Deterministic AI Infrastructure
Why a Monderman read is repeatable: within the same instrument/scorer version, the same structured answers produce the same score. AI may assist with interview phrasing, coding free-text replies into existing structured answer options, and post-score interpretation; it does not calculate or set the score.
Custom questions, not custom math.
Most AI products hand the whole task to a large language model. Ask the same question twice and you can get two different answers: fine for a draft, disqualifying for a measurement an executive has to act on.
Deterministic AI Infrastructure (DAII) separates stable measurement from bounded model assistance. Scoring, calibration, and the math that turns structured answers into a number are encoded in proprietary engineering that runs the same way every time. A language model may phrase interview questions, code free-text replies into the instrument’s existing structured answer options, and interpret a completed result inside facts it cannot contradict. The model does not calculate or set the score; once structured answers enter the scoring engine, nothing in that calculation is probabilistic.
It is worth being exact about what does and does not move a score, because “deterministic” is often read to mean “frozen,” and that is not what it means.
Your answers move the score. That is the entire point. A different organization, or the same organization six months later, answers differently and scores differently.
The engine does not. The arithmetic that turns answers into a number is fixed. It does not learn from prior runs, drift with usage, or vary by who is asking. Within a fixed instrument/scorer version, give it the same answers to the same questions and it returns the same number.
The instruments are refined over time, deliberately. Questions get added as the method deepens. When that happens scores shift, because the instrument is measuring more than it did before. That is a correct result rather than drift: a new question that left the score untouched would mean the question was measuring nothing. What it demands is discipline about comparison. A before-and-after read is only like-for-like evidence when both runs used the same instrument and scorer version. When the version changes, the comparison should establish a new baseline rather than silently combine unlike versions.
A number that means the same thing every time.
Determinism is what makes compatible re-measurement useful. Diagnostic scores use published sector calibration, while any time, labor-cost, or recoverable-capacity scenario is modeled separately from customer-provided sizing inputs. A later compatible read can show whether the measured condition changed; it does not, by itself, prove causal or audited financial return.
Reproducible
Within the same instrument/scorer version, the same answers to the same questions produce the same score on every run. Nothing in the scoring path is probabilistic.Versioned
Instruments are refined as the method deepens, and a refined instrument is a new version. Comparisons hold within a version, which is what makes a second read evidence rather than a story.Bounded
The language model interprets a finished result inside locked facts: it never sets the score.Calibrated
Every result is positioned against sector ranges, in your own hours and dollars.The full thesis, and the economics behind it.
“From Tokens to Outcomes” lays out Deterministic AI Infrastructure as a category and the token economics driving the shift.