The declared threshold is 2% of image width (12.8 pixels at the 640-pixel render width). Rays at a 320×180 grid sample the background under the head/hair convex envelope and an expanded buffer. The nearest included surface is rejected if it belongs to a rock object or a terrain face steeper than 45°. The corrected clips have no forbidden sample within that buffer; clearance beyond the buffer is not measured. This is a sampled proxy, not a subpixel contour guarantee.

The scene test includes 1,319,672 triangles from 78 rock/mountain instances and three terrain meshes. Rock geometry was checked during the initial native renders; the unchanged paths retain those receipts. The new path verifies head/hair geometry against the extraction record. Terrain triangles are checked at passage endpoints. The monitor uses a static environment snapshot.

Flatter ground, shadows, plants, lighting contrast and emotional isolation are not part of this predicate. A clear rock gap can still be an unappealing composition. Native visibility checks sample rays; they do not certify artistic readability. Review the actual film separately from these proxy measurements.

Rendering uses a common foreground-defocus setting in all four clips, applied after the geometry-based selection. No generated frames, source retiming, character reshaping or environment relocation. Preview render quality is unchanged.

The first check bound only named rock objects. Its selected approach passed numerically but still placed the character against a rocky terrain bank. Adding convex terrain patches overcorrected: their envelopes filled empty space. The final check uses nearest intersections with the actual surface triangles. The terrain binding and measurement representation were corrected; the camera pool, separation threshold and native guards stayed fixed.

Watch the initially misleading comparison → · Read the correction history →

That failure matters more than a clean score alone: a preserved rule can preserve the wrong interpretation. This example records an agent making and correcting that mistake; it does not impersonate a human teacher.

This phrase, correction, entity bindings and numeric threshold were specified by the agent under the user’s instruction to proceed. They are not recorded user preferences, a human teaching session or a learned linguistic distinction. The initial protocol preceded the search. Visual inspection exposed missing terrain bindings; a convex terrain approximation then overfilled empty space. Both failures were retained. The final native-triangle rule was fixed before its repeat search and carries the same payload and SHA-256 identifier into both passages. This is a development experiment, not a preregistered benchmark.

RFS’s Changing alphabet and Second chamber motivate making interpretation correctable and keeping the source of a decision visible. Here that idea is tested as an explicit geometric rule in native video. The system does not discover a new concept or train a model from the correction.

The remaining research question is whether an author can introduce a useful distinction without an engineer writing its predicate, then carry it into unfamiliar work with less repair than a competent direct workflow. This experiment supplies a replayable test case; it does not settle that question.

The original plan: frozen-blue environment, then warm amber dawn, carrying the same performance and layout commitments. The input archive is the preserved pre-run preparation snapshot, not the current result record.

Runway Aleph 2, SDK 5.20.1, seed 20261001, MP4, five timed guides at 0, 0.75, 1.75, 2.5 and 3 seconds. No edit range or aspect-ratio override. Default moderation. One accepted generation per revision; no rerolls. Both raw outputs are retained. Input, guide and parent-output hashes are in the receipts. Stage one takes a 1280×720 upload; stage two takes the exact 1920×1080 returned file, including its compression. Those inherited differences are recorded, not normalized away.

The initial final guide at 3.79 seconds was rejected: the service validated a maximum of 3 seconds for this 3.83-second clip. Two identical unpaid validation rejections preceded a documented amendment to source frame 155 at 3 seconds. No generated result had been viewed. The final jump has no direct guide.

The documented workflow uses timed edited frames. The official SDK exposes no regional preservation mask, CFG, denoising strength or sampling-step setting. Internal preprocessing and model weights are not controlled here. The seed promises similar results for an identical request, not an exact replay.

The older text-only run used a different prompt and stricter appearance constraints. It remains historical context; this is not an isolated causal test of adding keyframes.

Transformation: the requested frozen world and amber direction must be visible. Preservation: character shape, facial identity, action and prop continuity must survive at matching times, including unguided frames. Continuation: the second edit must carry the selected decisions forward. These are separate judgments; polish cannot compensate for changed action or identity.

The conventional reference is the existing native-character restoration in a generated frozen plate, then fixed color grading. It is not physical relighting. It has all-frame native masks while the model receives five images, so it is not an equal-information algorithm benchmark. Its inherited edge/contact limitations remain. No independent artist completion, labor comparison or superiority claim is supplied.

One development shot and one seed per stage establish a bounded workflow observation. They cannot establish generalization or a new capability beyond existing tools. The next evidentiary step would require untouched shots and independent judgments under the same stated brief.

The requested world changes are visible. The girl, hair and staff remain recognizable at sampled times; the hand-to-mouth gesture, blink and head turn broadly correspond. No gross identity replacement was identified in the declared samples. These are agent observations, not independent acceptance or exact all-frame equality.

The conventional alternative uses source-derived foreground masks on a generated frozen background, followed by fixed grading. It receives all-frame source information while the model receives five images. Approximate masks, edge blur, planar registration and lighting integration limit that reference. This is not a skilled-compositor benchmark or a physical relight.

This Spring passage, 107.75–110.75 seconds, is outside the recorded monitoring-corpus ranges and was not used in Afterimage's earlier generative trials. The film itself was already in the research library. This is a new transfer check, not a globally unseen or randomly held-out benchmark.

It uses the same model, seed, five-guide structure and two-revision lineage. Source aspect, duration and semantic subject differ. The native masks used for Settlers are replaced by approximate image segmentation for Spring. A single generated frozen background supplies the look; source girl/staff pixels supply each guide. A preflight hair-mask defect was corrected before any video submission. All final guides and conventional reference films were frozen before the new outputs.

Fixed keyframes are at 0, 0.75, 1.5, 2.25 and 2.75 seconds. Aleph 2, SDK 5.20.1, seed 20261001, MP4, no range/aspect override, default moderation. The second input inherits the first output's resolution and compression. Model internals and exact stochastic replay remain uncontrolled.