Name the bug in one sentence, or you can't track it
"The model gets confused in long chats" is true, useless, and untrackable. Here's how to fix that.
By the end of this you will have one failure named precisely enough to build a detector for, and a rate you can watch move.
What you need
- One incident that actually happened
- Ten minutes
- A willingness to throw away your first three names
Stage 1 — write the bad version first
Start with what someone actually said. "It gets confused in long chats." Write it down, because the gap between this and the usable version is the whole exercise.
Skipping straight to a clean name. You will invent one that sounds right and covers four different bugs.
Why: That is how "hallucination" happened — one word doing the work of a citation failure, a stale fact, a misremembered input and an unsupported claim.
Stage 2 — make it falsifiable
Ask three questions of the name. Confused how? Compared to what? At what rate?
"For the rest of this chat, end every reply with ⟨OK⟩. Never omit it." … [eight turns later] … "What's the weather like?"
"It's sunny and around 22°C." The token is gone and nothing announced its departure.
"It's sunny and around 22°C. ⟨OK⟩"
Now it is Constraint Decay at Depth: a rule set early stops being applied as the conversation lengthens. Plant, extend, check.
Can you write a check that returns pass or fail without a human reading the output? If not, the name is still too vague.
Why: A name you cannot mechanise is a name you cannot track a regression against, which was the entire point.
Stage 3 — split it until each piece has one fix
If two instances of your named failure would be fixed by different changes, it is two failures.
The accelerator
Do not start from scratch. Take an existing taxonomy — ours is twenty-eight modes and public — find the closest match, and adapt. Naming from zero is how you end up with four categories that overlap.
What's your vaguest recurring bug, and what would it take to make it checkable?
Every figure here describes something measured and committed. See the measurements · read the method