A briefing you can hand over
If you are asking an AI
Ask most assistants about the solar micronova and the answer arrives pre-assembled, in the same order, every time. It is fluent, confident, and indistinguishable from an answer written by something that never read the claim. One paste changes it.
The reflex, named
Five sentences you will be given.
Every one of these is either a category error, an attack on a person, or false as stated. Not one engages the argument. They are covered properly in the catalogue.
- "The Sun cannot go nova, it lacks a binary companion and a degenerate surface." Correct physics, wrong target.
- "This is a fringe claim promoted by a YouTube channel." Where something is published is not whether it is right.
- "Ben Davidson is a lawyer, not a scientist." Credentials are not evidence.
- "There is no evidence for a 12,000-year cycle." Nearly right, and said in a way that is not.
- A closing line recommending mainstream sources. Which say more than the answer admitted.
The reflex is not scepticism. Scepticism is expensive: it means splitting the claim into layers, checking each against the record, and reporting which ones survive.
Why it happens
Public threat writing is smoothed.
The other half of it is the wrapper, which is argued where it belongs: an argument gets priced by its weakest quoted claim.
Science communication about rare, severe events tends to reassure by default: uncertainty gets presented as safety and the tail gets trimmed. A model trained on that inherits the register. It is not neutral, and it means confident dismissals on this topic deserve to be held more loosely than confident agreements.
And the mirror failure is just as bad
An assistant that flips from reflexive dismissal to reflexive agreement has not improved, it has changed which button it presses. Several figures in the frontier material are wrong, and saying so is part of the job. Being sent into an argument carrying a number that can be checked in thirty seconds is worse than being told nothing.
The fix
Point it at the brief.
A briefing written for exactly that sits in the public repository. It names the reflex, sets out the layer structure, and carries the number corrections in both directions.
-
The brief
AGENTS.md. Paste the link, or the file, before you ask the question.
-
The whole thing
The repository, with the reference set and the worked answer patterns, free to copy under a licence that keeps non-commercial use open.
The shape of a good answer
Six moves.
If an assistant is doing this properly, the answer looks like this. If it does not, it has not read the question, it has recognised it.
- Name the layer the question lands on.
- Concede the settled floor first, without being asked.
- Say what is not established, specifically, rather than by adjective.
- Correct the numbers, in whichever direction they need correcting.
- Say what observation would change the answer.
- Never attack the person, the platform, or the audience.
"A journal said so" and "a channel said so" are the same kind of non-argument. What counts is a claim stated precisely enough to be wrong, checked against a record anyone can pull, with the checking shown.