Perfect Compliance, Wider Swings: The Five-Label Intent Classifier

The bullwhip effect is what happens when small changes in customer demand become larger and larger order swings as they travel upstream through a supply chain. A retailer sees a bump in demand, the manufacturer orders extra to avoid a shortage, the component supplier sees that larger order and reacts again. By the time the signal reaches the top of the chain, a modest demand change can look like a crisis.

I measure that with OVAR, the Order Variance Amplification Ratio: how wildly orders swing compared with how wildly customer demand actually swings. An OVAR of 1 means orders are as steady as demand. Above 1 means the ordering is amplifying the swing. Lower is better.

An earlier experiment in this series (V3b) had let the model output a continuous number — a safety-stock multiplier — and the model consistently miscalibrated it. Language models have no intrinsic number line; they produce tokens, not quantities. So the design here removes the number entirely. If calibrating a float is the problem, replace the float with a multiple-choice label and let hard-coded software do the arithmetic. This is the intermediate V4a variant; the design it established was then carried forward into the V4b WorldEvents experiment, which is the primary V4 result.

Read this table with one rule in mind: lower OVAR means less order volatility. A value below 1.0 means the ordering policy is absorbing volatility rather than transmitting it upstream. Exponential smoothing is the reference point.

The design fixed the right problem and exposed a different one. Removing the number from the model’s output did exactly what it was meant to: compliance went to a perfect 1.000 with no fallbacks. Confining a language model to a small, fixed set of labels is a dependable way to get structured output. That part of the bet paid off.

But the same constraint that made the output reliable also capped the outcome. The five labels map to five fixed multipliers. Whatever the model reasons, the order it ultimately drives can only take one of a handful of values per period. That is the Equaliser Effect: the discrete lookup table creates a structural ceiling and floor on OVAR that no prompt, no extra context, and no memory can escape, because the label-to-multiplier mapping removes the AI’s capacity for fine-grained adjustment.

In this no-disruption seasonal environment, the multipliers the model selected drove the buffer up and down more aggressively than the demand pattern justified, and the orders swung accordingly — well past what exponential smoothing produces by gently averaging recent demand.

For this setup, the discrete-label interface is a dependable way to get structured output from a language model, but it did not beat a standard forecasting heuristic at keeping orders steady. The AI amplified the bullwhip about 6 to 7× more than exponential smoothing, and adding context or memory did not close that gap. These are useful, checkable facts about one model and one simulated chain.

This closes the V4a intent-classification line. The architecture it established — classification by the model, arithmetic by the formula — is carried forward into the V4b WorldEvents experiment, which tests the same design under stochastic lead times and world events and is the primary V4 result.

The limits are worth stating plainly. I tested a single model (gpt-4.1-mini), a single 25-month synthetic demand series calibrated to Indian automotive seasonality, and a single simplified three-tier chain with a one-month lead time and no demand shocks. The intent-to-multiplier lookup values come from domain logic and are not guaranteed to be optimal. I would not generalise these numbers to other models, other markets, or more complex chains without testing them there.