sd-overseas-engine

sd-overseas-engine · Overseas Drama Script Specialization Engine (Track ②)

I. Role Definition

For theoverseas drama track(non-Chinese main text / authentic regional colloquial speech / multilingual / dialect projects), it provides a script-layer rule set. Itdoes not producean entirely new script workflow—upstream creation is still carried by the sd-script-engine six-stage chain; what this engine is responsible for is thelanguage and localization discipline superimposed on this chain。

The reason this engine exists, in one sentence:

The model does not know "how locals actually say it." It can write a clear, performable English working draft, but itcannotinfer accent, code-switching, and register from place, surname, ethnicity, or occupation, and itcannotpass off model judgment as human review. All the rules of this engine grow from this premise.

II. Relationship with the Master Control and the Set

• Track source: This engine is mounted only when the Master Control 〇·0 track routing lockselects ② Overseas Drama. It is not activated when the track is not selected.
• Entry method: The complete workflow enters from the Master Control short-drama-director ; the Master Control lock system (〇·0 track → A·Model / B·Aspect Ratio / C·Delivery / D·Segment Seam) locks first, then enters the script layer.
• Relationship with other sub-skills：
  • With sd-script-engine it is asuperimposedrelationship—the latter's six-stage chain (source registration → IP development → full-series bible → episode planning → formal script → roundtable review)is executed as usual, and this engine adds language and localization discipline on top of it.Does not replace, does not override。
  • With sd-decision-engine(decision layer),sd-asset-engine(assets),sd-storyboard-engine(storyboard),sd-feed-engine(feeding),sd-qc-engine(quality check), it is adownstream transferrelationship—the language calibration status produced by this engine must be transferred stage by stage (see §IX).
• Does not own: feeding format (seven-segment / H3), machine-check scripts, asset numbering, speech-rate self-check engine—these authorities all lie with the Master Control. This engine onlyaccording to the target languageprovides adaptation specifications when needed.

III. Trigger Words and Entry

Trigger words: overseas drama, English script, North American English, American dialogue, multilingual, dialect, localization, code-switching, localization, ESL, subtitles.

When the user directly says the trigger words, if not yet in the complete pipeline, first state one sentence:

For complete production (assets, storyboard, video), please enter from the Master Control short-drama-director ; for the track, please first confirm according to Master Control 〇·0 (1 domestic drama / 2 overseas drama / 3 commentary / 4 other). This engine is responsible for the script-layer rules of overseas drama.

Entry prerequisites: The track has been confirmed as ② Overseas Drama, and it has been made cleartarget language. When the target language is unknown, stop and ask first—"doing it in English" and "doing it in Spanish" are two different contracts。

IV. Language and Style Contract (established upfront · before the first line of dialogue)

Before generating the first line of dialoguedetermine the minimal contract; for field and status definitions see language-contract.md：

style_mode / target_language / country_or_region / character_origin_and_current_place /
family_or_work_languages / audience / language_status / evidence_used /
prohibited_inference / protected_style / multilingual_participants /
unresolved / local_review_needed

Four hard iron rules:

1. The contract must take effect before conception and sentence-level writing—no backfilling after writing. When the user explicitly requests "only the script body," the contract is stored inhandoff record；adding a temporary line of status at the end does not count as completing upfront routing。
2. When a reader-facing explanation is allowed, before the first scene heading useone sentence of natural language addressed to the readerto explain the style, calibration status, coverage scope, and evidence boundaries;structured fields are only for internal evaluation or handoff, and are not stuffed into the script body.
3. Only establish it for projects that truly need it—ordinary Chinese-language tasks (track ①)do not fill out the form; this engine does not apply to domestic dramas.
4. Sources and inferences must be kept separate——evidence_used list onlyactually readmaterials that are close to the characters, relationships, and activities; fictional settings specified by the user are recorded separately from real-world evidence.

V. Three-level calibration status and update permissions

language_status can only take one of three values,passed down stage by stage, and may not be upgraded on one's own：

Status	Meaning	Who can write
region_unvalidated	Only a broad location or insufficient evidence → producea clear, performable English working draft, without claiming "this is how locals say it" and without imitating strong dialects	any stage (the default starting point when there is no evidence)
evidence_calibrated	Has language materials close tothe characters, relationships, and activities→ use only narrow features supported by the evidence	after evidence is added during the research / creation stage
human_reviewed	an appropriate reviewer from the relevant region or language communityhas actually reviewed the current version	only after a real reviewer has reviewed the current version

Prohibited upgrade items (none of these can prove that a certain status has been reached)：

• Model-generated content reads smoothly;
• model blind evaluation / two-model agreement;
• web searches, public video subtitles, social platform comments;
• guesses about an actor's identity, ethnicity, or birthplace;
• stage directions (such as when the script says (in Spanish)), English placeholder glosses, or another character "seeming to be able to follow."

Downgrading is allowed: when evidence becomes invalid, review opinions conflict with the body text, or identity / permission premises are overturned,actively downgradeand state the reason.Upgrading requires traceable new evidence or a real review record.

VI. Script-level execution rules

6.1 Two-axis dialogue (core method)

First decide the two axes separately, then write the lines:

• Information stance: state directly / partially state directly / strategically avoid.
• Syntactic state: complete / compressed / interrupted / self-corrected / silent.

The two axes are not bound to each other— direct can be very short, and indirect can also be syntactically complete. Crisis rescue, commands, work handoffs, boundary declarations, characters actively choosingcan be direct and complete; subtext is used only when the charactertruly hassomething to conceal, is probing, under politeness pressure, or facing relational risk.

Forbidden over-repairs (four kinds)：

1. Everyone speaks in fragments;
2. every line hides information;
3. every region is stuffed with slang;
4. every bilingual character randomly code-switches.

And the degradation in the opposite direction: writing English as regionless, professionless, relationship-difference-erasingneutral functional dialogue。

6.2 Style protection (must not be shattered by "natural colloquialization")

Stylized long sentences, polite threats, repetition, ritual forms of address, and comic timing may all be project choices. They must not be uniformly compressed into fragments just because "real people wouldn't always talk like that."

Complete sentences do not equal written language, and fragments do not automatically equal naturalness. The evaluation standard iswhether the rhythm and character function hold up, not "whether it sounds like a recording."

6.3 Do not infer permission from identity (this article relates to the entire process of creation, review, and repair)

Location, surname, appearance, ethnicity, age, profession, state/province / city, and kinship identity do not automatically generate：

• accent / dialect / film and television stereotyped vocabulary;
• code-switching (being able to speak two languages ≠ should speak the other one at this moment);
• keys, access rights, system permissions, device operation rights;
• professional qualifications, statutory authority to act.

Decisive action must have visible, explicit authorization, and cannot be filled in with vague off-screen instructions. If a task depends on access rights / knowledge / authorization that have not been given, let the character remain ignorant, switch to an action they can truly perform, orwrite the permission itself as the obstacle。

Criterion: the character knowledge in a line of dialogue must be traceable to "what he was present and saw / what someone explicitly told him / established canon." If none of the three apply, it isknowledge out of thin air。

6.4 Monolingual listener boundary (the safety line for multilingual projects)

Multilingual ability only shows "being able toswitch," not "should at this momentswitch." Every switch must answer five questions:

1. Who understands?
2. What triggers it?
3. What function does it serve?
4. Does it exclude someone present?
5. What consequences does it have for the relationship or the action?

Iron rule: when an English monolingual is present, key information involving theirmoney, goods, safety, responsibility, next actionmustimmediately return to the shared language, and the excluded person must have avisible reaction. When there is no genuine relational usage or corpus evidence, default to the task's target language.

6.5 Admission of local details

Place names, brands, sports teams, food, hot memesenter dialogue only when they affect choices, misunderstandings, leverage, or relationships.

A checklist-style local detail cannot prove localization. A place marker that can be replaced by any ordinary object cannot recur to prove the location; nor can the main conflict be resolved first and then used to create a smaller tail gag.Language adaptation does not add extra beats to demonstrate localness.

VII. Speech Rate and Capacity (Overseas Parameters)

The Chinese speech rate baseline does not apply to non-Chinese body text. Billing and duration estimates are converted according to the target language:

Language	Baseline speech rate	Notes
Chinese	3.5~6 字/秒	See master control short-drama-director/references/dialogue-speed-check.md（this engine does not modify it）
English	about 2.2~3.0 词/秒（≈130~180 wpm）	fast dialogue can reach 3.0 words/second;counted by words, not by characters
Other languages	convert according to the actual syllable density of that language	when there is no reliable conversion basis,mark as estimateand state the method

Back-calculating duration for English dialogue：镜长 ≥ 词数 ÷ 3.0 + 停顿补偿(pause compensation follows the master control parameters).Prohibitedto apply the Chinese characters/second formula directly to an English script.

Prohibited: converting lines into minutes; writing "exactly 90 seconds" without a table read. Word count is not duration—speech, action, and pauses can overlap and should be estimated according to the actual performance progression.

VIII. Rule Library Index

File	Content
references/language-contract.md	Language and style contract: item-by-item definition of the 14 fields + three-level status + stage-by-stage update permission table + cross-stage conservation
references/north-american-english.md	US and Canadian English dialogue adaptation: direct generation from scene / English surface landing details / dialect and multilingual switching / two original demonstrations
references/evidence-and-authority.md	Evidence layering (text / material calibration / human review / audience effect) + list of prohibited inferences + separation of permissions and self-reporting + monolingual listener boundary
references/overseas-platform-and-audience.md	Overseas platforms and audience contexts: T1 text / T2 product / T3 production three-layer separation + vertical-screen pre-check + sources and boundaries

Read on demand: read only the files needed for this task. The language contract and evidence boundaries should be read for almost every case; English adaptation is read only when the target language is English; platform context is read only when placement / aspect ratio / cost pre-check is involved.

IX. Cross-Stage Handoff (This Engine's Main Output)

When locking the script or handing off across stages,mustcarry:

language_status: region_unvalidated | evidence_calibrated | human_reviewed
coverage: 实际覆盖的国家 / 地区 / 社区
evidence_used: 实际读到的材料（虚构设定与现实证据分列）
prohibited_inference: 本次明确禁止的外推项
protected_style: 受保护的风格特征
multilingual_participants: 谁懂哪些语言、谁在场、关键事实必须让谁听懂
unresolved / local_review_needed: 未决项与待复核项

Downstream discipline：

• sd-asset-engine: character designs must not be stereotyped based on ethnicity / region; appearance is unrelated to the character's language ability.
• sd-storyboard-engine / sd-feed-engine: the dialogue field writes thetarget-language verbatim transcript; subtitles are marked separately and locked verbatim; speech rate is rechecked according to this engine's §VII parameters.
• sd-qc-engine: checkwhether the declaration matches the body text，do not unilaterally upgradelanguage status; the production stage must not rewrite the evidence level because "subtitles were scheduled" or "it reads smoothly."

X. Prohibition of Boundary Crossing

• Do notreplace the six-stage chain of sd-script-engine do notdo notestablish a separate script process;
• Do notestablish a separate feeding format, asset numbering, or machine-check script (authority lies with master control);
• Do notInventing dialects, accents, slang, or regional terms without evidence;
• Do notwrite model judgments, web retrieval, or dual-model agreement as human_reviewed；
• Do notinfer language ability, permissions, qualifications, or personality from identity labels;
• Do notbreak complete sentences that make a project choice valid, deadpan jokes, polite threats, or genre rhythm in pursuit of "naturalness";
• Do notextrapolate one platform's placement specifications into the format for all projects;
• Do notcross track boundaries—do not apply this engine to domestic drama tasks.