One-sided annotations
What took the same model and prompt from 0 of 5 to 10 of 10?
Reading the reasoning of the lost episodes showed a pattern. At each decisive turn, the model cited an annotation that stated one side of a trade-off and left out the other, followed it correctly, and was wrong. The thesis calls this a one-sided annotation and found five kinds:
- a saving without the cost of reaching it: gems behind a door "save 155 HP", but reaching them costs 198;
- a value without its cost: a potion counted as the value of the fight in front of it, although it can be collected in any order;
- an efficiency without affordability: "take this fight before …" for a fight that costs 46.5 % of the HP;
- a deficit without scope: a budget that summed optional monsters and declared a winnable floor unwinnable;
- a count without multiplicity: two gems behind one door listed once.
Each was repaired in the engine by stating the missing side, never in the prompt. At turn 9, the same model in the same state reasoned like this before and after the net stat-route and fight-order lines were added:
Row 4 is the largest step: with model, prompt, temperature and lookahead fixed, two annotation lines took Qwen3.5-397B from 0/5 to 10/10. Rows 5 and 8 show that a repair can itself introduce a one-sided line, which the next row then fixed.