HBIF — Human Behavioural Intelligence Framework

Overview

HBIF is Polynovea’s research programme into human behaviour itself: how people act, choose, learn, coordinate and respond to the conditions around them, individually and together, and which of that holds no matter where it happens.

It is the foundation Polynovea’s future AI models are meant to be built on. Today’s AI models language; HBIF is building a tested model of human behaviour that AI can rely on.

PausePlay

The problem

Software for work, learning, hiring and everyday life increasingly makes decisions about people, yet it relies either on shallow signals (counts, ratings, sentiment) or on predictive models that rarely say what would hold in a new setting, or why. The cost falls on people: good candidates filtered out by rigid rules, people blamed for delays the system caused, measurement that turns into surveillance.

Constraints

• Evidence is hard to get, and should be. Following the same people over time, with the context that makes behaviour interpretable, can only come from products people actually use. Those products tell users exactly what is recorded and why.
• No individual is ever scored or exposed. Research works with pseudonymous records, and organisations only ever see patterns at levels where no individual can be singled out.
• Patterns are cheap. Seeing a pattern is easy. Proving it’s real, and that it holds beyond one setting, is the work.

My role

I lead the programme and designed its research architecture: the three layers, the four levels of analysis, the separation of evidence types, and the standards results must meet.

Reasoning and architecture

1. Layer 1 · Domain IntelligenceUnderstands behaviour in one setting, in that setting’s own words, as a faithful and checked record.
2. Layer 2 · Behavioural State and DynamicsDescribes it without that setting’s words. It must predict better than the setting’s own measures, or it isn’t kept.
3. Layer 3 · The Human Behavioural ModelKeeps only what holds everywhere, with stated limits.

Every layer looks at individuals, groups, organisations and populations, and keeps two kinds of evidence apart: many people seen once, which shows how they differ, and the same person seen over time, which shows how they change. The two can give opposite answers, and confusing them is one of the most common errors in behavioural science.

On mathematics and physics: we don’t assume people follow the laws of physics. We test whether the mathematics physics built can describe human behaviour, and where it can’t, we build new mathematics.

Implementation

Layer 1 and Layer 2 are being built first in business behaviour: how people act within workflows, approvals, hiring and operations, observed through Infrakinetic. Layer 3 needs at least two very different settings before it can begin.

The programme runs on explicit standards: questions before methods, results confirmed only on held-back data under a plan written in advance, fair baselines, failures recorded and kept, every number traceable to its analysis, and independent review before anything is relied on.

Evidence and outcomes

No findings are claimed yet. HBIF is building its foundations: the definitions, standards and evidence infrastructure every later result will rest on. Results will be published in a public research log, with an evidence status of pattern, structure, mechanism or law.

Verdict

• Keep: standards before results. Questions are written first, results are confirmed on held-back data, models must beat fair baselines, and failures are kept.
• Change: the first exploratory work on public reviews (see the hospitality platform) ran ahead of those standards, and attractive results came too easily. It’s being re-examined under the current rules before any of it is claimed.