# Reference-led reconstruction

Load after production-method selection, for `is_generation:false` reference-led
creative work. Reuse existing evidence for unchanged source bytes.

Before proposing reference-derived visuals, obtain source frames through
`video_studio` `reference.inspect` and inspect its attached color sheet. Retain
the returned source identity and sample times with observations. Binary file
text, probe metadata, and rejected extraction commands are not visual evidence.
If inspection fails, keep source-dependent design decisions unknown, recover
access through supported tools, and continue only independent product work;
report a remaining access blocker instead of inventing a reference analysis.
The initial sparse samples establish only what is visible at those times;
request relevant neighboring times before making motion or transition claims.

Understand the dimensions that actually constrain this job:

- **Content and persuasion:** subject/product, audience, hook, demonstration or
  proof, information order, payoff and CTA. Separate observed facts from inferred
  creative intent; a reference's offer, testimonial, demonstrated capability or UI
  is not evidence about the target product. Borrow visual organization without
  asserting those features: use supplied target facts, and label any conceptual
  interface as illustrative rather than presenting it as a verified product screen.
- **Visual design:** framing, hierarchy, composition, typography, palette,
  product placement, backgrounds and text-safe areas.
- **Movement and timing:** shot boundaries, subject/camera movement, transitions,
  beat lengths, reading time and emphasis. Sample around meaningful changes;
  inspect neighboring frames to locate boundaries instead of inventing timestamps.
- **Audio and text:** speech, captions, music and their relation to visible action.
  Transcribe when wording or speech timing drives the local edit. A silent
  reference has no transcript to recover.

Persist concise observations in `project/ingest.json` with input identity,
source times/frame evidence, unknowns, preserve/change decisions and intended
target scenes or segment ids. Cover applicable dimensions, marking absent ones
as absent; do not manufacture dialogue, product facts or motion from one frame.

Realize those observations in the existing canonical artifact: COMPOSE scene
copy, layout/motion and reference contracts; or EDL sources, references, temporal
anchors, segments and tracks. A separate analysis without corresponding authored
behavior is incomplete. Reference timing is evidence, not an instruction to copy
seconds blindly: when target wording/language changes, use its actual narration
timing and reading time, then align the corresponding visual beats.

Keep exact user assets and product facts where fidelity matters; generate only
the missing visual work. Reuse accessible references and completed media. A single
video-model request can contain several visual beats; reference shot boundaries
do not each require a paid generation. Preserve prior outputs on revision and
change only affected copy, tracks, scenes or inputs. Text embedded by a video
model is not an editable HTML/subtitle layer; disclose that distinction before
promising a cheap local text replacement.