The assayer's incentive

The assayer
owns the ore.

We hold equity in the companies we grade. Any assay office in that position should be distrusted by default. We would rather state the problem than let someone else discover it.

Structural defenses

Five constraints that make a flattering verdict hard to produce and easy to catch.

01
Thresholds are registered before the run

The revenue and usage levels that produce each verdict are written down at day zero, with the predictor version attached. A threshold set after the numbers arrive is a rationalization.

02
The verdict is computed, not argued

Day-ninety verdicts come from billing data against the registered thresholds. There is no partner meeting where a company gets talked up into a GO.

03
KILL is published like GO

Same page, same prominence, same cohort record. A studio that publishes only its winners has published nothing.

04
Our own accuracy is graded

The calibration ledger tracks every verdict to month eighteen, including every GO that died. That is the number a dishonest assayer would most want buried, so it is the one we lead with.

05
Selection criteria are public and versioned

Anyone can read what admits a founder and what does not, and see when it changed and why.

The economic argument

The record is
worth more than
any position in it.

Flattering a single company buys a marginally better mark on a marginally better position. It costs the credibility of every verdict we will ever issue. That is a terrible trade even before anyone catches it.

The output we are actually selling is a grade that a downstream investor will price. The instant those grades are understood as marketing, they price at zero, and so does everything downstream of them: the next cohort's applicants, the capital that follows a GO, the willingness of a founder to accept a KILL.

Which is why the calibration ledger exists and why it leads with our misses. Trust in an assay office is not built by being right. It is built by being checkable.

Check the terms
for yourself.