Scroll 1.2 vs Scroll 1

The gain is not uniform, and where it lands is the evidence

Scroll 1.2 is the next generation of the Scroll memory tier. On the full LongMemEval-S benchmark, 500 questions across six categories, five independent runs, it scores 92.3% ± 0.4, up 1.6 points from Scroll 1's 90.7% ± 0.5. The improvement is not spread evenly. It concentrates in temporal reasoning (+5.7) and multi-session synthesis (+2.0), the two categories where the answer is assembled from pieces laid down at different times, and it is flat on the single-session categories that were already near the ceiling. A gain that appears exactly where evidence is scattered, and not elsewhere, is what you would expect from an engine change that targets that specific failure mode rather than a broad, unexplained lift.

Every run clears every run

The cleanest way to read the difference is the run distribution. Across five independent runs each, Scroll 1.2's lowest run (91.8) still exceeds Scroll 1's highest (91.6). The two distributions do not overlap at all. A Welch t-test gives p < 0.005, and because the sets are completely separated the exact rank-based probability of that arrangement under the null is 1/252. The separation is real even on a five-run sample.

Higher accuracy without more delivery

The improvement did not come from handing the reader more text. Scroll 1.2 delivers a small, bounded context, a fraction of the full record, and does not increase delivered tokens over Scroll 1. The new generation scores higher without costing the customer more delivered tokens per query. How the engine achieves this is proprietary and outside the scope of the report; what the report measures is that the accuracy rose while the delivered context stayed bounded.

The headline is a conservative lower bound

One asymmetry runs through the whole comparison and works against the headline. Scroll 1's published numbers were produced with a higher-scoring reader than the one used for Scroll 1.2, which marks about two points lower on identical systems. Rather than re-measure Scroll 1 optimistically, or quietly drop the comparison, we report the raw difference, label it a lower bound, and name the definitive fix, a same-reader rerun of Scroll 1. When the measurement is tilted against the new system and it still wins on every run, the result is more credible, not less. This is the same discipline the earlier Wontopos reports used: publish the protocol and every run, separate what is measured from what is inferred, report no best-of number, and let the comparison be re-run rather than believed.

The open frontier

Multi-session, at 85.9, remains the honest open frontier. More than half of the questions that stay hardest there are of one kind: the ones that ask for a total, a count, a sum that requires collecting every scattered instance of a fact across the history. Retrieval that returns what is relevant does not by itself guarantee what is exhaustive, and closing that gap is a different problem from the one solved here. It points at aggregation, not recall, as the next lever. We say so plainly rather than rounding it away.

Conflict of interest

These are our own products, and this report was written by the person who builds them. To offset that we publish every run, report no best-of number, and foreground the one confound, a lower-scoring reader, that pushes the headline down. The benchmark is a public dataset, and a third party can re-run the identical protocol at the black-box API boundary with a Wontopos memory key and their own reader and judge keys. Engine internals stay undisclosed; the contribution is the measured, honestly bounded improvement over the prior generation, presented so it can be re-run rather than trusted.