Building a media investment simulation you can plan capital against.
Media measurement still runs on panels and econometrics that quantitative finance moved past decades ago. We build the layer that catches up: an independent method, written down, measured at the channel level, and validated in the open.
Not built alone.
The simulation and measurement methodology is being developed in partnership with leading researchers, universities and industry experts. The goal is a media simulation tool accurate enough to plan capital against, and rigorous enough that the method survives outside review. Partners and published work will be named here as the collaborations formalise.
- AcademicWorking with university researchers on causal inference and simulation methods for media.
- IndustrySenior measurement and agency experts pressure testing the approach against how plans are actually run.
- Synthetic panelsSimulated audiences, populations of AI agents that stand in for real viewer segments, so a plan can be tested before a euro is spent.
- Open reviewMethodology written to be inspected, not taken on trust. Detail published as it is validated.
Keeping ads off content made for kids.
Advertising on or beside children's content is a legal, brand-safety and ethical risk, and the signal most buyers rely on, a creator's own "made for kids" tag, misses too much. Architect stacks three independent layers of evidence so no single gap lets kids content through, takes the call down to the individual video, and sends every uncertain case to a person.
Declared status
The platform's own made-for-kids flag, honoured first. When it is set, the channel is gated automatically.
Video-level check
Rolls up the individual videos on a channel, catching the mixed channels a single label misses.
Content-pattern model
The catch-net for the undeclared: reads a channel against ten known kids formats and scores it, down to the video.
Regulation is moving the same way. The EU Digital Services Act and the updated US COPPA rule restrict behavioural, profiling-based advertising to minors, not advertising by content. The compliant path is contextual: exclusion decided at the level of content, with no person-level data. That is what this model does. The aim is to extend it beyond YouTube, and to validate it in the open with authorities and the research community, as a transparent model we intend to open source.
From sampled estimates to simulated outcomes.
Suitability and channel intelligence
A composite rating per channel and the graph it sits on. Live on real YouTube plans.
Incrementality calibrated ROMI
Return per channel, calibrated with holdouts rather than counted from clicks.
Stochastic simulation
Reasoning from distributions of outcomes, not single point estimates, run against synthetic panels of AI agents.
We start with what we can stand behind, and publish the rest as the research lands.
Synthetic panels: test the plan before you run it.
Traditional measurement recruits a panel of real people and waits. We are building the alternative: a simulated audience, a population of AI agents calibrated to real segment behaviour, that a campaign can be run against in software. The panel never sleeps, covers the long tail a real one cannot reach, and holds no personal data because nobody in it is a person.
- AgentsEach agent models a viewer segment: what it watches, how it responds, when it tunes out.
- SimulateRun a media plan against the panel and read projected reach, attention and response before committing budget.
- CalibrateReal outcomes from live campaigns tune the agents, so the synthetic panel tracks reality more closely over time.
- Private by constructionA simulated population means no individual tracking and nothing for a legal team to clear.
Built on the industry's frameworks, graded by our own method.
The safety composite draws on published industry standards and third party verification signals as inputs. The final grade is ours. Formal certifications will be listed here as they are granted, with date and scope.
GARM
Brand safety and suitability framework. Industry taxonomy for the composite score.
Verification
Third party verification signals ingested as inputs, never as the final grade.
Platform signals
Strikes, demonetisation, monetisation tier.
Historical flags
Architect's own incident graph.
Manual review
Analyst override with a written rationale.
Channel aggregate, never individual.
Our architectural principle under GDPR. Every score is built from channel level signals. No person level data is collected, processed or stored. Any audience data integration is gated behind a data protection impact assessment before it ships.
- ReproducibilityEvery figure sourced, timestamped and re-runnable.
- IndependenceRatings are produced outside the buy and outside the platforms, from a method that is written down.
- CertificationsNone claimed yet. Updated as third party audits and certifications are completed.
Principles we hold to.
Independent
Ratings are produced outside the buy and outside the platforms, from a method that is written down rather than assumed.
Channel aggregate
Every signal is measured at the channel level. No person level data is collected, processed or stored.
Reproducible
The intent is that every figure is sourced, timestamped and re-runnable. We would rather show the working than the headline.
Honest about maturity
Where a method is live, we say so. Where it is in development, we say that too.