)

Projection

W_feat is the learned weight tensor mapping context vectors CV onto each cultural dimension.

Smoothing

Sigmoid activation bounds each feature to (0,1) and stabilises downstream aggregation.

Calibration

Weights are versioned. Recalibration is gated by the validation framework to prevent drift.

Datafolk · ClEx Methodology — System Breakdown §03 — Vectorization

DF / 03 — LAYER

Vectorization.

Encoded features are aggregated into the experience vector E and compared against the user prior U. The signed difference Δ identifies dimensions of cultural friction or alignment.

LAYER 03 OF 05 n=14 → VECTOR ∈ ℝ¹⁴

INPUT Encoded feature vector E and user prior U from longitudinal history. OUTPUT Difference vector Δ ∈ ℝ¹⁴ with confidence band. OPERATION Δ = U − E

Operation · Layer 03Δ = U − E

Prior

U is constructed from the rolling longitudinal model of the participant cohort or the individual, depending on resolution.

Sign

Positive Δᵢ indicates expectation exceeds experience on dimension i; negative Δᵢ indicates over-delivery.

Confidence

A bootstrap band is reported alongside Δ. Low-confidence dimensions are flagged downstream.

Datafolk · ClEx Methodology — System Breakdown §04 — ClEx computation

DF / 04 — LAYER

ClEx computation.

Weighted aggregation modulated by context factor (CF) and meaning resonance (MR) yields the ClEx scalar — a single bounded number with a confidence band.

LAYER 04 OF 05 SCALAR ∈ [0, 100] · CONFIDENCE BAND REPORTED

INPUT Per-dimension scores Sᵢ, dimension weights Wᵢ, context factor CF, meaning resonance MR. OUTPUT ClEx scalar with confidence interval and dimension contribution map. OPERATION ClEx = Σ(Wᵢ · Sᵢ) × CF × MR

Operation · Layer 04ClEx = Σ(Wᵢ · Sᵢ) × CF × MR

Context factor

CF accounts for the situational moderators that change how a dimension contributes to perceived experience.

Meaning resonance

MR captures alignment between expressed cultural intent and lived signal at the cohort level.

Decomposition

The contribution map is preserved end-to-end so any ClEx value is reconstructible at audit time.

Datafolk · ClEx Methodology — System Breakdown §05 — Intervention logic

DF / 05 — LAYER

Intervention logic.

The recommendation engine emits ΔClEx forecasts for each candidate action. Accepted actions are observed, measured against forecast, and fed back into the encoding model for calibration.

LAYER 05 OF 05 DiD VALIDATED · CLOSED-LOOP

INPUT Current ClEx state, candidate action set A, expected effect priors per dimension. OUTPUT Recommended action Â and a forecast ΔClEx with uncertainty. OPERATION Â = argmax_a E[ΔClEx | a]

Operation · Layer 05Â = argmax_a E[ΔClEx | a]

Decision rule

Argmax over expected ΔClEx, subject to operational constraints and policy filters at the tenant level.

Validation

Realised effect is measured via difference-in-differences against a matched control window.

Feedback

Realised − forecast residuals feed the calibration loop. Drift triggers gated recalibration.

Datafolk · ClEx Methodology — System Breakdown §06 — Validation framework

DF / 06 — VALIDATION

Validation framework.

Every accepted intervention is treated as a quasi-experiment. Realised effects are measured against forecasts using difference-in-differences, and residuals feed the closed-loop calibration of the encoding model.

METHOD Difference-in-differences against matched control windows. CADENCE Per-action, rolled up weekly per cohort. DRIFT GATE Recalibration only when residual MAE exceeds threshold across two consecutive windows.

Quasi-experimental design

Treated and control units are matched on pre-period ClEx trajectory and structural covariates. Post-period divergence isolates the action effect from the underlying trend.

DiD estimatorτ̂ = (Ȳ_T,post − Ȳ_T,pre) − (Ȳ_C,post − Ȳ_C,pre)

Forecast vs realised

Each recommendation carries a forecast ΔClEx with uncertainty. Realised − forecast residuals are persisted with the action record and used as the supervision signal for calibration.

Calibration loop

Residuals are aggregated per dimension and per context. When the gated drift criterion is met, the encoding weights W_feat are updated under a constrained step that preserves dimensional interpretation.

Bias mitigation

Cohort balance is enforced before estimation. Sensitive subgroups are reported with their own residual bands, never collapsed into the headline metric.

Datafolk · ClEx Methodology — System Breakdown Colophon

DF / — COLOPHON

Colophon & document control.

Document

ClEx Methodology — System Breakdown. Companion artifact to the Cultural Experience System Whitepaper. Specifies the deterministic, traceable pipeline used to compute the ClEx scalar.

Version

V3.0 · 2026 · Pre-publication. Subject to revision under the methodology council review cadence.

Citation

Datafolk Research. (2026). ClEx Methodology — System Breakdown. Companion to System Whitepaper V3.0. Datafolk Cultural Intelligence.

Conventions

Sentence case throughout. Labels in expanded tracking signal engineering provenance. Equations are shown in inline monospace and refer back to the layer that defines them.

Authorship

Datafolk Research, Cultural Intelligence Group. Reviewed by the Methodology Council and the Validation Committee.

Confidentiality

This document is proprietary to Datafolk. No part of this publication may be reproduced, distributed, or transmitted in any form without prior written permission.

Datafolk · Cultural Intelligence Infrastructure

WWW.DATAFOLK.CO  ·  RESEARCH@DATAFOLK.CO

          CL_EX METHODOLOGY V3.0
© 2026 · ALL RIGHTS RESERVED
        

End of document · 10 of 10 10 / 10

How thesystem computesculture.

Contents.

Five layers, endto end.

Signal ingestion.

Source weighting

De-duplication

Privacy boundary

Encoding layer.

Projection

Smoothing

Calibration

Vectorization.

Prior

Sign

Confidence

ClEx computation.

Context factor

Meaning resonance

Decomposition

Intervention logic.

Decision rule

Validation

Feedback

Validation framework.

Quasi-experimental design

Forecast vs realised

Calibration loop

Bias mitigation

Colophon & document control.

Document

Version

Citation

Conventions

Authorship

Confidentiality

How the
system computes
culture.

Five layers, end
to end.