When two models disagree, most systems average them. We measured ours and found that averaging would have destroyed the only result worth having.
The setup
AION Sentinel screens listed Indian companies for accounting risk. Two methods examine the same company independently and each produces a ranking.
The obvious next step is to combine them. One number is easier to display, easier to sort, easier to explain. Every product instinct pushes toward a single score.
Before doing that, we measured how the two actually relate.
What the measurement showed
Across 2,715 companies on the same snapshot:
| Measure | Result |
|---|---|
| Spearman rank correlation | -0.07 |
| Companies shared between their top fifties | 1 |
| Agreement at the 99th percentile | 0 |
A correlation of -0.07 is not weak agreement. It is no relationship at all. The two methods are, statistically speaking, looking at different things and telling you different things.
Why averaging fails here
Averaging assumes your inputs are noisy estimates of the same underlying quantity. Under that assumption the noise partly cancels and the average is better than either input.
That assumption is doing all the work, and here it is false. When two measurements are independent, the average is not a better estimate of either one. It is a new number that is worse than both, because it discards the information that made each distinctive.
The concrete case that settled it for us:
Moksh Ornaments sat at the 59th percentile on the statistical model and the 99.7th on the rule engine. Sentinel flagged it on 8 June 2026. SEBI issued an adjudication order for market manipulation on 1 July 2026, twenty-three days later.
Average those two percentiles and you get roughly p79. Buried in the middle of the distribution, indistinguishable from hundreds of unremarkable companies. The one case with external regulatory confirmation would have been the one the blend hid.
What we did instead
We inverted the design. Rather than treating disagreement as noise to be smoothed away, we treat agreement as the signal.
Because the two methods almost never concur, concurrence is informative when it happens. On the same snapshot only 11 of 2,715 companies were ranked in the top 10% by both methods independently. That rarity is exactly what makes it worth reading.
Both component scores stay visible in the output. Nothing hides behind a composite, and a reader can always see which method fired and which did not.
The part that generalises
Three things transfer to any system combining multiple signals:
Measure the correlation before you combine anything. It takes minutes. We nearly shipped a blended score on the assumption that two risk measures must broadly agree. They did not, and no amount of design reasoning would have revealed that.
A single number is a product decision, not a statistical one. It is worth wanting for interface reasons. It is not automatically more accurate, and if your inputs are independent it is measurably less useful.
Rare agreement between independent methods carries more information than strong agreement between correlated ones. If two methods always agree, the second is nearly redundant. If they rarely agree, concurrence means something.
On calibration
One thing we will not claim: that these tiers are probabilities.
There is a single externally confirmed outcome in the record. One. That is nowhere near enough to calibrate against, so the output is an explicit rank composite and the payload says so in a field named not_a_probability. A ranking within this universe on this date is a genuinely different object from a probability of anything, and conflating the two is how confident, wrong systems get built.
Two confirmed outcomes is not a track record either, and we do not present it as one. It is a reason to look, not proof of anything.
Sentinel is informational and educational only. Not investment advice, not an allegation of wrongdoing about any company, and not a SEBI-registered Research Analyst product. Statistical divergence from peer disclosure is a reason to read further and nothing more.
You can try a company lookup at dashboard.aiondashboard.site/demo — no signup, no card.
Top comments (0)