For researchers
We're building the substrate. A research-grade infrastructure for capturing, structuring, and serving the wisdom that does not appear on the indexed internet — to a generation of AI systems that need it.
Get in touchMost frontier AI is trained on the indexed internet. It can pass the bar exam. It cannot tell a teenager what their grandmother would say when their first relationship ends. The gap between those capabilities is not a gap of model size. It's a gap of training data.
Waterfall captures consented life experience from seniors — directly, with their permission, in their own voice. The corpus is structured, attributed, revocable, and grows. We do not scrape it. We do not buy it. We are given it.
We do not train a frontier model. We build a sidecar that any host LLM can call per turn for the human take. This composes with whatever the rest of the field is doing. We are accompanying AI, not replacing it.
Episodic memory and encoding specificity. The questions are designed to surface specific scenes, not generalised accounts — because that is where the lesson actually lives.
Self-distancing, construal-level theory, confession framing. Lower the defensive register; raise the substantive content.
Default mode network as the substrate of narrative recall; embodied cognition as the substrate of moral intuition. The protocol uses both.
Eldercare, hospitality, retail. The robot has the smarts; the companion gives it warmth without falsity.
Passenger interaction. The calm voice when a passenger gets nervous, in a register that fits.
Where being a thing talking to a person is exactly the wrong shape. The companion provides the human residue.
The most vulnerable conversations in healthcare. Warmth + plural attribution + careful refusals.
Honest apology in the seniors' register. Reduced escalation rates as the falsifiable outcome.
Kiosks, transit, retail, civic services. Where the conversation has to feel local rather than global.
The empirical study
n=200, balanced across age cohorts. Companion-accompanied responses vs. unaccompanied baselines. Primary outcome: the warmth dimension of the Cuddy-Fiske-Glick scale. Pre-registered hypothesis: at least 0.8 standard deviations of warmth lift, with no more than 0.3 standard deviations of competence loss. We will publish the result regardless of outcome.
Read the full design16-page research brief with diagrams, citations, and the pre-registered study design. Written for cognitive scientists, neuroscientists, and social scientists.
Download PDF14-page narrative on the substrate, the architecture, and the long arc of the project. Written for collaborators and patient capital.
Download PDFJoin the project
Study replication, cross-cultural extensions, independent corpus audits, methodological critique. If any part of this resonates, we want to hear from you. The substrate compounds; collaborators compound it faster.
Email usinfo@waterfallwisdom.com