How thirteen agents read one fixture
Keeping research passes blind reduces cross-pass anchoring, while shared prompts, models and sources remain correlated failure modes.
On this page
The research half of an assessment is not one long read. It is thirteen short ones, each with a single question, each blind to what the others found.
Why blind
A reader who has already decided that the home side is undercooked can judge later evidence in a way that favours that view. Experimental work on biased assimilation found that people scrutinised disconfirming evidence more closely than confirming evidence. Blind passes reduce that route for influence; they do not remove every shared bias.
When eleven passes find nothing and two find the same injury from distinct primary sources, that corroboration is useful. When six find one thing and seven find the opposite, the disagreement is also useful, and it says the fixture is less knowable than it looked. Agreement alone is not enough if the passes have repeated the same underlying report.
What the disagreement is worth
The tempting move is to resolve a disagreement by weighting the more confident pass. Self-reported confidence is not enough: forceful prose is not evidence that a source is reliable or independent.
So disagreement is kept. It widens the interval on the final number, which usually keeps the assessed probability close to the market price it started from, with the margin removed, and that is the correct behaviour: when we do not know, the honest answer is not to move.
The failure this does not fix
Independence between passes does not create independence between sources. If every outlet is working from the same agency wire, thirteen blind passes will find the same claim thirteen times and it will look like overwhelming agreement. Checking whether sources are actually distinct is a separate problem, and it is the one this design is weakest on.
What this does not tell you
Nothing here says the passes are right. Separating their outputs blocks one route for social influence, not every correlated error, and a fixture with no reliable published evidence returns thirteen empty hands.
Common questions
Why thirteen separate passes instead of one?
Separate blind passes reduce cross-pass anchoring. They do not make the errors independent, because the model, prompts and source pool can still be shared.
What happens when two passes disagree?
The disagreement is kept and widens the uncertainty on the final number. It is not resolved by picking the more confident one.
Do the passes share their findings?
No. Each runs blind to the others, and they are only combined at the end. That preserves one form of independence, although shared sources can still produce correlated errors.