Jervis Labs

01. CINEMIND · SWARM-INTELLIGENCE PREDICTION ENGINE

CinemindALPHA

From concept to global hit,
test market reaction before release

Cinemind is a content decision engine that simulates market reaction to your work in advance, inside an ecosystem where thousands to hundreds of thousands of virtual consumers live. Upload a script, synopsis, trailer or webtoon manuscript, change the ending, trailer or release cadence, and compare which choices create interest and buzz before release.

  • Film
  • Video
  • Webtoon
  • Short drama
Collaboration inquiry See how it works
Illustration of a crowd of virtual consumers connected by a network, reacting to a work on a cinema screen

THE PROBLEM

Sound familiar?

01

Decisions are made before release, but reactions only show up after.

Money for production, rights and marketing is committed before release, yet at that point there is no real audience-reaction data.

02

Surveys and FGIs are small and slow.

Collecting samples costs money, and every change to the plan means re-running the research. They simply add up individual answers and can't capture how audiences influence one another.

03

Comparing past hits can't answer “what if we change this?”

They can only explain results that already exist, and for new genres, formats or overseas markets there is little to compare against to begin with.

04

You only learn where word of mouth spreads after release.

Differences in reaction by country and taste, and the buzz and controversy that spread through fandoms, reviews and social media, only become visible after release.

WHY CINEMIND

How Cinemind is different

01

Ask a crowd, not one person

A group of hundreds of virtual consumers from different countries, cultures and tastes recommend, share and argue with one another. We plan to scale up to thousands to hundreds of thousands. Alternatives are compared by interest, immersion, drop-off and spread metrics.

02

Understand the work and the market separately, then connect them

Characters, events, conflicts and emotional shifts in the work become a content network, while audience groups, fandoms, critics, influencers, media and platforms become a market network — and the two are linked. You can trace which scene drives which audience's behavior.

03

Virtual consumers remember what they see and hear

Even the same viewer judges differently depending on whether they saw the trailer, got a friend's recommendation or ran into a controversy. Each virtual consumer has a disposition, prior experience, current state and memory, and chooses on its own to watch, hold off, drop off, pay, recommend or criticize.

04

Watch reactions spread across time and channels

Runs across multiple rounds — pre-release, early release, spread and sustain — applying different exposure and behavior rules for social media, fan communities, review spaces and content platforms.

05

Evidence, not just numbers

Get the moment reaction turned, the audience groups that drove spread, and the scenes and lines behind it in a report. If you're curious, ask the virtual consumers directly.

06

Continuously calibrated with real results

After release, we compare actual audience, viewing and payment data with predictions and correct the model by the gap. The more experiments accumulate, the more calibration baselines grow.

Two-layer infographic linking the work's character-event-emotion network with the audience-group and platform market network
Content network × market network The upper layer is the content network of characters, events and emotions; the lower layer is the market network of audience groups, fandoms, critics and platforms. Linking the two by taste and narrative fit and by exposure history, we trace which scene drives which audience's behavior.

HOW IT WORKS

How to use it

From inputting materials to a comparison report, an experiment is done in five steps.

  1. 01

    Input materials

    Upload scripts, synopses, trailers or webtoon manuscripts and choose target countries and audience groups

    OutputCharacter, event and emotion structure data
  2. 02

    Review the networks

    Review and edit the extracted characters, events, emotions and episodes

    OutputContent & market networks
  3. 03

    Set experiment conditions

    Define the baseline and variant, and specify variables such as ending, trailer and release cadence

    OutputExperiment design
  4. 04

    Run

    Set the number of agents and rounds, run it and watch progress

    OutputPer-round reaction logs
  5. 05

    Report & deep queries

    Compare results and follow the evidence through virtual-consumer interviews

    OutputComparison report
Cinemind start screen: an area to upload plan, scenario and synopsis files, settings for cast groups, reaction paths, period and real films to compare, and a list of recent projects

KEY FEATURES

Key features

Content analysis that turns a work into structured data

From scripts, video and webtoons we extract characters (relationships, motives, personality changes), events (causality, order, twist positions), conflicts (axes of opposition, escalation, resolution), emotional shifts (scene- and episode-level emotion curves), and scenes and episodes (length, placement, cliffhangers). Episode-level tension and likability curves and character relationship maps show drop-off risk zones and twist points at a glance.

Narrative elements × audience-group fit

A table shows how well narrative elements such as a fast opening, share of romance, twist ending, fidelity to the original and cliffhangers fit audience groups like original fans, new viewers in their 20s, overseas romance fans and thriller lovers. Elements where fit diverges can go straight into a condition-change experiment.

Five conditions you can change

  • WorkOpening, character setup, event order, ending
  • MarketRelease country, target audience, share of original fans
  • MarketingTrailer, poster, promotional message, initial exposure group
  • DistributionRelease cadence, free episodes, price, release timing
  • External eventsCompetitor release, negative reviews, a particular scene going viral

Prediction report

Along with per-episode interest curves by condition, we summarize five things for you.

  1. Reaction turning pointsWhen interest rose or fell
  2. Leading & suppressing audience groupsThe groups that drove or blocked spread
  3. Connected elementsScenes, lines and promotional elements tied to shifts in reaction
  4. Key assumptions & missing dataSurfaces assumptions that change results significantly
  5. Follow-up queriesIn-depth questions asking specific agents why

Agent interviews

Ask virtual consumers directly, such as “Why did you hold off watching at episode 2?” Answers can be checked against the exposure and behavior history actually recorded for that agent, and are labeled separately from real consumer interviews.

Experiments you can reproduce anytime

Each run automatically accumulates agent states, exposure and behavior history and records the experiment version and seed. With the same conditions, you can run the same experiment again.

USE CASE

Decisions it supports

Example: should we add a twist ending? — A single twist revealed in episode 8 sends each audience group down a completely different path.

Original fans

  1. Take it as a violation of the setting
  2. Post negative reviews
  3. Controversy spreads in communities
Interest falls · polarization

New viewers

  1. Immersed in an unexpected turn
  2. Recommend and share
  3. New inflow increases
Interest rises

Run Ending A (twist) and Ending B (original) repeatedly with the same audience mix and initial conditions, and you can see where paths diverge by audience group after the episode 8 reveal — and why. (Example)

Questions it answers

  • Which ending stays with overseas romance fans longer?
  • If the trailer leads with the revenge storyline, will episode-2 drop-off shrink?
  • Which creates bigger word of mouth: twice-a-week releases or binge release?
  • If a competitor launches the same week, how much will our initial interest wobble?

BUILT FOR

Who it's for

Film production & investment companies

Teams facing hard-to-reverse decisions such as production approval, rights purchase, casting and marketing budgets. Compare reactions to each alternative before release and decide on solid evidence.

Short-drama, OTT & video production companies

Teams that must set episode structure and release cadence. Short dramas in particular have short release cycles, so predict → verify → calibrate runs quickly.

Webtoon & IP holders

Teams who want to see in advance how original fans and new audiences will react differently when adapting a work to video or taking it overseas.

Platforms & distributors

Teams that want to attach reaction prediction to whole IP portfolios and campaign decisions. An annual-contract API and a dedicated analysis environment can connect to your own content, viewing and purchase data.

TRUST & SAFETY

So you can rely on it

  • Predictions are recorded first, verified later

    Predictions are submitted before release to lock the timestamp and are never edited afterward. Hits and misses are reported together.

  • We state clearly that these are simulation results

    Differences in a report are effects inside the model. We advise verifying actual causal effects with separate experiments such as A/B tests.

  • Performance is checked against real data

    We compare against real interest, viewing and purchase metrics on a held-out validation set that contains only pre-release information, and evaluate how much better we do than simple baselines such as similar titles or genre averages.

  • Your materials are stored separately

    Uploaded manuscripts are stored isolated per customer, and separate consent is required before they are used for training.

  • We don't identify individuals

    Audience and market data are used only in aggregate, with personally identifiable information removed.

  • Data is never mixed

    Content manuscripts, audience and market data, real results and simulation logs are stored separately, and training and validation data are kept apart.

ENGAGEMENT

How we work together

We are currently running a pilot on the IP of Korean film production companies.

Pilot

Pick one or two real choices on a work with a set release date and work with you from prediction through post-release verification.

  1. 01

    Select decisions

    Choose one or two among ending, trailer, release cadence and target country

  2. 02

    Record predictions in advance

    Submit predictions before release to lock the timestamp. No edits afterward

  3. 03

    Release

    The customer actually releases the work and runs the campaign

  4. 04

    Compare results

    Compare audience, viewing and purchase metrics with the predictions

  5. 05

    Calibrate the model

    Analyze causes of error and adjust parameters

  6. 06

    Case report

    Summarize as a reference within the scope the customer consents to

Deliverables
Pre-release prediction report · Post-release comparison report · Calibration log
Success criteria (agreed in advance)
Direction of prediction matches · Improvement over simple baselines · Used in real decisions · Intent to subscribe
Participation requirements
Confirmed release date · Able to share performance data · Two or more decision alternatives
Duration & pricing
Negotiated based on pilot scope

COMING SOON

Coming soon

  • Multi-country prediction engine beta — expanding personas by country and culture
  • Scaling simulations up to thousands to hundreds of thousands of agents
  • Cost and speed guidance by scale, calibrated with real data

We will let you know the release schedule once it is confirmed.

This product is currently in developer testing (ALPHA). Features and schedules may change.

GET STARTED

Before you decide,
experiment with reaction

Preview the market reaction you used to check only after release, right at the planning stage. From concept to global hit, prepare with Cinemind.

Collaboration inquiry Request an overview