The Healer's Question

GlagalCloud had been running for two years. Fifteen customers, eight velociraptors (five permanent, three rotating temporaries), a protocol system, encryption, digital signatures, and a version-controlled instruction library. The cave complex had expanded to three interconnected chambers — the main processing hall, the storage wing, and the protocol room where incoming and outgoing messages were sorted.

Glagalbagal, now in his late years, spent most of his time in a chair near the entrance, occasionally offering advice and more frequently offering opinions that no one had requested. The day-to-day operations were managed by Blortz (who had, against all biological probability, outlived every other sheep-dinosaur in the region) and by a young woman named Trviksha.

Trviksha

Trviksha had arrived at GlagalCloud eighteen months earlier, sent by the regional trade bureau to learn the system. She learned it in three weeks and spent the remaining fifteen months finding its limitations. She was practical, impatient, and unimpressed by elegance for its own sake. When Glagalbagal explained the mathematical beauty of the key exchange protocol, she asked how long it took. When he said fifteen minutes, she asked whether there was a faster way. There usually was not, but Trviksha asked anyway.

She ran GlagalCloud's "hard problems" desk — the department that handled customer requests too complex or too unusual for the standard instruction tablets. Most of these involved custom queries: "compare my inventory across three warehouses, weighted by distance to the nearest market, adjusted for seasonal price variation." Trviksha would decompose the request into a sequence of standard operations, write a temporary instruction tablet, run it, and return the result.

Then Grothvik the healer walked in with a question that could not be decomposed.

The Question

Grothvik had been a GlagalCloud customer since the early days. Her patient records — symptoms, treatments, outcomes — were stored on dedicated shelves with proper encoding, headers, and indexes. She could query any specific patient, any treatment, any outcome. The system answered factual questions perfectly.

Her question was not factual.

Grothvik: I have two hundred and fourteen patient records. Some patients got sick during the rainy season. Some did not. I want to know why. What makes a patient likely to get sick?

Trviksha: What factors are you considering?

Grothvik: Age, diet, location, occupation, whether they live near water, whether they have been sick before, how many people are in their household.

Trviksha: And you want a rule that says: if these factors have these values, the patient will get sick?

Grothvik: Yes.

Trviksha looked at the data. Seven factors, each with multiple possible values. To write a rule tablet covering every combination would require thousands of rules — most of which would have no supporting data, since two hundred and fourteen patients did not cover every possible combination of seven factors.

More fundamentally, she did not know which factors mattered. Did age matter? Did location matter more than diet? Did some factors matter only in combination with others — say, living near water mattered only for people over fifty? Writing rules required knowing the answer first. Grothvik was asking because she did not know the answer.

Trviksha: I cannot write a rule tablet for this. I do not know the rules.

Grothvik: Neither do I. That is why I am asking you.

The First Attempt

Trviksha's instinct was to look for patterns in the data manually. She pulled all two hundred and fourteen records and began sorting them: sick patients in one pile, healthy patients in another. She compared the piles.

Among the sick patients, she noticed that a disproportionate number lived near water. Seventy-two percent of sick patients lived near water, compared to forty-one percent of healthy patients. This looked significant.

She checked age. Among sick patients, sixty-one percent were over fifty. Among healthy patients, thirty-eight percent were over fifty. Another pattern.

Diet showed something too. Patients who ate primarily dried fish were over-represented in the sick pile, but only slightly — fifty-three percent versus forty-five percent. A weaker signal.

The Mess

Each factor, examined on its own, told a partial story. But the stories did not fit together neatly. Among patients over fifty who lived near water, the sickness rate was very high — higher than either factor alone would predict. Among patients under thirty who lived near water, the rate was barely elevated. The factors seemed to interact, amplifying or muting each other in ways that simple sorting could not untangle.

Trviksha: Water matters. Age matters. But how much each one matters seems to depend on the other. I cannot examine them one at a time. I need to look at them all at once.

Grothvik: Then look at them all at once.

Trviksha: I do not have a method for that. I have piles. Piles work for one factor. For seven factors interacting, I would need — she paused and calculated — thousands of piles, most of them empty. The data is too thin to fill that many categories.

Glagalbagal (from his chair): You are trying to find rules that the data is following. But the data does not know it is following rules. It simply is what it is.

Trviksha: That is not helpful.

Glagalbagal: Most true observations are not.

Trviksha returned to the piles. She had found something: the data contained patterns. Living near water appeared alongside sickness more often than chance would predict. Being older appeared alongside sickness. These co-occurrences were real. But extracting a usable prediction from them — one that accounted for all seven factors simultaneously — was beyond what manual sorting could accomplish.

She needed a systematic method. She did not yet have one.