We use AI where it genuinely helps — and we never let it own a number that has to be right.
We use AI to suggest, to assist, and to queue work for your approval — never to decide unsupervised. Two places it earns its keep:
Here's the honest part: a single AI model, asked once, is statistically likely to be confidently wrong on hard tasks. And the human review that's supposed to catch it is exactly the kind of tedious checking that — let's be real — rarely gets done well.
So when a task matters and AI is prone to slip, we won't trust one answer. We'll ask three different foundation models the same question, each phrased a little differently, then compare and contrast their answers before a human ever sees it. Where they agree, confidence is high. Where they diverge, we surface the disagreement — which is exactly what a reviewer needs to look at. The human reviews a pre-vetted, cross-checked result, not a raw guess. Less tedium, fewer misses.
Rolling out now — the 3 AI Rule goes live across PackOps within days.
The math that must be exact stays exact. Sub-pixel measurement, real-world dimensions, anything dimensional — that's deterministic geometry, not a language model's opinion (see the Mars Rule). AI perceives, suggests, and orchestrates; it never owns the number.
Our automatic marker calibration is the model: computer vision measures the marker and ruler to the sub-pixel — auditable, repeatable — while AI handles the fuzzy parts (finding the card in a cluttered photo, reading labels, triaging shots). AI proposes, the deterministic math measures, and you approve.
AI where it helps. Determinism where it must. Humans approve.