Response to

Reviewer btN3

• Q1. How often does a scored chunk actually retrieve a relevant first visit, and do independent training runs support the 200 s optimum?
• Does the bank content matter at long spans? (bank-content ablation at inference, new)
• Q2. Do the external comparisons hold when realized motion is matched, and what does CaR look like in the camera-following diagnostic?

Q1. How often does a scored chunk actually retrieve a relevant first visit, and do independent training runs support the 200 s optimum?

We replayed the training sampler on the 380 training segments (4,000 spans per setting, ten scored chunks each) with the

training retrieval score, and counted a latent as relevant when it sees at least half of the chunk's view frustum

(0.3 and 0.7 give the same picture). Pre-window history only, because the window is read directly.

Two readings follow. (i) Longer spans help because more first visits are in range and retrieved visits are older, not because retrieval

gets better: conditional hit rate is flat to slightly lower. (ii) Raising the cap from 200 s to 400 s also halves the share of

short spans (10.2% -> 5.0%) and doubles the mean span (106 s -> 206 s). So the 400 s decline can come from coverage, from this change

in the span distribution, or from both. We cannot separate them with the runs we have, so we now state in Sec. 4.3 that 200 s is best

*in our tested setup* and present saturation only as a hypothesis.

Variance: the numbers in the paper vary rollout seeds on one trained model. [PENDING: G2] we trained 200 s and 400 s again with a different

training seed; [fill in: ordering preserved / not preserved, with the new Revisit values]. With two trainings per configuration we do not claim a standard deviation.

Does the bank content matter at long spans? (bank-content ablation at inference, new)

We kept the trained 200 s / 400 s models and changed only what the bank holds at inference (20 web walks, first 60 s, one rollout seed; the top-6 row is the paper's five-seed mean):

Both long-span models depend on the retrieved content: removing the bank, or filling it with frames from another video, takes Gain to about zero, and random frames from the same video recover only about a third. Same-seed top-6 values (0.464 / 0.363 and 0.390 / 0.299) are close to the five-seed means. This is evidence that the 400 s model uses its memory; it does not explain why 400 s is below 200 s.

Q2. Do the external comparisons hold when realized motion is matched, and what does CaR look like in the camera-following diagnostic?

*Camera following, five rollout seeds, CaR included* (first 60 s, v5 web walks; direction agreement / correlation with prescribed yaw / implied horizontal FOV, target 90 deg):

CaR follows the path as well as we do; its implied FOV of 60 deg is consistent with a narrower camera, not with a worse path. The other four are clearly worse.

*Returns stratified by realized yaw error.* Yaw is estimated from the generated video (phase correlation); it sees rotation, not translation, and is not a pose estimate.

We report two versions: one that assumes every camera has a 90 deg field of view, and one that rescales each video so its turn amplitude matches the prescribed path (this removes the effect of a different field of view, e.g. CaR's 60 deg, and keeps only unsteady turning).

Revisit / Gain over five rollout seeds, returns with error <= 10 deg, per-video calibrated:

Without calibration (90 deg assumed) the picture for the other four models is the same; CaR then has only 200 of its pairs in this bin (its 60 deg field of view reads as large error), with Revisit 0.490 and Gain 0.341.

Honest summary: among well-aligned returns, CaR is within seed noise of Memorizon 100 s and 400 s, and Memorizon 200 s is ahead of it by about 0.06 Revisit and 0.07 Gain; HY-World, Matrix-Game, Infinite-World, LingBot and DreamX remain far lower.

Most returns of the other systems fall in the > 25 deg bin, which is the confound the reviewer describes. We therefore (a) move the camera-following discussion into Sec. 4.2, (b) describe the cross-system ranking as indicative and the within-backbone ablations as the primary evidence for memory, and (c) no longer claim a clear margin over CaR from the 60 s table alone.

Fold-back and other returns give the same picture (split in the supplementary tables).