1. 01PDS survivorship bias

   The Planetary Data System contains only observations the mission actually downlinked. DEEPSIFT cannot see images that were never sent or were deleted onboard, so every result is about re-prioritizing archived observations, not about the rover’s full onboard image stream.

2. 02Archived, downlinked observations only

   The replay treats archived products as if they were waiting onboard. It is a retrospective experiment; it does not reconstruct what the rover actually held in storage at any time.

3. 03Simulated compressed stereo tier

   Scheduler V3’s COMPRESSED_STEREO_PAIR is DEEPSIFT’s own JPEG q50 ground re-encoding of both eyes — a SIMULATED PRODUCT TIER, not a NASA flight product. Traverse results use only NASA label estimates (full-quality and thumbnail pairs).

4. 04PLACES interpolation

   Rover positions come from the PLACES localized_interp table, matched by site/drive/pose; most frames use the nearest pose within their drive, not a per-frame onboard localization. Spatial metrics inherit that resolution.

5. 05One rover, one camera

   Curiosity Navcam only. Nothing here is tested on other cameras (Mastcam, Hazcam, ChemCam RMI), other rovers or orbiters.

6. 06Limited mission windows

   Four sol windows (412–430, 779–820, 1100–1129, 950–979) and traverse sequences only. Test: 12 traverses, 679 frames.

7. 07Visual-change proxy ≠ scientific value

   Visual-change coverage is a MobileNetV2 embedding similarity. It is not a judgement of scientific interest, and spatial coverage is geometric: neither metric measures whether science value was preserved.

8. 08No flight hardware test

   Everything ran on a laptop CPU. There is no test on radiation-tolerant flight processors, flight software frameworks or real storage hardware.

9. 09No power, thermal or compute-budget model

   The simulation models bytes, not energy, heat, CPU time or memory on a spacecraft.

10. 10No NASA validation

    DEEPSIFT is an independent research prototype. It is not reviewed, endorsed, used or validated by NASA, JPL or the MSL mission team.

11. 11Failed components are excluded, not fixed

    QUALITY_V2 (validation false-positive rate 10.6 % vs 3.6 % in development) and constrained pHash (false-merge 0.15–0.19) failed validation and are not part of the claim. A quality detector would need a new validation interval.

12. 12The held-out test is used up

    The final test interval has been analysed once. Any change to the method now needs a new, untouched test interval.

On a held-out Curiosity Navcam interval (sols 950–979), position-based traverse sampling with a stereo-safe progressive scheduler kept 1/4 of traverse frames at full quality, using 26.1% of full-quality traverse bytes, while every archived traverse frame stayed within 5 m of a kept frame (maximum 2.05 m) and no stereo pair was broken.