Latest#048
The bias came first. The algorithm just made it cheap.
Issue 022 argued that ranking distorts what gets made. This is the correction: ranking did not invent the preference, it inherited ours and removed the cost of applying it. The case is Jack Parsons, flattened in 1950 with a file and a committee, no feed required. Our own instrument carries a rule written against exactly this inheritance.
◆ In the News▲ SILT Analysis & Response● What We're Watching
01In the News
In issue 022 we argued that ranking algorithms bend what gets produced. We stand by the finding and want to narrow it. The algorithm did not author the preference it enforces. It absorbed a human habit and stripped out the marginal cost of applying it at scale.
Consider Jack Parsons. He co-founded JPL and Aerojet and invented the first rocket engine using a castable composite propellant. In 1950 his security clearance was revoked, permanently, for removing confidential documents on explosives and propellant manufacture without permission. He died in 1952 at 37 in a chemical explosion in his home lab, ruled accidental. In 1972 the IAU named a crater on the far side of the Moon after him.
1950 had no feed, no embedding, no recommender. Parsons was flattened anyway. He was distrusted from the technical end as a security risk and, at the occult end, by the head of his own order. Aleister Crowley, writing to Karl Germer about the Babalon Working, called the participants idiots producing a Moonchild. The surviving text of that letter varies between two words for them, so we paraphrase rather than assert one.
02SILT Analysis & Response
A classification system needs you to be one thing, because a single label is what makes a judgement affordable. Clearance boards, professional bodies, tenure committees and credit references all did this before software existed. Parsons could not be both a propellant pioneer and a man who kept strange company; the file required him to resolve to a risk. An algorithm is that same habit with the marginal cost taken out.
This is where issue 022 implicates us rather than the machines. Our own finding there was that frequency is substantially constitutive of a model's apparent competence: popular over true. The model did not invent that preference either. It is a mirror with a gain knob. The useful part of that framing is that it tells you where to aim a remedy, at the inheritance rather than the amplifier.
We built one such remedy into our own instrument. C.I.B.'s D-scorers are forbidden from reading a model's prose. The documented reason is that register, a human style habit, would otherwise enter the denominator of a cross-model ranking as per-model bias. We wrote a rule against inheriting exactly the thing this issue describes.
03What We're Watching
We are keeping 022 and this issue distinct on purpose. 022 is about ranking distorting production, what gets made. This is about ranking inheriting a pre-existing human bias, who gets filed. The second is the older problem and the harder one to measure, because the preference predates the instrument that applies it.
No new measurement is required to make the argument, and we have added none. Our standing figures are unchanged: 38 models scored across the current cohort, 44 with history, 18 retired, 2,290 scored cells, and a mean judge spread of 2.59. That spread is itself a reminder that a single score flattens disagreement the way a single label flattened Parsons.
What we are watching is our own denominator. The register rule guards against one inherited habit. We do not assume it is the only one. The next audit we run on C.I.B. will ask which human preferences we have failed to keep out of the scoring, not which the models invented.