NPS 80 to 95
Research-to-product translation. Kudos as labeled training corpus. Six surfaces from a HEXACO personality substrate.
HR buyers wanted flight-risk prediction. The data supported it. We said no. A person's declared values beat any model's inference about them. Predicting who quits trains a model on suspicion. We refused that surface and built growth potential on the same data. Declared beats inferred was a product constraint, not a slogan. Every surface started with what someone said about themselves, not what the system guessed.
Read the case What shipped, how it worked, the lines that land
Role and team
Senior / Founding Product Manager (Apr 2022 to May 2023). Eight people total. General Manager Willie Litvack (direct line to CEO). Partnerships (Dan Kasper), one senior fullstack, two frontend, two backend. Advisor bench of five led by Dr. Galen Buckwalter: dual PhD ML and Psychology, psychometrician behind eHarmony's matching algorithm.
The science partnership
Galen ran the science. HEXACO as the validated six-dimension instrument (honesty, emotionality, extraversion, agreeableness, conscientiousness, openness), plus IPVS (Instill Personal Values Scale) and ICVS (Instill Company Values Scale). My job was translation: take peer-reviewed work and turn it into product surfaces a team of eight could ship and a company of fifty could feel.
Six surfaces, one system
- HEXACO assessments: the spine. Everyone took it. Everyone got their map.
- User Profiles: HEXACO map plus declared work preferences (response times, feedback style, place in the org). Built for the coworker who needs to ping you.
- Directory: every profile in one view. Org as greenhouse, not chart.
- Slack integration: surfaced informal org structure beyond the formal chart.
- Onboarding 2.0: IPV and ICV handshake ceremony mapping personal values against declared company values. +30% activation. Converted enterprise waitlist (US Army, Deloitte) to active.
- Kudos: monthly value-tagged peer recognition. Each kudo mapped against the recipient's HEXACO profile and the company's declared values. Became the values-affirmation training data for the recommendation engine.
The decision that mattered
Refused to build flight-risk prediction. Built identify growth potential instead. Same data, opposite product surface. The values-affirmation surface (Kudos) doubled as the labeled corpus. Every act of recognition was a peer-validated training signal of what humility or courage actually looked like inside that company.
Craft: how it actually worked
Six surfaces, composed
Each surface had a specific role and reinforced the others. The system was the product. HEXACO at the spine, Kudos as both surface and feedback corpus. Strip a surface and the loop collapses.
Week to week
Take peer-reviewed work from a five-advisor science bench. Ship a product surface a team of eight could build. Measure with NPS, not vanity metrics. Use the result to scope the next surface. Repeat.
Kudos as training data
Every recognition mapped against two coordinates: the recipient's HEXACO profile and the company's declared ICV values. The recommendation engine had a peer-validated, instrument-anchored signal for what each value looked like inside that company, not in the abstract.
Same data, opposite surface
The flight-risk model and the growth-potential model could be built from identical inputs: HEXACO, tenure, manager input. One trains on suspicion. The other on opportunity. The architectural decision was which surface the data became.
A person's declared values beat any model's inference about them. Not as ethics, as architecture.
See the proof Composition, primitives, first principles
Recognition data and risk data have the same shape and opposite valence. Building both from the same store is a category error.
Primitives: the reusable architecture
Peer-validated, value-tagged, instrument-anchored signal
Three properties that compose into a high-quality human-data primitive. Strip any one and the signal collapses.
- Peer-validated: the rater is the coworker, not the algorithm.
- Value-tagged: every signal carries declared semantic shape (which value, which trait).
- Instrument-anchored: every signal is locatable on a published, peer-reviewed scale.
The handshake ceremony pattern
IPV times ICV before access. Onboarding as declared mutual scope, not orientation. Reusable wherever two parties need to bind to a frame before transacting.
Refusal as architecture
Refusing flight-risk wasn't ethics-as-vibe. It was a primitive. Every product surface declares what it predicts and what it explicitly refuses to predict. The refusal becomes part of the spec, not a side conversation.
How this composes
- Amazon Refusal-as-data at survey scale.
- Clause Declared values operationalized per individual.
- Oracle The Kudos corpus is a semantic event store at company scale.
- Declaration Your portrait is your own, including the algorithmic one.
Same data, opposite product surface. The most consequential PM decision isn't what to build. It's what to refuse to build with the data you already have.
First principles
Declared values beat inferred values
Not as ethics, as architecture. The declared signal carries provenance and consent. The inferred signal carries neither. Anywhere you can swap inference for declaration, the resulting system is more legible, more correctable, and more legitimate.
Same data, opposite surface
The most consequential PM decision is not what to build but what to refuse to build with the data you already have. Refusal isn't blocking the roadmap. It's authoring it.
The training corpus is the product
What you measure is what you optimize is what you become. Recognition data and risk data have the same shape and opposite valence. Building both from the same store is a category error.
Why this matters
Constitutional AI's labeled preference data has the same architectural shape: peer-validated, value-tagged, instrument-anchored. The discipline works whether the rater is a human or another model. The Instill build is a worked example of producing this kind of signal at company scale.
The training corpus is the product. What you measure is what you optimize is what you become.