Workshop

Where audience intelligence multiplies.

Workshop connects governed audience sources, guides analysis with Dora, and turns every tabulation, model, and visualization into a living dossier.

Request a Demo See how it works

Workbook

Base query

Joins

Composites

Models

Q5 scoring model

Model file Audience propensity score

Summary Recipe Script

Predictive score for Q5

Estimates the likelihood that an audience member belongs in the target response group, using matched survey and consumer variables.

Training data

Matched survey rows and consumer variables.

3.7k

Outcome

Binary response class from Q5.

Ready

Validation

Best fold mean squared error.

0.076

Top predictors

10.18

20.14

30.11

Status Trained / Review Fold 3

Dora

Can we turn the new consumer variables into a Q5 audience propensity model?

Yes. I matched them to 3.7k survey rows, held out validation, and ranked the strongest predictors.

Audience propensity model is trained.

Output field: audience_propensity_q5. The recipe, MSE notes, and staged score are written to the workbook dossier.

Review model dossier

How a workbook works

Connect sources. Ask Dora. Build the dossier as you work.

A workbook is not a spreadsheet or a dashboard. It is a compounding intelligence environment where every source, model, and analysis becomes part of the knowledge layer.

01 Increase the surface area for insight. Join surveys, voter files, consumer data, behavioral signals, and modeled scores so each source can strengthen the next. Workshop helps compute new variables, generate model-driven variables, and track every merge and query back to validated source context.

Workbook 1278 Audience build sequence

Workbook Project

Source Data

Full Dataset 121 / 1.8k

Market Survey 45 / 1.8k

Consumer Data 61 / 1.8k

Computed Fields 15 / 1.8k

Snapshots

Persuadables 121 / 480

Female 18-34 121 / 311

Joined Consumer Data

Dataset Tabulations Agent Dossier

Variable profiles Value profiles

1

Household Composition Family structure appended from consumer data.

SingleFamilyMulti

2

Children in Household Count of children in the home.

0123+

3

Veteran in Household Flag used for audience segmentation.

YesNo

4

Home Ownership Owner/renter signal for household models.

OwnRent

5

Estimated Income Band Modeled household income range.

<50k50-100k100k+

6

Purchase Category Category affinity from acquired behavior.

HomeAutoRetail

7

Length of Residence Tenure signal used in response models.

0-2y3-7y8+y

Respondent ID HH composition Children Income band Score

R-10482 Family 2 100-150k 0.74

R-10917 Single 0 50-75k 0.58

R-11306 Multi-gen 1 75-100k 0.81

R-11844 Couple 0 150k+ 0.66

R-12039 Family 3+ 75-100k 0.72

R-12671 Single 0 <50k 0.49

R-13128 Family 1 100-150k 0.69

R-13652 Couple 0 50-75k 0.52

R-14086 Family 2 150k+ 0.79

Join query Scrolled to fields 55-61 of 61

-- 54 selected fields above
  consumer.home_ownership_status,
  consumer.length_of_residence,
  consumer.household_composition,
  consumer.children_in_household,
  consumer.veteran_in_household,
  consumer.income_band,
  consumer.consumer_match_score
FROM market_survey AS survey
LEFT JOIN consumer_database AS consumer
  ON survey.respondent_unique_id = consumer.respondent_unique_id;

Search 121 vars

Dora

Children CHILDREN_IN_HH

Veteran VETERAN_HH

Income INCOME_BAND

Residence LENGTH_RESIDENCE

☝️ I found 61 consumer fields that can help deepen the intelligence in your survey and be used for modeling downstream. Would you like me to join this data to your survey?

go ahead. JG

Done. Added 61 consumer fields, computed 15 modeling fields, and saved two audience snapshots.

The truth is in here... ↑

Custom variable builder

Recipe Preview

Inputs Turn joined fields into analysis-ready signals.

children in household estimated income purchase category digital engagement

household_opportunity = income_band + family_stage + engagement_score

Generated variables 3 variables ready for review

household_opportunity price_sensitivity_band engagement_depth_score

Snapshot workspace

Frame Compare

Locked snapshot High-propensity audience

1,750 records 132 variables 4 joined sources

Delta readout What changed against base?

+12Income

+8Intent

-5Age

Predictive model workspace

Model Scoring

Best MSE Q5 audience propensity

0.076

Training rows: 3.7k Outcome: Q5 response class Status: trained / review fold 3

Model logic Top predictors and scoring output

household_opportunity engagement_depth_score consumer purchase category Output: audience_propensity_q5

02 Generate Intelligence for Humans and Agents. Turn workbook data into weighted tabulations, readable findings, and structured dossier artifacts that people can inspect and agents can reuse downstream.

Workbook Project

Source Data

Full Dataset 121 / 1.8k

Market Survey 45 / 1.8k

Consumer Data 61 / 1.8k

Computed Fields 15 / 1.8k

Snapshots

Persuadables 121 / 480

Female 18-34 121 / 311

Persuadables

Dataset Tabulations Agent Dossier

Q1

Not at all 15% (72)

Not very 25% (118)

Somewhat 36% (176)

Very likely 24% (114)

Q2

Yes 63% (302)

No 37% (178)

Q3

Strongly agree 20% (96)

Agree 35% (168)

Disagree 30% (142)

Strongly disagree 15% (74)

Q4

Low 21% (101)

Medium 39% (187)

High 40% (192)

Q5

Strongly agree 17% (82)

Agree 33% (158)

Disagree 30% (144)

Strongly disagree 20% (96)

Total 480

Gender

Male 235

Female 245

Age

18-34 160

35-54 170

55+ 150

HH Composition

Single 150

Family 230

Multi-gen 100

Gender nested within age n480

18-34

Female n42

Male n38

35-54

Female n56

Male n51

55+

Female n28

Male low n

Single Family n96

Female n31

Male n29

Search

Dora

Dora, I want to understand gender differences across segments like age and household composition, but I’m worried some cells may be too small for a reliable read. JG

Yes. Some three-way cells will get thin if we keep the raw response scales.

Create composites Convert bipolar questions into Top 2 / Bottom 2 box variables.

Run nested tabs Compare gender inside age and household-composition segments, with low-n cells flagged.

That sounds great. Go ahead. JG

Done. The nested tables and charts are ready for review.

Review output Three-way views are available in the workbook.

Agent dossier updated three-way-tabs.md now includes JSON plus the mechanical readout for future agent access.

The truth is in here... ↑

03 Build predictive models you can inspect and deploy. Train models from workbook snapshots, validate the recipe, generate scoring scripts, and deploy model-driven variables back into the audience intelligence layer.

Workbook Project

Trained Models

Q5 Propensity 0.076

Persuadable Likelihood 0.081

High-Intent Likelihood 0.092

Response Lift Model 0.104

Activation Readiness 0.118

Segment Expansion draft

Q5 Propensity Model

Model Deployments

0.076 Good best MSE

Q5 propensity model

Ranks audience members by likelihood of landing in the Q5 target group. MSE measures prediction error, so lower is better; this run is good for a 0-1 target, with one fold flagged for review.

Trained Review fold 3 Score field: audience_propensity_q5

How to read it Two folds are good; one fold needs inspection.

0.081 fold 10.076 best0.149 review

Model build Trained on a locked audience snapshot.

Training rows3.7k OutcomeQ5 class Avg. error~0.28

Signals Strongest predictors remain inspectable.

Engagement0.18

Household0.14

Consumer0.11

Deployment path The model becomes a reusable audience variable.

Workbook score fieldready

Model dossier + recipeupdated

Gateway packageafter review

Search

Dora

Dora, use the Persuadables snapshot to build a Q5 propensity model we can understand and deploy. JG

I can do that. I’ll train against the Q5 target class, compare MSE across folds, and keep the deployment path visible.

Run it. JG

Done. The best fold is in a good range, but one validation fold is higher and should be reviewed before activation.

Score field created audience_propensity_q5 has been written back to the workbook.

Deployment staged The score and model dossier are ready; Gateway deployment waits on fold review.

The truth is in here... ↑

THE DOSSIER

Every analysis leaves behind structured intelligence.

Every time you run an analysis, Workshop produces two things: the visualization you see, and a structured artifact in the dossier. The dossier is the machine-readable record of everything the workbook contains — every variable, every cross-tab, every model output, with provenance and validation metadata. It grows continuously as you work.

Workbook Dossier File structure

manifest.md

Dossier identity, validation status, and snapshot frame metadata — population, filter set, weight, lock timestamp. First file a consuming AI reads.

dataset-context.md

Source origin and provenance, data creation method, and interpretive context. Updated as the workbook evolves.

variables.md

One entry per variable — species monograph and specimen record. Grows as new variables, joined sources, and modeled scores are added.

one-way-tabs.md

Frequency distributions for every variable. JSON and mechanical readout per entry. Grows as variables are added.

two-way-tabs.md

All pairwise cross-tabulations. JSON and mechanical readout per entry. Canonical direction fixed. New pairs computed against all existing variables when a new one is added.

custom-tables.md Workbook only

User-directed analyses — three-way tabs, bespoke cross-tabulations, custom computations. JSON and readout per entry. Validated before entry.

Gateway access

Let external AI systems reason from the workbook.

Every workbook dossier is accessible through the Connectivity Gateway, initially via MCP, with API and A2A access coming. Claude, ChatGPT Enterprise, Copilot, Glean, and custom agents can use classified variables, weighted cross-tabulations, model outputs, provenance, and interpretive context without touching the raw data.

Workbook dossier Validated audience intelligence

AI work systems Agents, tools, apps, workflows

MCP API coming A2A coming Claude ChatGPT Enterprise Copilot Glean

Build intelligence that keeps getting smarter.

Bring sources, models, and questions into Workshop so every analysis becomes reusable knowledge for your team and your AI systems.

Request a Demo