## Fresh-context asset adaptation

On September 5, 2026, a separate agent received only the downloadable project
and its guide, without repository or private product context. It replaced the
fixtures in **CameraPush** and **MotifMontage** with an original fictional trail
map and four distinct route-badge treatments, then rendered a new 13-second
sequence. The map is an illustration, not geographic or navigation data.

The trial exposed three gaps: a concrete asset-replacement example; complete
two-recipe registration, font loading and local clocks; and diagnosis of a
restricted-host failure. The project now includes `ASSET_ADAPTATION.md` and
five complete files under `examples/trail/`. The example has a separate entry
point and does not change the fourteen default compositions.

A fresh extraction of the revised **26-file ZIP** passed `npm ci`, typechecking
and all tests. The agent rendered the documented `TrailSequence` entry without
editing any archived source file. Its final export has **390 frames, 13 seconds,
1280 × 720, 30 fps, H.264/yuv420p/BT.709 and no audio**; full decoding completed
without errors. The original thirteen recipe exports and 50-second sequence
retain their previously inspected source and rendered bytes.

### Host-qualified rendering result

On the restricted test executor, the documented Node interface-enumeration
preflight and unassisted render both failed with `os.networkInterfaces()` /
errno 97 before evaluating a composition. The successful retest used an
explicitly disclosed, external process-only loopback-interface adapter.
No project or Remotion dependency was patched; the adapter is not included in
the download. This is **not an unassisted render success on that host**.
The guide distinguishes that restriction from missing Chrome libraries and
directs users to a supported host. The earlier default-project clean render
and browser-install checks remain separate evidence.

### Rendered comparison

The first adaptation exposed a map layer leaking behind a fixed footer during
the push. Full-width frame-space masks fixed it before acceptance. The complete
corrected sequence was played at 1×; the final packaged-example render was then
compared with that accepted export across all 390 decoded frames and 21 selected
boundary, motion and hold frames.

All 210 montage frames match exactly. The camera shot has small raster
differences: whole-sequence SSIM is 0.999557 and frame 84's mean absolute channel
difference is 0.2824 on a 0–255 scale. Side-by-side and amplified differences
show no material timing, geometry, text, mask or continuity regression in the
inspected frames. The precise cause of the raster differences was not established.
**The exports are not pixel-identical.** The failed strict-equality check is
retained in the local evidence; it is not reported as a pass.

These checks resolve the three guidance gaps. They do not prove that arbitrary
replacement assets will fit, that every host can render, or that an actual
talking-head recording has been synchronized. The separate-video workflow,
real-recording timing and storyboard approval boundary still apply.