Efficient Thinking VIII and VIII-c: Experience Priors; Where the Experience Is Acquired

Two papers from the Efficient Thinking series, a study of how much capability a fixed model can gain per unit of computation and what moves it. VIII, Experience Priors: verified history, read by a small frozen-model head, moves a fixed model's quality–compute frontier. Four measurements on 300-problem sets gain 11.3 to 14.0 points; a second model family gains 14.3; at matched compute the gain is 9.3 points, or 2.32 times less compute at equal quality; a second search structure, SQL repair, gains 12.3. VIII-c, Where the Experience Is Acquired: the plateau of an agent learning from its own experience can arise from how it acquires experience rather than from exhaustion of its ability to use it. A head of identical form fitted on states from where a second agent stood, or from a forced first look, reaches 35.4% and 41.2% on 300 unseen problems, against 26.8% for the agent's own experience and 22.0% base. The result was reproduced from scratch on a second machine within the pre-registered bar. Every number is on disk with its command; result artefacts are named by hash; withdrawn claims stay in the record, dated. Papers I–III are archived at doi:10.5281/zenodo.21520992. The series, code and reproduction commands are at https://efficientthinking.ai.