Kerala Rising: From Exporting People to Exporting Ideas · Page 84 of 353

KERALA RISING · FROM EXPORTING PEOPLE TO EXPORTING IDEAS

KERALA RISING · FROM EXPORTING PEOPLE TO EXPORTING IDEAS
and to generate the labelled data that retrains the
classifier. The tiers exist so that expensive
judgement is spent only where it is needed, which is
what keeps the average cost per verification near
the floor.
Three design commitments govern the layer. It is
conservative by default: when the model is
uncertain, it under-credits rather than over-credits,
because a falsely denied point costs a citizen a small
annoyance while a falsely granted point costs the
audit its credibility. It is inspectable: every verified
fact carries its provenance; which model, what
confidence, what checks passed, so a quarterly audit
can reconstruct any decision. And it is correctable:
because the corpus is public, systematic errors are
visible to outsiders and can be challenged, which is
a feature rather than a risk. The truer framing for a
technical reviewer is that the layer does not need to
be perfect; it needs to be cheap, conservative,
auditable, and improvable, and it is all four.
This tiering is also why the cost numbers in Chapter
5 hold up under scrutiny. A program does not run
every submission through an expensive frontier
model; it runs almost all of them through the cheap
classifier and escalates only the hard minority. The
blended cost is therefore much closer to the
classifier’s floor than to the frontier model’s ceiling,
and it falls every year as both tiers get cheaper. A
reviewer who assumes every verification costs the
84