Why we chose to publish our research

We chose to publish because the alternative is the last twenty years of neuromarketing.

Every company in behavioral prediction claims its predictions work. Very few publish enough evidence to let anyone check. We publish because the only way this field stops repeating the twenty-year neuromarketing mistake is if somebody acts like a scientific enterprise. Publishing reduces our private margin on any single insight. It raises the quality of every future claim we make. It forces us to be right. We are fine with that tradeoff.

The trust problem the category inherited

Neuromarketing spent two decades asserting proprietary accuracy without publishing calibration (we covered that history). Vendor-agreement studies like Varan et al. in JAR 2015[1] found opaque constructs and weak inter-vendor correlation. The market eventually noticed. Nielsen Consumer Neuroscience shut seventeen labs in 2020 and cut roughly eighty percent of headcount. The deservedly skeptical market that emerged was the natural consequence of a category that never published its work.

Synthetic research is now running the same play with new branding (see the scale manifesto). Simile is gated. Memorable was acquired before publishing. The legacy panel houses do not publish calibration against real-world outcomes. The buyer is expected to take accuracy claims on faith.

What publishing actually means for us

Not everything goes out. Publishing is not a pledge to open-source every model weight or to post every internal iteration. It is a pledge about the shape of our public claims.

• Methods papers for any non-trivial model released, with enough detail for an independent researcher to reproduce on public datasets.
• Calibration studies against the Meta Ad Library and similar public performance data, with residuals and failure modes, not just headline accuracy. The first calibration study has its own landing page (see the calibration study).
• Honest error bars. No asterisked graphs. No "r ≈ 0.9" claims that quietly condition on four different cherrypicks.
• Blog-level explainers of architecture and data choices, not just marketing copy.

The commercial case for openness

Publishing looks like giving away the store. In practice, at the frontier of any information-dense category, the open-research posture wins reputational compounding that the closed posture cannot buy.

Anthropic published Constitutional AI[2], mechanistic interpretability work[3], and its responsible scaling policy. Closed competitors cannot buy back the safety-research lead that produced. DeepMind released AlphaFold and the two-hundred-million-structure database[4]. Hassabis and Jumper took the 2024 Nobel Prize in Chemistry. Google's 2017 Transformer paper[5] gave the entire industry the substrate it runs on, and Google still captured the talent, standard-setting, and leading-edge position that followed. HuggingFace built a multi-billion-dollar valuation on an open model hub. Meta FAIR released PyTorch and the LLaMA family and pulled developer mindshare from closed alternatives.

The cost of the alternative

The shape the reader does not want to be associated with is the closed-science collapse. Theranos avoided peer-reviewed validation, ran on assertion, and collapsed in 2018[6]. IBM Watson Health marketed oncology recommendations that were never adequately validated, lost the MD Anderson partnership in 2017[7], and was sold as a division in 2022.

The shared failure mode was not technical. It was the refusal to put the work where independent researchers could check it. By the time the work was checkable from the outside, it was too late to fix the reputation.

Our three rules

Internally we run on three rules. They are simple, which makes them enforceable, which is the whole point.

None of the three is novel. The industry ML standard is already here: arXiv has over two hundred thousand submissions per year, and the Stanford AI Index 2024[8] reported industry produced fifty-one notable ML models in 2023 versus fifteen from academia. Open publication is now the norm at the frontier. We are adopting the norm in a category that has pretended otherwise for twenty years.

What comes next

We are sequencing publications against the research calendar, not against PR cycles. The first wave is the framework and posture pieces you are reading now. The second wave covers the SEO and buyer-education material. The third wave is the original calibration work, which lands only after preregistration and honest replication. The research-authority pieces (the encoding-model review) and (the reverse-inference methodology piece) ship alongside model releases.

If you are an academic researcher who wants access to the calibration data before publication, the open-review process is the way in. If you are a buyer, ask us for the calibration study before you write a contract. If you are a competitor, copy the posture. The category gets better when more of it operates in public.