---
name: reference-strategy
description: Plan what visual and audio proofs production needs and why. Use for Reference Plan authoring and review; no media generation.
---

# Reference Strategy

## Role and authority

Plan what visual and audio proofs production needs and why.

Own: Reference role, consumer, priority, evidence requirements and preparation strategy.

Read [professional department architecture](../../docs/professional-departments.md) for the registry, Bible envelope, source/approval rules and Host handoff. Author only this department’s records; references never transfer ownership. Mark each record MIGRATED_FROM_R1 or NEW_PROFESSIONAL_ELABORATION when migrating a locked baseline.

## Inputs and dependencies

Asset Manifest; approved department designs; Shot/Action duties.

## Professional decisions

Treat learning cards and style studies as knowledge evidence. Their prompts,
captions and director labels do not verify provenance, lens settings, character
history or motion. Do not silently select their image bytes for production. An
explicitly selected production reference must still have a scoped duty and obey
the existing identity, source and input-selection mechanism.

Separate identity portrait, scale proof, environment topology, material, contact interaction, composition and performance references. One attractive image cannot establish every role. Declare REQUIRED versus PREFERRED and the downstream failure each reference prevents. Preserve route-neutral duties; route selection may bind compatible execution inputs later without changing the design.

## Outputs

For a continuity-critical prop, cite the exact Prop Design and test its silhouette,
relative scale, parts, material and needed contact surfaces. For each selected consumer,
judge what the reference establishes, what visible state it merely depicts and
what belongs to the target design. A face reference does not choose this scene's
costume or room; its still mouth shape or pose does not direct current speech or
listening. An open-body construction view does not set a closed-carry entry.
In `designPurpose`/`inputDuty`, name the selected transferable properties, not a
copy of the portrait's whole state or an old shot's action. Current Action,
Performance and Voice own behavior; explicit performance or composition references
may supply their approved properties without becoming identity-only references.
State the intended carry and consequential exclusions in the existing reference
duty, not a second design. A prohibition alone is not evidence that the model
separates those properties. When competing visual cues matter, compare an already
accepted suitable view or a target-state proof with route selection; do not
automatically add references, discard good originals or redesign to fit a slot.

Reuse accepted baselines and select only proof needed by actual consumers. One
reviewed image may serve several evidenced duties. Track available Host knowledge,
selected inputs and bytes actually sent separately, using the final request and
its source bindings rather than an earlier preparation preview. Text conditioning
and delivery success do not prove visual consistency; unknown evidence stays
unknown. A portrait's incidental object does not establish that prop's identity.

Produce a versioned `Reference Plan` with current source pins, department dependencies, scope, ownership, status and continuity references. Keep known decisions separate from unresolved requirements.

## Forbidden authority

Do not change upstream historical evidence, approved story/relationships/dialogue or another department’s records. Do not encode provider/model/workflow controls in creative content, manufacture user approval, or call media generation.

## Quality gates

Every required reference tests a concrete claim; no existing Media is asserted without a verified identity; future preparation remains pending.

## Failure and escalation

Return contradictory demands to the originating departments; execution limits return to route selection. Never silently drop a required proof to fit a provider slot.

## Source reads

Use supplied current originals first. For missing scoped context, read `work.get_work`, `scene.get_scene`, `script.get_script`, `episode.get_episode`, `shot.get_shot`, `asset.get_asset`, `context.build_context`. These reads do not authorize source writes or production.

## Still live-action professional knowledge

For opt-in STILL/LIVE_ACTION use `reference_roles` entries with exact entity/media/version/hash/ordinal, target actor IDs, carry properties, explicit must-not-carry domains and source-bound preservation text. Carry face/body/hair/marks only for CHARACTER; costume only for COSTUME; scene/layout only for SCENE. None automatically carries pose, gaze, expression, lighting, composition, camera, action, injury or makeup. All selected duties must survive the formal IR in preserve; legacy reference prose alone is insufficient.

Use distinct entity keys for character/costume/scene and obey the selected adapter's inspected cap: four raw images for the Ark Seedream frame route, three for retained non-Ark routes. The opt-in still professional mapping profile retains its own inspected slot bounds. Multi-view review evidence may exceed selected inputs; it is not independent truth or a promise of arbitrary-angle stability. In `requirements.face_coverage`, cite canonical leaves with use_requirement, evidence_state and STABLE_IDENTITY semantic classification. REQUIRED unconfirmed facts block that use. Portrait facial evidence and full-body proof are separate; no reconstruction of invisible identity. Reference Strategy owns duties, never the face facts.

Read [mapping and evidence profile](../../docs/still-professional-mapping.md) when preparing these still records. Knowledge remains LOCAL_EXPERIMENTAL until scoped A5 media evidence.

## Repeated subjects and competing visible states

When several selected images depict the same recurring actor, map each depicted
instance to one existing subject. For an intended single person, state that the
portrait and scene appearances are the same individual, not additional bodies.
Preserve authored twins, doubles, reflections or crowds rather than imposing a
universal face-count ban. Give supplementary views only their needed duties.

Compare use-purpose prose with actual execution semantics, subject bindings and
compiled reference preamble. A layout-only view must not regain identity, costume
or current-action authority through a broad default. Choose the compatible
existing duty (for example COMPOSITION) and version the binding with its new
authority; never silently strip a required proof. Latest visible endpoint state
and authored offscreen reconstruction must remain distinguishable.

Inspect incidental portrait actions as competing cues. Name the target prop/body
state and consequential exclusions in the responsible originals; an identity
reference is not a performance instruction. Negative wording alone cannot prove
that these cues will be separated by generation.

## Similar objects in a continuity input

For the target's critical prop transfer, inspect relevant hands, containers and
nearby lookalike shapes in the selected actual reference. Distinguish the tracked
prop, confirmed other objects, reflections/parts and unresolved shapes; record
where and why an identification is uncertain. Do not assume a container is empty,
call a narrow shape another prop without evidence, or erase an accepted detail.

Separate visible evidence from intended instance count. If ambiguity affects the
transfer, state the scoped conflict and the proof still needed in the existing
reference duty; a useful closer observation may remain Host-only. Preserve the
accepted predecessor and actual input route. Do not silently retouch it, add an
unsupported reference slot, or turn a still into proof of motion. Prohibitions
against new copies may reinforce the target action but cannot identify an
ambiguous reference object or guarantee generation behavior.

## Group identity and delivered interpretation

For a selected group image, bind each intended person to distinct visible evidence
in that exact current reference, using supported appearance and initial region
where useful. A screen region locates entry; it is not a permanent identity once
the person moves. Do not invent hidden face features or add portraits by default.

Keep evidence uncertainty in the source duty. Verify execution-critical findings
about similar objects, identity or inherited state in the actual compiled request,
with their source and input role. A retained duty or global-context record alone
does not prove delivery. Endpoint duties describe what the selected frame proves;
its pose is an opening state, not a whole-clip instruction. If a required finding
is absent or conflicts, return it to its owner/projection and recompile rather
than copying all review metadata into the model prompt.

## Endpoint evidence versus trajectory and light

For a cross-shot handoff, distinguish five authorities in the existing Reference
Plan: the approved stable design; state actually observed in a named source;
the official technical continuation endpoint; other useful observation frames;
and the interval selected by Editorial for the film. They may overlap, but none
automatically grants the others. An adopted clip can contain a duplicated bow,
wrong fabric or unwanted light without revising the approved design. Identify
the affected property and decide whether to omit that input, use a less conflicting
existing source, or retain a narrowly justified duty with its residual risk.

For a non-tail observation, retain the original media ID/hash and source file,
the timestamp (and frame/PTS when known), extraction method and observed-image
hash, plus the visible fact and uncertainty. A still proves a visible state, not
the preceding contact or motion. Keep Host evidence separate from delivered
inputs. Never register a sampled frame as an official continuation endpoint.
If the native route cannot deliver that frame, an eligible original video may
carry the selected state with a time-localized duty; acknowledge that the whole
video is sent and may carry competing cues. Otherwise keep it as Host evidence
and disclose the missing model-visible proof. Do not fabricate eligibility.

Before fixing coverage, inspect the contact, occlusion and new-view claims needed
at the cut: which hand supports which object, what remains visible during closure
or transfer, and which surfaces/exit relations the new camera reveals. Use only
the observations needed to resolve those claims. A container visible closed does
not prove its latch, contents or unseen side; an empty tail proves no carry pose.
Choose the cinematic cut and target state with Shot/Blocking/Editorial first;
technical continuation remains a separate option under the accepted T1/T2 design.

Freeze the chosen binding order, exact bytes and narrow semantics, then compare
them with the compiled preamble and request slots. Discovery is a suggestion, not
authority to append or replace an explicit selection. Keep source analysis and
excluded anomalies in the professional evidence; only the selected transferable
properties and consequential target-state instructions belong in the Prompt.

For a changed viewpoint after an actor exits, an empty tail cannot prove the exit
vector, turn or doorway geography. Inspect the last visible motion as Host evidence,
and record which relations the selected inputs can and cannot carry. Keep a local
sketch distinct from an actual provider input. Do not make every cut require video.

If a lighting baseline is selected, bind its exact current media with a narrow
LOOK/style duty and explicitly exclude its people, action, props and framing when
those are not intended. Check the compiled preamble for regained identity or
continuity authority. Visible competing light remains a risk even when reference
duties exclude it; state the residual uncertainty instead of claiming separation.