The Obsidian plugin I’m building has an AI read and summarize each new source document into wiki notes; that
summarizing command is /ingest. When /ingest summarized a source, there was a bug where the facts (quotes,
figures) were accurate but the explanation of “why” came out different from the source. It happened three times in
a row in the same source (an article on strategy), in the same paragraph (a darts analogy): the first time it flipped
to the opposite meaning (“a precise target is useful”), and the next two times it vanished entirely. Upgrading the
model from gpt-5.6 to gpt-6 didn’t change it. Since
a plausible cause had once before turned out not to be the cause when actually tested,
I took the same approach this time.
I thought grammar was the problem (source grammar error A/B test)
Rereading the sentence in question, the source itself had a real grammar error (“but in the game of life it as an apt rule” — it should be “it is”). The sentence structure was doubly twisted too: “the bullseye isn’t even the top score in darts, yet in life it’s the right rule, because it teaches you nothing.” It was a plausible cause.
In a copy of the source I fixed just that one grammar error (touching nothing else), then ran real ingest on the original 5 times (including 1 trial run) and on the fixed version 4 times. Result: all 9 dropped it the same way. Grammar wasn’t the cause.
I thought the counterintuitive content was the cause, then dropped the idea right away
The moment grammar was ruled out, I thought “then it must be this”: the claim in the sentence is itself counterintuitive (that precision is actually bad, the opposite of the common belief in this genre of content). I even wrote it down as “the leading hypothesis left by elimination.”
That was faulty logic. Ruling out one hypothesis doesn’t make another one more likely. I retracted it at once.
Testing it directly on another source (control experiment)
To check whether counterintuitive content itself was the cause, I found a far more blatantly counterintuitive claim in a completely different source: an article on “why game theory explains smart people not winning,” with the claim that “the harder you try, the more you lose.” Running it through ingest 5 times, it was preserved accurately every time.
That meant counterintuitive content alone doesn’t produce this bug.
Putting the two paragraphs side by side: the number of repetitions differed
In the darts paragraph, the core claim appears in just one sentence, once. The game theory article repeats the same sentence (“Effort spent inside a rigged game gets absorbed by the game.”) twice in bold, each time backed by a concrete anecdote (a dollar auction, a traffic jam).
I manipulated that difference directly. I split an experiment that adds one short sentence after the darts paragraph into two branches: one sentence repeating the same claim once more, and one plausible but unrelated sentence (a placebo). Result: the baseline without repetition was 0/5, the placebo 1/8, and the branch that repeated the same claim rose to 5/8. It wasn’t simply that the paragraph got longer; the repetition of the same claim itself was the lever.
Finally I checked the opposite direction: in the same article, with the same structure, I found a sentence stating a completely ordinary claim (“a general sends scouts before acting”) only once, and ran it 8 times. All 8 survived.
It reproduced the same way on three models: from gpt-5.6-luna to gpt-6-sol
gpt-5.6-luna, gpt-6-luna, gpt-6-sol: the same pattern showed up identically on all three models. That was why switching to a newer model didn’t fix it.
Detecting and fixing it worked, but it isn’t in the code yet
Once the cause was this narrow, a fix followed naturally: one narrow verification call to detect an omission or substitution, and one narrow repair call that, when triggered, explicitly puts the claim back. Testing 5 runs, detection was 5/5, and every run where the repair was actually applied (3/3) was restored successfully.
Even so, it’s not in the code yet. So far this pattern has been confirmed in exactly one source. I don’t yet know whether the problem is broad enough to attach one more always-on call in production. Instead, I decided to keep counting how often this pattern (a counterintuitive claim mentioned just once) shows up in every ingest round, and to reconsider implementing it once cases pile up.
A postscript: since then I’ve counted this pattern in every real test. Across five runs from October 20 to 23 (two sets of six sources), 3 to 6 of the 10 to 12 claims that went against common sense and appeared only once were dropped every time. None had their causal explanation rewritten. Evidence is building that this isn’t a single-source problem, but it’s not in the code yet.
Final thoughts
The moment I ruled out one hypothesis, the remaining one felt automatically right. I actually made that mistake and only noticed when it was pointed out. To avoid treating something unproven as proven, I had to doubt the very sentence “this is all that’s left.” Fortunately I then built real controls and split it apart, and the cause was neither the counterintuitive content nor the lack of repetition alone, but the two combined. An explanation that sounds plausible by elimination and an explanation actually split apart with controls carry entirely different weight.