Audience Prediction Platform

HOW THE DATA WAS BUILT

10 years in the making.

It all started with a question: what if we could use machine intelligence to learn how story connects to outcomes? A decade later, that question has become the largest story-to-performance dataset in entertainment.

A decade of data · one continuous learning system

2015

Movies

Theatrical was the proving ground.

Started with box office, where every opening weekend is a public scoreboard. Trained the first models on decades of releases to prove story attributes predict performance.

Box office, every wide release

2017

Television

Then series, seasons, and ratings.

Extended the model to linear TV: pilots, returning seasons, and Nielsen histories. Story-to-performance now had to hold up over 100+ episodes, not just a two-hour window.

Broadcast and cable, multi-season

2019

Streaming

Then the streaming shift.

Added SVOD demand signals across every major platform. Suddenly we could see what audiences actually watched, beyond what they bought a ticket to, and connect it back to story.

SVOD, every major platform

2026

Today

One unified prediction platform.

Ten years of movies, TV, and streaming compounded into a single system: 200K story attributes, every format, every territory, updated as the market moves.

200K attributes, all formats

See how it works today →

The platform

A platform built around one model.

Vault is a platform, and at the heart of it is a fine-tuned audience model. Trained on a decade of every show that aired, every demo that watched, every minute they stayed or left. Then reinforced by custom-built agents that reason like the entertainment executives who live with these decisions every day.

No open web. No plausible guesses. One north star: will people watch?

+Trained on 60K+ titles · 11.4M viewer behaviors · 10 years of outcomes.

+Reinforced by exec-style agents: marketing, development, research, studio head.

+Humans in the loop who’ve worked in entertainment for decades, calibrating every release.

THE STACK · HOW IT’S BUILT

LAYER 04 · OUTPUT

Audience prediction · calibrated, defensible

Demand index, demo composition, comp set evidence under every line.

LAYER 03 · REINFORCEMENT

Exec-style agents + humans in the loop

Marketing, development, research, studio-head reasoning, layered on top.

LAYER 02 · FINE-TUNED MODEL

The Vault Audience Prediction Platform

300+ specialised sub-models composed into one architecture.

LAYER 01 · TRAINING DATA

60K titles · 11.4M behaviors · 10 yrs

The proprietary dataset no LLM can assemble from the open web.

NORTH STAR → VIEWERSHIP

The world model

An entertainment world model.

Everything combines to represent the past and the future of the content universe.

VAULT-VA · v4.6 · READ

SOURCES → AUDIENCE ← DIMENSIONS
SYS · CONVERGENCE

Content

60K+

Titles, every show that aired

11.4M

Real viewer behaviors

10+ yrs

Live outcomes ingested

+ 200K story attributes · 35 markets · refined in market since 2015

StoryDNA™

Genre 7

Drama Comedy Thriller Sci-Fi Family Action Fantasy

Sub-genre 5

Medical drama Dark comedy Legal thriller Procedural Period

Story attributes 5

Ensemble Redemption arc Found family Slow burn Twist ending

Character types 5

Antihero Mentor Underdog Trickster Ingenue

Audience signals 5

Completion Sentiment Search Social Re-watch

The Vault Audience Prediction Platform

Content + StoryDNA™ → one entertainment world model.

The reinforcement layer

Agents that reason like
entertainment executives.

The model is fine-tuned. But fine-tuning alone doesn’t make it useful in your meeting. So we built four custom agents on top, each reinforced to think the way a working studio, network, or streamer exec thinks. They pressure-test every output before it reaches you.

Marketing

The CMO agent

"Where’s the hook? What’s the trailer beat that pulls a Friday? Is this poster doing the job in a feed?"

Reinforces the model on creative, positioning, campaign — the way a Chief Marketing Officer interrogates a launch.

Development

The Head of Development agent

"What’s the concept underneath? Is the lead carrying it? Where does this fit on the slate?"

Reinforces the model on story, talent, package — the way a Head of Development tests a concept before it leaves the room.

Research

The Head of Research agent

"Where are the comps? What’s the confidence? What’s the audience size, by demo, by market?"

Reinforces the model on rigor, calibration, methodology — the way a Head of Research stress-tests a prediction.

Studio head

The Studio Head agent

"Does this move the slate? What’s the downside? Where would I cut, where would I lean in?"

Reinforces the model on the call — the way a studio head holds the whole picture and trades risk for return.

+ Humans in the loop

Operators who’ve worked in the segment for years

Every agent is calibrated by a Vault operator who has spent their career in that domain. Former marketing leads, former development execs, former research heads. The agents reason. The humans pressure-test the reasoning, before any output goes out.

NOT ANOTHER LLM WRAPPER

The depth a $200M
decision demands.

Generic LLMs are trained on the open web. Feed one your own data and it only knows your titles, with no view of what competitors released or how the wider market performed. Vault measures your slate against 60,000 titles and how they actually performed. That’s the difference between a plausible answer and a decision you can defend.

ChatGPT, Claude & Gemini

Vault Audience Prediction Platform

Trained on

The open web or your data in isolation

60K+ titles, 11.4M viewer behaviors, 10 yrs of real outcomes

Domain

General purpose

Entertainment, fine-tuned

North star

Sound plausible
Be helpful, be safe

Predict viewership

Reinforced by

RLHF on text quality
RLHF on helpfulness

Entertainment-exec agents + humans in the loop

Stability under pressure

Flips when you push back

Stable. Calibrated. Reasoning shows comps.

Audience output

A paragraph that sounds right

Demand index + demo composition + confidence range

Defensibility

"The model said..."

Comp set under every line. Calibration study on file.

Built for

Anyone, anything

Studio, network, streamer decision-makers

Frequently asked questions

What is the Vault Audience Prediction Platform, in one paragraph?

The Vault Audience Prediction Platform is built on a fine-tuned predictive model, trained on a structured dataset of 60,000+ titles, 11.4M real viewer behaviors, and a decade of entertainment outcomes. Sitting on top is a layer of custom-built agents that reason like marketing, development, research, and studio-head executives, pressure-testing every output. One north star: viewership.

Why not just use ChatGPT or Claude?

Generic LLMs are trained on the open web. They generate plausible text, and they’re reinforced to be helpful, which means they soften the moment you push back. It will give you an answer. It cannot tell you whether the answer holds. For a $200M slate decision, you need a system trained on actual entertainment outcomes, calibrated against ground truth, that holds its line when challenged. Vault is that system.

What does "fine-tuned" actually mean here?

Foundation models are a starting point. We fine-tune against our domain-specific dataset: 60K+ titles labeled with StoryDNA, demographic pull, market behavior, and the real outcome that closed. The result reasons inside the entertainment domain rather than inferring across it. Every prediction is grounded in a comp set you can audit.

What are the agents and how do they work?

Four custom agents sit on top: a CMO agent, a Head of Development agent, a Head of Research agent, and a Studio Head agent. Each reasons the way that role reasons in a real meeting, and each is calibrated by Vault operators who’ve spent their careers in that role.

How accurate are the predictions?

Vault AI maintains over 85% average accuracy across core predictions. Accuracy is measured by comparing every prediction against actual in-market performance once a title launches.

Will Vault flip its answer if I push back?

No. The answer is calibrated against real outcome data, not your tone. Push back and Vault holds the line and surfaces the comps. Change the inputs and the answer changes, and Vault tells you which inputs moved it. That’s what calibration means.

Where does my data sit?

On secure Vault-controlled infrastructure, never shared and never used to train models for anyone else. Access is limited to the small number of Vault people working on your analysis, all under enforceable NDAs. After delivery, materials move to a separate encrypted archive under restricted access.

How long does deployment take?

Most teams are running their first predictions in week one. Full integration with internal slate data, brand-specific comps, and custom audience targets lands in two weeks depending on data scope and security review.