Making Mushy Qual Mathematical

Mathematical clarity has always been the exclusive domain of quantitative research. Conjoints, statistical probabilities, significance, confidence intervals, regression models — quant has developed an elaborate machinery for demonstrating how it gets from data to conclusion. Qualitative research has so far operated in a different epistemological realm. It relies on interpretation, experience, and judgment. The researcher is the instrument. You give them a pile of conversations and, if they are any good, they disappear into the material and come back with something that nobody else had seen. There is a certain amount of magic and/or witchcraft to it. The client has to trust that the person standing in front of them knows what they are doing.

In traditional staged qual — focus groups, IDIs, ethnography — this is not necessarily a bad arrangement. The researcher is close to the material and, importantly, the client can often be close to it too. They sit behind the mirror, watch the conversation, hear the language, see the contradictions, and feel it. A good moderator can pull them along and help them see the world behind the obvious answers. The researcher is interpreting, but the client is also experiencing the evidence.

Social Intelligence creates a different problem. We are not dealing with sanitized, framed, and manufactured conversations. We are dealing with natural discourse: people talking to each other, often without any awareness that their conversation might eventually become research. The language is messy. People contradict themselves, change the subject, use irony, repeat things they have heard somewhere else, misunderstand each other, and occasionally say something extremely important only once. There is no moderator keeping the conversation on topic. There is no neat research stimulus. There is simply reality.

And reality is messy and very, very large.

At that point, the traditional qualitative model starts to break. When you have thousands or hundreds of thousands of conversations, the researcher cannot simply read everything and then tell us what it means. More importantly, we have no particularly good way of knowing whether the researcher has actually understood the landscape or has simply found the conversations that support the emerging story. Qualitative analysis has always had a cherry-picking problem. Not because qualitative researchers are dishonest, but because humans are extremely good at finding evidence for patterns they have already begun to see.

The perfect quote is the simplest example. A researcher can find the conversation that perfectly expresses an emerging tension, put it on a slide, and make the whole room nod. The quote may be completely genuine. The person may genuinely feel exactly that way. But what we do not know is where that conversation sits in the wider system. Is it central or peripheral? Is it part of a large cluster of similar meanings, or is it an isolated observation? What other things tend to appear alongside it? What contradicts it? How much of the surrounding conversation have we actually explored?

This is where I think qualitative research needs an overhaul. Not because interpretation has become obsolete, and certainly not because machines can replace qualitative researchers. The problem is that interpretation has historically carried almost the entire burden of credibility. We ask the researcher to find the meaning, decide which meanings matter, construct the themes, select the evidence, and finally tell the client the story. Quantitative research has built mathematical machinery around these steps. Qualitative research has mostly built better researchers.

What if we could give qual some mathematics of its own?

I don’t mean turning qualitative research into quantitative research. I don’t mean counting words, assigning sentiment scores, or pretending that a social media corpus is statistically representative. I mean introducing mathematical structure into the process of understanding meaning.

This is what Bakamo's Reading Machine is about.

Imagine a way to take hundreds of thousands of opinions—and without interpreting or changing them—being able to file them according to their similarity to each other. No interpretation, no flattening—just an accurate and universal filing system. This is vector technology—or the amazing mathematical solution Google came up with a few years ago. A vector gives language a position in a multidimensional semantic space. Conversations that mean similar things tend to occupy similar positions; conversations that mean different things move further apart. This is not about keywords, it is not about which language—it is about the semantic meaning. This changes everything.

Suddenly, the job of qualitative analysis—and its quest for qualitative saturation—is not simply to find an interesting and compelling narrative to feed the client. We can map the structure in which those conversations exist and see the discourse holistically, even if we’ve not read or felt our way through a single post.

This is the thinking behind the Reading Machine.

The important thing about the Reading Machine is not that it uses AI to read lots of text. That is increasingly unremarkable. Anyone can throw a wrapper around an LLM and generate unremarkable summaries of the text. They are as bland, and as superficial or straight-out flat as it gets. An LLM can read thousands of posts and produce five themes. But it cannot feel, it cannot find hidden, dark, or nasty things—LLMs are programmed to be shiny, happy people. It just covers the problem, rather than resolving it. Moreover, most of these tools come from people who believe that you can automate annoying qual away. Good riddance to that.

the Reading Machine is based on a different philosophy. The machine is stupid. It should not do the work that it cannot do, but rather give qual researchers hundreds of eyes. It is an exoskeleton with guardrails for qualies. It forces a systematic traverse of the semantic space rather than relying on a researcher to decide which parts look interesting.

Let me reiterate. The human analyst remains absolutely central. But their role changes. Instead of being responsible for inventing the map from a selection of observations, the analyst can work with a map of the discursive data and decide what the structures mean. This flips the game. No longer does the quali need to demarcate the field, and determine what it means. Rather it can focus on the latter - math takes care of the first.

That distinction is important because the objective is not to remove subjectivity. That would be both impossible and undesirable. Qualitative research exists precisely because human interpretation matters. The objective is to enable subjectivity to encompass everything. Discover and explain everything.

How does it work?

To understand how the human and the machine interact, let’s first look at the Aboriginal myth of the Wawilak Sisters. As anthropologist Bradd Shore notes in Culture in Mind, the myth tells of two ancestral sisters walking across the unformed, raw Australian landscape during the Dreamtime. As they trek, they encounter plants, animals, and natural features. With each encounter, they call out the names of these things, giving them language, classification, and place within the world. Shore uses the Wawilak Sisters myth to illustrate what he calls the “two births of culture”—the idea that meaning exists simultaneously in the external world and the internal human mind.

This is exactly how Bakamo's Reading Machine operates.

The vector math creates the unformed landscape. It maps the topography of the conversation, grouping millions of data points into clusters and structures based purely on semantic similarity. But until a human looks at it, it is just empty geography.

We analysts are these sisters. We wander across this mathematical landscape, looking at the clusters the math has surfaced. We use our empathy, the cultural context, and our judgment to understand what they are looking at. Then, they name it. By calling out these themes, they bring the raw data into meaning, building an accurate, mathematically backed model of the conversation.

So what?

Besides keeping qual human, it gives qual something it has never had: a defensible answer to “how do you know you’ve seen enough?” Saturation stops being a judgement call the researcher makes when the material starts feeling familiar, and becomes an observable property of the corpus — the point at which new conversations stop opening new semantic ground. We can show our coverage, not assert it.

This piece first appeared on Daniel's Substack.