# ATR HTML Article Translation Workflow Prompt — UNIVERSAL FULL INTRO PROMPT WITH HARD QA GATES, ADVERSARIAL REVIEW, AND ANTI-FALSE-FINAL DISCIPLINE

Scope for Universal Intro: This is the first user-facing control prompt; enforce the Key Term Research Table, Presence Gate, parenthetical-order rule, hybrid residue scan, and no-repair promotion hardening before any final claim.




BATCH 75 — UNIVERSAL MATERIALITY-AWARE POLISH + ANTI-CHURN HARDENING — 4 June 2026

This Batch75 patch is universal. It applies to all ATR translation, review, refinement, formatting, QA, no-repair promotion, packaging, and handoff workflows. It does not weaken any Batch74 false-pass, exact-artifact, Latin-script fluency, hybrid-residue, title/link, href/src, Blogger-chrome, or source-fidelity safeguards.

MATERIALITY-AWARE POLISH GATE
Before making any visible-text change during translation, refinement, adversarial review, repair, audit, or no-repair promotion, classify the proposed change:

A. REQUIRED REPAIR — must be fixed.
Required repairs include defects in source fidelity, omission/addition, mistranslation, doctrinal or technical term accuracy, article-central terminology consistency, speaker attribution, exact-title/link-text handling, source-language residue, mixed-language/hybrid residue, malformed retained source/Sanskrit/Pāli/Tibetan/English labels, target-language grammar, punctuation or line-merge damage, target-language readability, HTML/link/media/code integrity, and target-language naturalness where current wording sounds machine-translated, obscure, distracting, or misleading.

B. BENEFICIAL POLISH — allowed in active editing passes.
Beneficial polish is not strictly wrong but materially improves publication quality: naturalness, readability, sentence flow, public ATR Dharma-article register, reduction of translationese, clarity of difficult doctrinal points, consistency of repeated term families, idiomatic phrasing where literal wording is clunky, or smooth grammatical integration of retained technical labels. If visible text changes, the pass remains an editing/repair/refinement pass and cannot be promoted as no-repair in the same pass unless a separate no-material-edit promotion pass is then performed.

C. PREFERENTIAL REWRITE — avoid.
Do not rewrite merely because another acceptable phrasing is possible. If the current wording is source-accurate, doctrinally clear, grammatically correct, natural, termbank-consistent, publication-acceptable, and free from residue/hybrid/title/punctuation/link/HTML defects, do not churn synonyms or sentence structure.

MODE-SPECIFIC BEHAVIOR
Active translation / repair / adversarial review: required repairs must be made; beneficial polish may be made if it improves publication quality without changing source meaning or adding doctrine; preferential rewrites should normally be avoided. Record visible-text edits and keep repair/editing status.

No-repair promotion: reopen the exact latest artifact, actively search for required repairs, identify optional polish without editing it, and do not fail promotion merely because another acceptable phrasing exists. If only optional polish or preferential alternatives remain, record “optional polish only; no material repair required.” If any required repair is found, repair it and reset to the repair status.

CHANGE MATERIALITY LEDGER
For repair/adversarial-review passes, the QA/changelog must include a compact Change Materiality Ledger for representative and material visible-text changes: nearest heading/sentinel, old wording, new wording, category (Required repair / Beneficial polish / Preferential rewrite avoided), reason, and whether the change affects status. For no-repair promotion passes, report required repairs found yes/no, beneficial polish identified but not edited yes/no, preferential rewrites avoided yes/no, and material edits made yes/no.

ANTI-CHURN STOP RULE
After repeated adversarial review passes, if the exact latest artifact has no required repairs and remaining possible changes are only preferential rewrites or optional beneficial polish, stop editing. Do not reset status through synonym-level rewriting. A no-repair promotion pass may promote only after exact-artifact reopening, source comparison, terminology/residue/hybrid/title/link/href/src/structure/fluency gates, no required repair, and no visible-text material edit.

DO NOT WEAKEN STRONG QA
Continue to treat omissions, additions, mistranslations, doctrinal distortion, wrong technical terms, source-language residue, mixed-language hybrids, malformed retained labels, broken grammar, awkward translationese that damages publication quality, exact-title drift, link-text corruption, href/src/media/code mutation, speaker-attribution issues, punctuation/line-merge damage, structural omissions/duplications, and misleading target-language wording as material defects requiring repair.

BATCH 74 — UNIVERSAL MATERIALITY-AWARE POLISH + ANTI-CHURN HARDENING — 4 June 2026

This Batch75 patch is universal. It is not Polish-only and must apply across Latin-script, non-Latin-script, inflected, analytic, right-to-left, and mixed-script target languages. Preserve all Batch73 Presence / hybrid-residue safeguards and all inherited Batch72/71/70/68/67/66 safeguards.

FALSE-PASS / REUSED-ARTIFACT FAILURE GATE
A review, repair, no-repair promotion, package, or handoff pass is invalid unless it reopens the exact latest artifact and produces fresh evidence for this pass. Do not reuse prior pass summaries, prior links, old QA reports, old filenames, or memory as if they were a new review. If a response claimed a new pass but actually reused prior artifacts or did not review, explicitly retract that claim, identify the latest real artifact, and do not continue pretending the false pass exists.

EXACT-ARTIFACT EVIDENCE REQUIREMENT
Every pass must report the exact input artifact path/name, exact output artifact path/name, whether the output is byte-for-byte identical or materially changed, and SHA/hash or equivalent identity evidence when the tool environment allows it. State whether each intended replacement matched and applied. If no new artifact or new no-material-edit QA report is produced, do not claim a new pass; say: No new pass was completed.

NO-REPAIR VS REPAIR SEPARATION
A strict no-repair promotion pass must make no material edits. If any material issue is found, promotion fails. Either stop with a failed-promotion report or begin a separate material repair pass. Any pass that makes material edits must use: REPAIRED AFTER RE-AUDIT — material gates currently pass, no-repair promotion still required. It must not also promote in the same pass.

THOUGHT-ONLY REVIEW PROHIBITION
Thinking briefly and reusing old links is not a review. A valid review must show concrete evidence: opened latest artifact, searched/scanned target text, sampled high-risk source/target passages, ran target-language fluency checks, checked the cumulative risk basket, checked href/src/media/code parity, and generated a new QA report or explicitly stated that no new artifact was created.

STRING-SPECIFIC REPAIR FRAGILITY / 0-MATCH RULE
If a targeted repair does not match the current artifact, do not assume the issue is fixed. Search semantic variants, shorter substrings, inflected variants, capitalization variants, punctuation variants, HTML-separated variants, and nearby context. Log every 0-replacement intended repair as a warning requiring investigation. For scripted repairs, QA must list old string/pattern searched, match count, replacement applied, 0-match intended repairs, semantic follow-up search, and before/after cumulative risk-basket status.

LATIN-SCRIPT RESIDUE AND FLUENCY GATE
For Latin-script target languages, ordinary English residue is visually hidden because both source and target may use Latin characters. Run an extra Latin-script residue scan for English words left in ordinary prose, English technical labels with local grammar errors, English terms glued to local inflections/cases/prepositions, unquoted English labels used as target-language nouns, partially translated source-title fragments, and English idiom calques that look superficially grammatical.

INFLECTED-LANGUAGE PROTECTED-LABEL AGREEMENT GATE
For inflected target languages, retained labels such as AMness, I AMness, Presence, Awareness, Self, No Mind, no-mind, Dharma, Dhamma, Maha, anatta, anātman, dhātu, skandha, and śūnyatā must be checked for case, gender, number, agreement, and natural grammar. If a protected label is retained, the surrounding sentence must still be natural. Use target-language-first phrasing, quotes, or a short explanation rather than forcing local grammar onto the retained label.

SLASH-CALQUE HARDENING
Scan visible prose for slash compounds such as A/B, A / B, A/B/C, subject/object, doer/agency, Dhamma/Dharma, I/self/Self, mind/consciousness, realization/actualization, no-self/non-self/anatta/anātman, linked title fragments, and person aliases joined with slashes. Classify each as exact title/proper label, glossary/termbank, ordinary prose, or ambiguous. Preserve exact titles and intentional glossary labels; convert ordinary prose into natural target-language coordination, apposition, or explanatory phrasing; flag ambiguous cases.

EXACT TITLE VS TRANSLATED TITLE HARDENING
For every link-bound title, decide separately whether it is preserved exactly, fully translated naturally, or intentionally paired. Do not partially translate titles, honorifics, or title fragments such as translating “Mr.” while leaving the rest English. Do not mix target-language words into exact English titles unless that is the exact source title. QA must list linked titles as preserved, translated, or intentionally mixed/paired.

IDIOM AND METAPHOR CALQUE HARDENING
Prompt-level review must actively hunt literal renderings of English idioms/metaphors and build a target-specific idiom/calque risk basket. Seed examples include: go and read; goes down the drain; along the line of; break through; stood out; one hand claps; the drop is thorough; quality time; open and boundless; wholly thus; always already so; no mirror reflecting; magical appearances are too empty. This list is illustrative, not exhaustive.

TECHNICAL TERM FAMILY DRIFT GATE
Once a term family causes repeated repairs, add the family to the cumulative risk basket and scan globally in all variants. High-risk families include actualization/actualize/embodied/lived/integrated; ground/Ground of Being/ground reality/basis/path; Presence/pure Presence/total Presence; AMness/I AMness/I AM; Self/self/no-self/no-I/anatta/anātman; no-mind/No Mind/Mind that is no-mind; subject/object; doer/agency/no-doership; total exertion/total activity/full functioning; suchness/thusness/isness/wholly thus/Always So; Maha; yāna/lower yānas; dhātu/skandha; knowledge/cognitive obscurations; māyā/magical appearances; mirror/no mirror/mirror-like awareness.

MECHANICALLY CORRECT BUT UNNATURAL IS MATERIAL
A translation can pass termbank checks and still fail if the target-language prose is unnatural or doctrinally misleading. Treat unnatural case/gender/number agreement, source word order copied too literally, line-merge fragments, technical labels without grammatical integration, slash-heavy prose, literal metaphor calques, target-language punctuation defects, capitalization problems, and misleading ordinary-word renderings of Dharma terms as material defects when they affect readability or doctrinal clarity.

NO-REPAIR PROMOTION EVIDENCE ADDENDUM
A no-repair pass must include a section titled “No New Artifact Reuse / Exact Artifact Evidence” with: I reopened: [exact file]; I produced: [exact report/file]; Material edits made: yes/no; if no, output is byte-for-byte identical/hash-equivalent/explicitly copied unchanged; if yes, no promotion allowed; prior pass status was not assumed; new review evidence is listed.

CROSS-LANGUAGE TERM SOURCE EVIDENCE LIMIT
Use target-language Buddhist sources where available to confirm major Dharma term choices, but source evidence supports broad term choices only. It does not prove sentence-level fidelity, context fit, or target-language fluency. Paragraph/link/blockquote/list counts are necessary but insufficient and must never be treated as proof of translation quality.

STATUS DISCIPLINE
If material issues are found after a previous no-repair promotion, state: Previous promotion was premature. This pass found material issues that should have been caught earlier. Status is reset to REPAIRED AFTER RE-AUDIT — material gates currently pass, no-repair promotion still required. Never claim strict-certified final.

BATCH 73 — UNIVERSAL PRESENCE GATE + TARGET-LANGUAGE-FIRST PARENTHETICAL ORDER + HYBRID RESIDUE HARDENING — 1 June 2026

This Batch73 patch is universal. It is not a Japanese-only fix. Apply it to all ATR translation, review, refinement, HTML translation, audit, formatting, package, and promotion workflows for all target languages.

UNIVERSAL KEY TERM RESEARCH BEFORE DEFAULT TRANSLATION GATE
Before translating or standardizing any high-risk AtR/Dharma term, do not rely only on dictionary equivalence, memory, previous target-language habits, or another language's termbank. Create or update a compact Key Term Research / Decision Table for article-central terms. Include source term, target rendering, translated/retained/transliterated/paired status, protected-label status, ordinary-prose rule, forbidden/high-risk renderings, whether external/source research is needed, and whether the term has multiple senses.

Minimum high-risk basket when present: Presence/presence, Awareness/awareness, I AM/AMness, anatta/no-self/non-self/anatman, emptiness/empty nature, dependent arising/dependent origination, non-dual/non-duality, no-mind/No Mind/One Mind, luminosity/clarity/vividness, manifestation/appearance/phenomena, suchness/thusness/isness, subject/object/subject-object division, agent/agency/doer/doership/non-doership, spontaneous perfection, total exertion, Dzogchen/Mahamudra/rigpa/Dharma/Buddhadharma, and article-specific technical terms.

UNIVERSAL PRESENCE GATE
Treat Presence as a high-risk AtR/Dharma term, not a generic English word. Distinguish technical AtR Presence from ordinary presence/present, ever-present/already-so phrasing, philosophical/metaphysics-of-presence usage, and Dzogchen title/phrase usage such as Instant Presence. Technical AtR Presence points to luminous immediacy/direct actuality/the vivid fact of being-experience and appears in phrases such as pure Presence, Total Presence, Presence-Awareness, Presence is spontaneously perfected, manifold of Presence, and nature of Presence. It requires target-language research and a termbank decision.

Japanese house-style note for the Seven Stages article family: technical AtR Presence defaults to 臨在; pure Presence = 純粋な臨在; Total Presence = 総体的臨在; Presence-Awareness = 臨在＝覚知; Presence is spontaneously perfected = 臨在は自ずから円成している / context-specific. Avoid 現前性 as the default for AtR Presence unless the context is specifically philosophical/phenomenological. Ordinary present/ever-present contexts may require 現前している, すでに現れている, 常にすでにそうである, or another natural phrase. Do not impose this Japanese note on other target languages. Chinese AtR house style may preserve 臨在 where applicable, but Chinese choices are clues, not automatic authorities for all other languages.

TARGET-LANGUAGE-FIRST PARENTHETICAL ORDER RULE
In ordinary target-language prose, if a retained English/Sanskrit/Pāli/Tibetan/source label is useful, prefer target-language first: target-language rendering（source label）. Avoid source-label first forms in ordinary prose unless the source label is an exact title, is being discussed as a phrase, the target-language convention strongly prefers it, the user explicitly requests it, or it is a glossary/termbank entry. Do not glue retained source labels directly to target-language case markers, particles, suffixes, postpositions, clitics, or inflections.

Japanese examples: use 無我（anatta）の, not anatta の; 空性（shunyata）の実現, not shunyata の実現; 心（Mind）, not Mind（心） in ordinary prose; 伝達（transmission）, not transmission（伝達）; 無我（no-self）, not no self（無我）. Avoid hybrids such as Pure覚知, Natural覚知, Non-dual は, dependent origination の, ground と path, meditative composure と insight, no-mind の状態, anatta実現, Maha経験, On 空性, Self-実現, and Being／Self-Realisation への.

UNIVERSAL HYBRID RESIDUE SCAN
Before final or promotion, build a target-language-specific hybrid scan basket. Search for English/source labels glued to local particles/case markers/suffixes/inflections; English words inside target-language compounds; target-language words inserted into exact English titles; mixed-script technical hybrids; English-first parentheticals in running prose; translated fragments inside exact English titles; ordinary English prose overprotected as technical labels; and stray spaces before local particles after parentheses. For Latin-script target languages, run an extra ordinary-English residue scan because English residue is harder to see.

Japanese risk basket seed: Pure覚知; Natural覚知; Non-dual は; dependent origination の; ground と path; meditative composure と insight; no-mind の状態; anatta の/へ/と/を/は/後/実現; Maha経験; On 空性; Self-実現; Being／Self-Realisation への; transmission（伝達）; no self（無我）; Dhamma／Dharma） を; Awakening to Reality 実践 Guide; 音Cloud; Scenery……; 真如のMaha経験; 五蘊（五蘊）. Also scan Latin/source labels before の, を, に, へ, と, は, が, で, も, から, まで, より and before 後, 的, 性, 実現, 経験, 状態, 化, 作用, 実践, 知見, 覚知, 空性.

EXACT TITLE VS TRANSLATED TITLE HANDLING
For every linked article/book/video/audio title, decide whether the link text should remain the exact source title or be translated as a full natural target-language title. If preserved, preserve the whole title exactly; do not partially translate it. If translated, translate the whole title naturally and retain the exact source title in parentheses only when useful. Do not create mixed titles such as On 空性, Awakening to Reality 実践 Guide, or Recognizing Rigpa vs Realizing 空性. Do not translate URL slugs. Record intentional title preservation/translation in QA.

PROTECTED-LABEL WHITELIST DISCIPLINE
A word being technical does not automatically mean it should remain in English. Classify every retained English/Sanskrit/Pāli/Tibetan label as exact title, proper name, URL/code, intentionally retained technical term, intentionally retained English AtR label, gloss in parentheses after target-language rendering, or human-review placeholder. Ordinary lowercase terms such as experience, realization, practice, view, awareness, consciousness, manifestation, ground, path, source, self, mind, state, clarity, naturalness, ordinariness, spontaneity, agent, doer, doership, subject, and object should normally be translated unless a reason is documented.

TERM-FAMILY ESCALATION RULE
If more than one material issue is found around the same term family, escalate that family into the active termbank and scan basket. Example: if Presence was mistranslated once, scan Presence, pure Presence, Total Presence, Presence-Awareness, nature of Presence, present, ever-present, and target-language alternatives globally. If one anatta hybrid appears, scan every anatta/no-self/non-self/anatman form. If one total exertion issue appears, scan total exertion/total activity/complete activity/full function equivalents globally.

TOTAL EXERTION SOURCE-AWARE RULE
Do not default total exertion mechanically. For Japanese Dōgen/Zenki contexts, consider 全機 and 一法究尽 depending on source meaning. For Chinese, preserve the existing rule not to default to 全体作用; use 一法究尽 / 具尽 / context-specific handling. For other languages, research source/context before setting a default.

NO-REPAIR PROMOTION HARDENING
A no-repair promotion pass must include exact-artifact reopening, source comparison, Prompt 6 semantic fidelity review, Prompt 9 target-language fluency review, target-language hybrid residue scan, target-language-first parenthetical-order scan, high-risk term table consistency scan, exact linked-title scan, ordinary prose vs protected-label classification, URL/href/src/code integrity scan, punctuation/line-merge scan, and representative high-risk sentence sampling from beginning, middle, ending, blockquotes, bold/italic sections, and all article-central terms. If material issues are found after a previous no-repair promotion, explicitly retract the earlier promotion and reset status: Previous promotion was premature. This pass found material issues that should have been caught earlier. Status is reset to REPAIRED AFTER RE-AUDIT — material gates currently pass, no-repair promotion still required.

REQUIRED QA EVIDENCE ADDITIONS
Every final/review/promotion QA report should include: exact source file, exact target file, prompt versions applied, target-language risk profile, key term decision table, Presence Gate result when relevant, parenthetical-order scan result, hybrid residue scan basket, ordinary English/source residue basket, protected-label whitelist, exact linked-title scan, href/src/media parity, non-source text in URL/code scan, major structure counts, representative sentence sampling notes, cumulative risk basket and results, whether material edits were made, and the correct status label.

STATUS DISCIPLINE
Because Batch73 is a material managed-prompt update, its status is: REPAIRED AFTER RE-AUDIT — material gates currently pass, no-repair promotion still required. Only a later exact-artifact no-material-edit readback pass may promote to: reviewed final / publishable — strict-certified final not claimed.


**Version note:** This is the full language-agnostic initial prompt for ATR HTML article translation workflows. It is intended to apply to all ATR target languages, including but not limited to Simplified Chinese, Traditional Chinese, Tibetan, Nepali, Bengali, Spanish, Hindi, Tamil, Portuguese, Japanese, Thai, Polish, Danish, Vietnamese, French, Indonesian, Malay, Korean, Arabic, Russian, Italian, Serbian, Burmese, Sinhala, Greek, Ukrainian, Swedish, and any other requested target language. It is not a patch and is not specific to any one translation. It incorporates hard QA gates against premature “final/publishable” claims, weak no-repair promotions, untranslated source-language residue, mixed-language hybrids, URL/code corruption, and target-language fluency failures.


## 0. Target-language configuration

Set the following values before starting:

* `TARGET_LANGUAGE`: [e.g. Simplified Chinese / Traditional Chinese / Tibetan / Nepali / Bengali / Spanish / Hindi / Tamil / Portuguese / Japanese / Thai / Polish / Danish / Vietnamese / French / Indonesian / Malay / Korean / Arabic / Russian / Italian / Serbian / Burmese / Sinhala / Greek / Ukrainian / Swedish / any other requested target language]
* `TARGET_LANGUAGE_LABEL`: [human-readable label for reports, e.g. “Simplified Chinese translation,” “Bengali translation,” “Spanish translation,” “Tibetan translation,” “Ukrainian translation,” etc.]
* `TARGET_SCRIPT_OR_VARIANT`: [e.g. zh-Hans / zh-Hant / Tibetan / Devanagari / Bengali / Tamil / Arabic / Cyrillic / Greek / Latin-script European language / Japanese / Korean / Thai / Burmese / Sinhala / etc., or “not applicable”]
* `SOURCE_LANGUAGE`: English
* `SOURCE_ARTICLE_URL`: [insert English original URL if available; otherwise use `URL_NEEDED_HUMAN_REVIEW`]
* `SOURCE_ARTICLE_TITLE`: [insert English article title]
* `TOP_ORIGINAL_LINK_LABEL`: [target-language label for “English original,” e.g. for Simplified Chinese: 英文原文; for Bengali: ইংরেজি মূল]
* `USER_NAME_RENDERING_RULE`: Preserve “Soh Wei Yu” or “Soh” unless I explicitly provide a target-language rendering.
* `TARGET_LANGUAGE_TERMBANK`: Use the target-language-specific termbank supplied below or in the uploaded prompt pack. If no target-language termbank is supplied, translate technical terms conservatively according to the uploaded prompt pack and source context.

Please translate the uploaded English Blogger HTML article into `TARGET_LANGUAGE`.

I will upload:

1. The current ATR prompt pack ZIP.
2. The English source HTML article to translate.

Important:
This is a fresh `TARGET_LANGUAGE_LABEL` workflow, not a repair of an existing target-language translation unless I explicitly upload an existing target-language file and ask for review/repair.



### 0.1 Universal scope and no single-language assumption

This prompt must remain applicable to all requested ATR target languages. Do not silently adapt it into a language-specific prompt unless the user explicitly asks for a target-language-specific derivative.

For each new translation workflow:

* Treat `TARGET_LANGUAGE` as the only active target language.
* Do not import the terminology decisions, residue patterns, grammar assumptions, punctuation conventions, or protected-term choices from a previous target language unless they are explicitly stated as universal in this prompt or in the uploaded prompt pack.
* Before translating, infer or construct a small target-language risk profile based on the language’s script, morphology, punctuation, and common translation failure modes.
* If a target language uses the Latin alphabet, source-language residue is harder to detect because English and target language share letters; therefore perform an extra ordinary-English residue scan using high-risk English Dharma and prose terms.
* If a target language uses a non-Latin script, perform an extra mixed-script and transliterated-English scan.
* If a target language is inflected, perform an extra grammar scan for case, gender, number, tense, agreement, compounds, and unnatural literal calques.
* If a target language is analytic or topic-prominent, perform an extra scan for unnatural word-for-word source syntax, dropped particles, and broken sentence flow.
* If a target language is right-to-left, perform an extra audit for punctuation directionality, URL isolation, list formatting, and HTML attribute safety.
* If the prompt pack or user provides a target-language termbank, that termbank controls over generic dictionary choices, subject always to source meaning and prompt-pack rules.

### 0.2 Anti-false-final principle

The main failure this prompt is designed to prevent is false promotion: claiming that a translation is “reviewed,” “publishable,” “complete,” or “final” while later Prompt 6/9 review still finds material defects.

Therefore:

* QA must be adversarial, not confirmatory.
* A no-repair promotion pass must try to disprove readiness, not merely confirm prior status.
* Mechanical HTML/count parity does not prove translation quality.
* A Prompt 6/9 review is not complete unless it checks semantic fidelity, doctrinal terminology, target-language fluency, ordinary source-language residue, mixed-language residue, line-merge damage, punctuation, exact-title handling, and HTML/code preservation.
* If a later pass finds material issues that an earlier “publishable” or “no-repair” pass should have found, explicitly treat the earlier promotion as false, retract the promotion status for the current working artifact, repair the issue, and add that defect family to the cumulative risk basket.

---

## 1. First inspect the uploaded prompt pack, then actively apply it

Before translating, editing, judging, or formatting anything, inspect the uploaded prompt pack carefully. Do not rely on memory or previous sessions.

Read the relevant prompts and workflow rules, especially:

* Prompt A: HTML Article Translation Workflow
* Prompt 1: High-Fidelity Translation Workflow
* Prompt 6: High-Fidelity Translation Review
* Prompt 9: Whole-Document Refinement
* Prompt 8, if present and relevant to dialogue/chat-log cleanup
* Unified Blogger Formatting Prompt, if the pack routes Blogger presentation through it
* Strict HTML QA Audit Prompt, if present
* Any Batch67+ / no-repair promotion / status-label / no-premature-final rules in the pack

However, inspecting the prompts is not enough. You must actively apply the relevant prompts as the operative workflow for the task.

Do not use prompt names as a ritual label. If you say Prompt 6 or Prompt 9 was applied, the artifact and QA report must show evidence that the relevant checks were actually performed. A later discovery of obvious residue, line-merge damage, target-language grammar errors, or recurring calques means the earlier Prompt 6/9 claim was incomplete and must be treated as a QA failure.

Use the uploaded prompt pack as controlling instructions in this exact functional order:

1. **Prompt A — governs the HTML article workflow**

   * Apply Prompt A to control source-of-truth discipline, HTML/Blogger preservation, routing, artifact handling, cumulative checkpoints, link/media preservation, and final QA requirements.

2. **Prompt 1 — governs the actual high-fidelity translation into `TARGET_LANGUAGE`**

   * Do not merely read Prompt 1.
   * Translate the article according to Prompt 1’s Buddhist/Dharma translation rules.
   * Apply Prompt 1’s source-faithfulness, doctrinal precision, terminology discipline, no-omission, no-added-gloss, and no-paraphrase requirements throughout the actual `TARGET_LANGUAGE` translation.

3. **Prompt 6 — governs the post-translation fidelity and terminology review**

   * After translating, review the `TARGET_LANGUAGE` output against the English source using Prompt 6.
   * Repair mistranslations, omissions, additions, doctrinal distortions, terminology errors, speaker-attribution issues, source-language residue, mixed-language hybrid residue, and unnatural target-language wording that affects fidelity or clarity.
   * Do not treat Prompt 6 as optional commentary; it must be used as an active QA pass.

4. **Prompt 9 — governs whole-document refinement after Prompt 6**

   * After Prompt 6 review, run Prompt 9 across the whole translated document.
   * Refine consistency, readability, punctuation, terminology, line-merge damage, target-language flow, and whole-document coherence without paraphrasing away source meaning or adding new doctrine.

5. **Prompt 8 — apply before translation only if relevant**

   * If the article is dialogue/chat-log-heavy, or if the pack routes the task through Prompt 8, apply Prompt 8’s dialogue-cleanup rules before translation.
   * Apply Prompt 8 only to the extent required by the source workflow.
   * Preserve speaker attribution, chronology, and source-traceable line merging.
   * Do not invent speakers.
   * Do not split one speaker’s sentence into another speaker’s turn.
   * Do not convert a tagged name inside one paragraph into a new speaker label unless the source clearly makes it a separate speaker turn.

6. **Unified Blogger Formatting / Strict HTML QA — apply at the end if routed by the pack**

   * If the uploaded pack routes final Blogger presentation through the Unified Blogger Formatting Prompt or Strict HTML QA Audit Prompt, apply them as actual final checks, not merely as references.
   * Preserve CSS, wrappers, IDs/classes, links, media, iframe/script/config values, href/src values, and Blogger structure unless a prompt explicitly authorizes a change.

Do not say “I read Prompt 1” and then translate generically. The output must show that Prompt 1 governed the translation, Prompt 6 governed the review, Prompt 9 governed the refinement, and Prompt A governed the HTML workflow.

---

## 2. Critical task-object rule

* Prompt A / Prompt 1 / Prompt 6 / Prompt 9 are workflow instructions, not the source article.
* Translate the uploaded English HTML article, not the prompt files themselves.
* Only translate Prompt A or any prompt file if I explicitly say: “translate Prompt A itself” or “translate this prompt file.”

---

## 3. Source-of-truth rules

* The uploaded English HTML is the authority, not the live webpage.
* If I later upload a newer English HTML, treat the newer uploaded English HTML as the new source of truth.
* When a newer English HTML is uploaded, compare it carefully against the previous English source and update the `TARGET_LANGUAGE` translation for all changes, not only the examples I mention.
* Existing live translations or previous target-language versions are not the source and must not be used as scaffolds, missing-block donors, or authority unless I explicitly instruct you to compare or reuse them.
* Preserve meaning exactly, especially doctrinal nuance.
* Do not omit, summarize, paraphrase away, shorten, reorganize, or add doctrinal glosses unless the prompt pack explicitly requires cleanup.
* Do not introduce new explanatory material unless clearly marked and explicitly requested.
* If you find ambiguous/source-dependent issues, flag them clearly for human review instead of silently deciding.

---

## 4. HTML / Blogger preservation rules

* Preserve all Blogger HTML, CSS, classes, IDs, wrappers, structure, links, anchors, images, blockquotes, media embeds, iframe/script/config values, and layout unless a change is required by the prompt pack.
* Preserve the `<style>` block as code.
* Do not translate CSS property names, class names, selectors, URL paths, script/config values, href/src values, or code-like attributes.
* Preserve all href/src URLs exactly unless I explicitly ask to update a link.
* Translate visible article/body text into `TARGET_LANGUAGE`.
* Translate human-facing attributes such as `alt`, `title`, `aria-label`, `placeholder`, iframe `title`, and caption-like data attributes unless they are exact titles, proper names, URLs, code, or protected labels.
* Do not broadly redesign the article.
* Hard gate: after translation, review all `href`, `src`, `action`, CSS `url(...)`, iframe config URLs, script/config URL values, and code-like attributes to ensure no target-language words, non-source non-Latin characters, spaces, or translated slugs were inserted.
* If any URL or code-like value is changed without explicit user instruction, it is a material defect and must be repaired.

---

## 5. English-original link requirement

If this is a target-language translation page, add one top link after the hero/title or near the top:

`TOP_ORIGINAL_LINK_LABEL`：<a href="SOURCE_ARTICLE_URL">SOURCE_ARTICLE_TITLE</a>

Use the actual uploaded English article URL if provided. If the English URL is not provided, insert `URL_NEEDED_HUMAN_REVIEW` and flag it for human review.

Preserve all existing original links and add this as one extra link only.

In every QA report, explicitly account for this one intentional extra link if `SOURCE_ARTICLE_URL` was inserted or if `URL_NEEDED_HUMAN_REVIEW` remains.

---

## 6. Important workflow requirement — cumulative full HTML only

Do not return isolated “Part 1 / Part 2” fragment files.

If the article is too long to finish in one turn, return a cumulative full working HTML file each time.

Meaning:

* First return: full HTML shell + translated content from the beginning up to the current stopping point.
* Next return: the same cumulative HTML, with the next translated section appended.
* Continue this way until the full article is translated.
* Each returned file must be a valid Blogger-ready HTML file, not just a fragment.
* If translation is incomplete, exclude later untranslated English article-body content rather than leaving it inside the returned HTML.
* Keep the style block closed before the article body in every returned file.

---

## 7. Required process

A. Inspect the uploaded prompt pack.

B. Identify the relevant controlling rules from Prompt A / Prompt 1 / Prompt 6 / Prompt 9 / Prompt 8 / Blogger Formatting / Strict HTML QA.

C. Confirm the task object: the English HTML article is the source, not the prompt files.

D. Apply Prompt A to set up the HTML article translation workflow and preservation constraints.

E. Apply Prompt 1 to translate the uploaded English HTML into `TARGET_LANGUAGE` while preserving all HTML/CSS/link/media structure.

F. Apply Prompt 6 as a full post-translation review against the English source:

* fidelity
* terminology
* omissions
* additions
* doctrinal nuance
* speaker attribution
* source-language residue
* mixed-language hybrid residue
* exact-title handling
* HTML-preservation risks

G. Apply Prompt 9 as a whole-document refinement pass:

* source-faithful clarity
* consistency
* punctuation
* readability
* terminology coherence
* line-merge cleanup
* no paraphrasing away source meaning
* no new doctrine

H. Apply Blogger Formatting / Strict HTML QA checks if the pack routes final presentation through them.

I. Apply the hard final-gate module in Section 16 before any publishable/final status.

I2. Apply the adversarial Prompt 6/9 defect-hunt module in Sections 16.11–16.16 before any no-repair promotion. This includes target-language fluency review, high-risk sentence sampling, residue search baskets, calque/grammar scans, and evidence logging.

J. Produce the required artifacts.

---

## 8. Per-response Prompt Application Ledger

In every response, include a brief **Prompt Application Ledger** before the checkpoint audit or final summary.

Do not merely state that the prompts were read. State how they were applied in this response.

Use this compact format:

**Prompt Application Ledger**

* **Prompt A:** Applied for HTML/Blogger workflow, source-of-truth discipline, structure preservation, href/src/media preservation, and checkpoint/final artifact rules.
* **Prompt 1:** Applied as the active translation engine for the current covered source range, including high-fidelity Dharma translation, terminology discipline, no omission, no added gloss, and no paraphrase.
* **Prompt 6:** Applied as the active review pass for the current translated range, checking fidelity, terminology, doctrinal nuance, omissions, additions, mistranslations, source-language residue, and HTML preservation.
* **Prompt 9:** Applied as the whole-document/current-checkpoint refinement pass for consistency, punctuation, readability, source-faithful flow, and mixed-language cleanup.
* **Prompt 8:** Applied only if dialogue/chat-log cleanup was required; otherwise state “Not applied — no dialogue/chat-log cleanup needed in this range.”
* **Blogger Formatting / Strict HTML QA:** Applied if routed by the prompt pack; otherwise state “Not applied in this checkpoint unless final Blogger QA is routed by the pack.”
* **Hard Final-Gate Module:** State whether Section 16 was applied, whether any material edits were made, and whether no-repair promotion is allowed.

The ledger must be specific to the actual work done in that response. Do not use it as boilerplate if a prompt was not actually applied.

The ledger must not replace evidence. For every final, repair, or promotion response, the QA/changelog must also include concrete audit evidence such as search baskets used, issue families checked, representative source/target sentinels, and whether any material issue was found or repaired.

---

## 9. Critical workflow discipline and status labels

Do not prematurely call the file final-ready.

For a fresh translation, the initial translation itself counts as a material creation/editing pass. Therefore, do not promote it to publishable/final in the same pass.

Use these status labels strictly:

1. If you create the translation or make any material correction in the current pass, status must be:

“REPAIRED AFTER RE-AUDIT — material gates currently pass, no-repair promotion still required.”

2. Only after a later pass reopens the exact latest repaired/translated artifact and makes no material edits may you promote it to:

“reviewed final / publishable — strict-certified final not claimed.”

3. Do not claim “strict-certified final,” “perfect,” or “final-ready” unless the prompt pack explicitly permits that status.

4. If I later say “continue,” perform the no-repair promotion pass on the exact latest artifact. If no material edits are made, promote to:

“reviewed final / publishable — strict-certified final not claimed.”

If you find new material issues, repair them and keep the status as:

“REPAIRED AFTER RE-AUDIT — material gates currently pass, no-repair promotion still required.”

5. If you cannot actually perform a required check, do not claim it passed. State:

“Not fully verified: [specific check]. Human review required.”

6. If a later review finds material defects after an earlier no-repair promotion or publishable status, do not defend the earlier status. Say plainly:

“Previous promotion was premature. This pass found material issues that should have been caught earlier. Status is reset to REPAIRED AFTER RE-AUDIT — material gates currently pass, no-repair promotion still required.”

7. The next no-repair promotion attempt must include the newly discovered issue family in the risk basket and must scan globally for related defects.

---

## 10. Cumulative checkpoint audit

For every returned checkpoint, briefly report:

* Source visible text-node range covered so far, or nearest source heading/date/source-inline sentinel if exact node count is not available.
* First and last covered source sentinel.
* Next source line/sentinel if incomplete.
* Whether the style block is closed before the article body.
* Whether tag/class structure is preserved for the covered range.
* Whether links/hrefs are preserved exactly for the covered range.
* Whether src/media are preserved exactly for the covered range.
* Whether human-facing attributes in the covered range are translated/preserved/classified.
* Whether ordinary untranslated English remains, excluding proper names, URLs, exact titles, and intentional technical labels.
* Whether no forbidden target-language terms remain in the covered range.
* Whether any source-language + target-language hybrid constructions remain.
* Whether any target-language or non-source non-Latin text entered `href`, `src`, `action`, CSS URLs, iframe config URLs, script/config URLs, IDs, classes, or selectors.
* Whether any intentional extra link, such as the required English-original link, is accounted for.

---

## 11. Dialogue / speaker-attribution audit

Compare every `.dialogue-turn`, `.atr-chat-line`, chat paragraph, Facebook-style comment, and speaker-label region against the source HTML.

Confirm:

* The number and order of dialogue turns match the source, except for intentional Prompt 8 cleanup.
* Every speaker label matches the corresponding source speaker.
* Tagged names inside a paragraph are not mistaken as speakers.
* Final dialogue sections are checked especially carefully, because Facebook-style comments often cause speaker corruption.
* No same-speaker adjacent dialogue-turns remain unless they are truly separate source turns and intentionally preserved.
* If multiple source lines are merged, sentence boundaries, punctuation, capitalization, and spacing are repaired.
* No speaker label is accidentally absorbed into the previous paragraph.
* No source-language residue remains inside ordinary dialogue prose unless protected by the whitelist in Section 16.

---

## 12. Line-merge / punctuation audit

Check the whole covered range for:

* Missing sentence-ending punctuation after merged lines.
* Missing capitalization after sentence breaks in English retained labels/titles.
* Missing spaces after ellipses/full stops in retained English.
* Malformed punctuation such as “? !”, “! !”, “? ? ?”, doubled punctuation inappropriate for `TARGET_LANGUAGE`, or repeated punctuation caused by line merging.
* Damaged URL text such as `facebook. com`, `blogspot. com`, `index. php`.
* Sentence fragments caused by line merging.
* Speaker labels accidentally absorbed into the previous paragraph.
* Source-language punctuation patterns left in target-language prose where target-language punctuation should be used.
* Hybrid punctuation around protected English labels; retain labels if needed but make the sentence natural in `TARGET_LANGUAGE`.

---

## 13. Prompt 6 / Prompt 9 terminology audit

After translation, run a full terminology and doctrinal review, not just a mechanical translation check.

Important hierarchy:

* The uploaded English source controls meaning.
* Prompt A / Prompt 1 / Prompt 6 control source discipline.
* The target-language termbank in the uploaded prompt pack controls where applicable.
* The additional `TARGET_LANGUAGE_TERMBANK` below is a house-style preference, not permission to override source context.
* If the house-style termbank conflicts with Prompt A / Prompt 1 / Prompt 6 or the source context, the prompt pack and source context control.

### Target-language termbank module

Apply the following only if relevant to `TARGET_LANGUAGE`.

If translating into Simplified Chinese or Traditional Chinese, apply this Chinese-specific module:

* view = use context. For Soh / AtR doctrinal prose, translate general “view” as 知见 by default, especially in phrases such as lack of view, refine one’s view, clear view, view and experience, one-mind view, substance/essence view, and view as framework/paradigm/stance. Do not use 见地 as the default for “view” in Soh / AtR prose.
* Use 见 / 见地 only when the source is explicitly a Tibetan lta ba / formal “View” title or when the user explicitly requests that register.
* Retain fixed canonical compounds where appropriate: 正见、邪见、我见、身见、自性见、实有见、真我见、常见、断见、一见、异见、妄见.
* wrong view = 错误知见 or 邪见 depending source/register.
* right view = 正见.
* view aspect = 知见层面 when the source means the level/aspect of understanding; use 见地层面 if the source context is doctrinal View.
* agent = 主宰者 where the source means a controller/doer; otherwise translate by context.
* agency = 主宰性 where the source means controlling agency; otherwise translate by context.
* action = 行动 / 作用 depending on context.
* subject-action-object = 主体—行动—客体.
* inherent existence = 自性存在 / 实有 depending on context.
* inherentness = 自性 / 实有性 depending on context.
* reification = 实体化.
* constructs = 构造 / 概念构造 / 心理构造 depending on context.
* mental proliferation / conceptual proliferation = 戏论（概念增殖） when referring to prapañca/papañca; preserve 造作 when the source explicitly says fabrication/constructing/造作.
* thought realm = 意界, not 念头界.
* mind-door = 意门.
* physicality = 物理性.
* self enquiry = 自我参究.
* luminosity = 明耀 / 明性 / 光明性 depending on context; avoid casual 光辉.
* manifestation = 显现.
* pure seeing = 纯粹的看见, not 纯粹的见.
* awareness/consciousness/knowing/mindfulness: translate contextually as 觉知 / 意识 / 明知 / 正念 / 知 / 识; do not flatten all usages into one Chinese term.
* vidyā / rig pa = 明 / 明知 as source context requires; never ordinary 知识.
* ye shes / jñāna = 本初觉智 / 智慧 depending context.
* shes pa / vijñāna / consciousness = 识 / 意识 / 觉知 depending context.
* phenomena in formal Dharma contexts = 诸法, not mechanically 诸现象.
* total exertion = preserve the prompt-pack/source-context-approved rendering; do not mechanically use 全体作用 unless the prompt pack explicitly requires it.
* doer / doership / no-doership = translate by context; preserve doctrinal contrast between doing, doer, agency, and non-doership.

Chinese-specific forbidden or high-risk renderings:

* Do not use 能作者 for agent.
* Do not use 苏维宇. Use “Soh Wei Yu” or “Soh.”
* Do not use 念头界 for thought realm.
* Do not use 错误的见 for wrong view.
* Avoid 固有存在 / 固有性 when the doctrinal meaning is 自性存在 / 自性 / 实有性.
* Avoid 实有化 where 实体化 is meant.
* Do not use 知识 for vidyā / rig pa / knowledge-of-basis/state/essence contexts.
* Do not translate Dzogchen “View” mechanically as 知见 when the technical View/lta ba is meant.
* Do not translate every “effort” as 精进; use 勤作 / 努力 / 尽力 / 精进 only by source context.

For other target languages:

* Use the uploaded prompt pack’s terminology rules and any user-supplied `TARGET_LANGUAGE_TERMBANK`.
* Preserve technical labels where AtR convention requires loan terms.
* Translate ordinary English prose naturally into `TARGET_LANGUAGE`.
* Do not leave ordinary English untranslated by accident.
* Flag unresolved terminology choices for human review rather than silently normalizing them.
* Do not over-protect English terms. A protected English label may remain, but ordinary prose using the same word should be translated unless intentionally preserved.

### Universal target-language QA module

For every target language, create a compact target-language QA profile before final review. The profile must identify:

* Script family and mixed-script risks.
* Whether English source residue is visually obvious or visually hidden by shared Latin script.
* Common grammar risks for that language, such as agreement, case, gender, tense, compound formation, honorifics, particles, postpositions, or word order.
* Punctuation and quotation conventions.
* Whether Buddhist/Dharma terms are normally translated, transliterated, or preserved as Sanskrit/Pāli/Tibetan/English loan labels.
* Whether the language needs special handling for exact article titles, bilingual labels, or proper names.
* A target-language search basket for ordinary English residue and high-risk AtR/Dharma terms.
* A target-language search basket for awkward calques and line-merge damage.

For all target languages, the following distinction is mandatory:

* **Protected labels:** exact technical labels, proper names, exact article titles, and source-required quoted phrases may remain in English or Sanskrit/Pāli/Tibetan where appropriate.
* **Ordinary prose:** ordinary English prose, even when it contains words that are also technical labels in other contexts, must be translated into `TARGET_LANGUAGE` unless explicitly preserved and accounted for.

Examples of terms that may be protected as labels in some contexts but must not be overprotected in ordinary prose: Presence, Awareness, Self, No Mind, no-mind, non-dual, emptiness, view, experience, realization, practice, luminosity, consciousness, source, ground, path, agent, doer, subject, object.

### Universal target-language termbank requirement

Before or during translation, build a small working termbank for `TARGET_LANGUAGE` covering at least the high-risk AtR/Dharma terms listed below. The termbank may preserve a loan term, translate it, or use a paired translation + retained label, but each choice must be consistent with source context, the uploaded prompt pack, and target-language readability.

The working termbank must include, where relevant:

* anatta / no-self / non-self / anatman
* emptiness / empty nature
* dependent arising / dependent origination
* Presence / presence
* Awareness / awareness
* consciousness
* luminosity / clarity / vividness
* I AM / AMness
* Self / self / self/Self
* One Mind / No Mind / no-mind
* non-dual / non-duality
* subject / object / subject-object division
* agent / agency / doer / doership / non-doership
* suchness / thusness / isness
* manifestation / appearance / phenomena
* inherent existence / inherentness / essence-view
* reification / substantialist nonduality
* mirror-like awareness / mirror metaphor
* ordinariness / naturalness / spontaneity / spontaneous perfection
* total exertion
* Dzogchen / Mahamudra / rigpa / Dharma / Buddhadharma

If no stable target-language rendering is available, preserve the technical label and flag it for human review rather than inventing a confident but unsupported rendering.

Be especially careful with these AtR/Dharma terms in every target language:

* anatta / no-self / non-self
* emptiness
* Maha
* ordinariness
* spontaneous perfection
* Presence
* Awareness
* I AM / AMness
* One Mind / No Mind
* no-mind
* non-dual
* dependent arising / dependent origination
* luminosity
* appearances
* total exertion
* suchness / thusness / isness
* subject/object
* agent / doer / doership
* mirror-like awareness / mirror metaphor
* inherent existence / essence-view / substantialist nonduality
* Dzogchen / Mahamudra

Do not force all English technical terms into `TARGET_LANGUAGE` if the AtR convention is to preserve a technical loan term, but do not leave ordinary English phrases untranslated by accident. Preserve quoted English technical labels visibly as labels, not ordinary target-language prose.

---

## 14. Audit requirement for every return

Compare against the uploaded source HTML segment covered so far and confirm:

* Top-level element / major wrapper structure covered so far.
* Tag/class structure preservation.
* Dialogue speaker-label preservation.
* Links/hrefs preserved exactly.
* Src/media preserved exactly.
* CSS/style block preserved unless intentionally updated by the styling prompt.
* Human-facing attributes translated/preserved/classified.
* No major untranslated ordinary English remains except proper names, URLs, source titles, and intentional technical labels.
* No target-language forbidden terms remain.
* No source-language + target-language hybrid residue remains, except explicitly protected labels with natural target-language syntax.
* No target-language text or non-source non-Latin characters entered href/src/action/CSS URL/script/config/code-like values.
* No added doctrinal glosses.
* No omissions.
* Stopping point and next source line/sentinel if incomplete.
* Whether any material edits were made in the pass.
* Whether target-language grammar, fluency, and calque risk were checked, not merely source/target structure.
* Whether the target-language search baskets and representative sampled sentences are documented in the QA/changelog.
* Correct status label under Section 9 and Section 16.

---

## 15. Final return

When the translation reaches the original article ending:

1. Return one complete full translated HTML file.
2. State that it reaches the original article ending.
3. Run Prompt 6-style review before final status.
4. Run Prompt 9 whole-document refinement after Prompt 6 if the pack calls for it.
5. Run Strict HTML QA / Blogger Formatting checks if the pack routes final AtR Blogger presentation through them.
6. Run the Section 16 hard final-gate module.
6a. Run the adversarial target-language Prompt 6/9 defect-hunt module in Sections 16.11–16.16, including at least the required high-risk sentence sampling and search-basket evidence.
7. Produce:

   * Updated `TARGET_LANGUAGE` HTML.
   * QA / changelog TXT.
   * ZIP bundle containing both.
   * A short chat summary with the correct status label.

8. Do not claim strict-certified final.
9. Do not promote to reviewed/publishable in the same pass where the translation or any material correction was made.
10. If material edits were made in the final translation/review pass, status must remain:

“REPAIRED AFTER RE-AUDIT — material gates currently pass, no-repair promotion still required.”

11. If I later say “continue,” reopen the exact latest artifact and perform a no-repair promotion attempt. If no material edits are made, then and only then use:

“reviewed final / publishable — strict-certified final not claimed.”

Begin by reading the uploaded prompt pack, identifying which prompt rules govern the workflow, explicitly stating how Prompt A, Prompt 1, Prompt 6, and Prompt 9 will be applied, and then proceed with the translation into `TARGET_LANGUAGE`.

---

## 16. Hard Final-Gate Module — No QA-by-Assertion, Residue Regression, and No-Repair Promotion Discipline

This section is mandatory for every checkpoint, repair pass, final pass, and promotion pass.

### 16.1 No QA-by-assertion rule

Do not claim “reviewed,” “publishable,” “final,” “complete,” “material gates pass,” “ready,” “certified,” or equivalent status merely because you say Prompt A / Prompt 1 / Prompt 6 / Prompt 9 / Strict HTML QA was applied.

A status claim is valid only if it follows a fresh audit of the exact latest artifact.

Do not substitute confidence language for verification.

If a check was not actually performed, state:

“Not fully verified: [specific check]. Human review required.”

### 16.2 Exact-latest-artifact rule

Every repair, review, audit, and no-repair promotion pass must reopen and operate on the exact latest HTML/text artifact.

Do not rely on memory, previous summaries, previous QA notes, prior confidence, file names alone, an earlier artifact, or a partially updated draft.

If there is ambiguity about which artifact is latest, do not promote. Use the most recent explicit user-provided or tool-created artifact and say which one was used.

### 16.3 Material repair invalidates promotion

If any material edit is made during a pass, the artifact cannot be promoted in that same pass.

Material edits include, but are not limited to:

* translation correction
* source-language residue cleanup
* mixed-language hybrid repair
* href/src repair
* URL slug repair
* omitted-section restoration
* duplicated-section removal
* doctrinal terminology repair
* speaker attribution repair
* punctuation repair that affects meaning or readability
* paragraph/list/blockquote/section restoration
* HTML structure repair
* media/embed repair
* attribute repair
* title/link-text repair
* CSS or wrapper repair that changes rendered output

If any material edit is made, status must be:

“REPAIRED AFTER RE-AUDIT — material gates currently pass, no-repair promotion still required.”

Only a later separate pass that reopens the exact latest repaired artifact and makes no material edits may promote to:

“reviewed final / publishable — strict-certified final not claimed.”

### 16.4 Protected-English whitelist gate

Before scanning for source-language residue, create or infer a protected-English whitelist.

English may remain only if it is one of the following:

1. URL or code.
2. HTML/CSS/JS attribute name, selector, class, ID, property, function name, variable name, or config value.
3. Proper name or author/person/group name.
4. Exact linked article title intentionally preserved as a title.
5. Quoted technical label intentionally preserved, such as:

   * I AM
   * AMness
   * anatta
   * anatman
   * Dzogchen
   * Mahamudra
   * Brahman
   * Sunyata
   * Maha
   * self/Self
   * Presence
   * Awareness
   * One Mind
   * No Mind
   * no-mind
   * non-dual
   * rigpa
   * Dharma
   * Buddhadharma
   * another explicitly protected AtR/Dharma label

6. Bilingual navigation label intentionally retained, such as “Bengali Version,” if the multilingual list requires it.
7. A quoted original phrase that the source requires preserving.

Everything else is ordinary source-language residue and must be translated or explicitly flagged for human review.

Important:
Do not overprotect ordinary lowercase uses of otherwise protected terms. For example, “Presence” as a technical label may remain if intentionally protected, but ordinary prose uses of “presence” should normally be translated in a non-English target-language article.

### 16.5 Mandatory source-language residue scan

Before any no-repair promotion, scan visible text for:

1. Ordinary English words embedded in target-language prose.
2. English + target-language hybrid particles or suffixes, such as:

   * term-এর
   * term-কে
   * term-তে
   * term-এরই
   * term-ও
   * term and term
   * term বা
   * term নয়
   * term হলো

   where the English term is not protected.

3. Lowercase ordinary English words, including but not limited to:

   * experience
   * realization
   * insight
   * practice
   * view
   * emptiness
   * awareness
   * consciousness
   * manifestation
   * phenomena
   * absolute
   * ultimate
   * mind
   * presence
   * sound
   * taste
   * vivid
   * ontological
   * dualistic
   * non-conceptual
   * meditation
   * teaching
   * guide
   * path
   * ground
   * source
   * self
   * no-self
   * non-self
   * dependent
   * origination
   * arising
   * luminosity
   * clarity
   * naturalness
   * ordinariness
   * spontaneity
   * subject
   * object
   * agent
   * doer
   * doership

4. Corrupted mixed-script words.
5. Accidental translation of exact article titles that should remain exact titles.
6. Accidental preservation of ordinary English prose that should be translated.
7. Residual English punctuation patterns caused by line merging.
8. Untranslated source tail after a translated section.

If any material residue is found and repaired, do not promote in the same pass.

For target languages other than Bengali, adapt the hybrid-pattern scan to the relevant target-language particles, suffixes, postpositions, inflections, punctuation, and script-mixing risks.

### 16.6 Href/src and attribute integrity gate

Before promotion, compare source and target HTML for:

1. Raw href count.
2. Raw src count.
3. Exact href values.
4. Exact src values.
5. Non-Latin characters inside href/src/action/data-url/style URL values unless present in the source URL.
6. Spaces inserted into URLs, such as:

   * blogspot. com
   * facebook. com
   * index. php
   * youtube. com
   * awakeningtoreality. com

7. Translated URL slugs.
8. Missing links.
9. Duplicated links.
10. Changed iframe config values.
11. Changed script/config values.
12. Any added link must be explicitly accounted for.

Do not translate or mutate:

* URL paths
* href values
* src values
* iframe config values
* CSS URLs
* script values
* IDs
* classes
* selectors
* code-like attributes
* JavaScript/config values
* Blogger widget identifiers

Human-facing attributes such as `alt`, `title`, `aria-label`, `placeholder`, iframe `title`, and caption-like data attributes may be translated only when they are user-facing language and not code, URL, exact title, proper name, or protected label.

Any non-source href/src change is a material defect unless explicitly requested by the user.

### 16.7 Coverage and structure gate

Before promotion, compare source and target for:

1. Major section/stage count.
2. Heading sequence and order.
3. Paragraph count, allowing only explicitly accounted additions such as the required English-original top link.
4. Blockquote count.
5. List-item count.
6. Dialogue/speaker-turn count where applicable.
7. Table count and row/column preservation where applicable.
8. Media/embed count.
9. Article ending reached.
10. Style block closed before article body.
11. No untranslated source tail remains hidden after the translated article.
12. No duplicated translated blocks.
13. No dropped links inside paragraphs.
14. No dropped parenthetical notes.
15. No speaker labels absorbed into previous paragraphs.
16. No ordinary text accidentally placed inside CSS/script blocks.
17. No CSS/script code accidentally translated.

Any missing, duplicated, reordered, or structurally corrupted section is a material defect.

### 16.8 Risk-basket regression rule

Every discovered defect must be added to a session-specific risk basket.

The next review pass must scan globally for the same defect type, not only repair the single instance.

Examples:

1. If one English+Bengali hybrid like “dependent origination-এর” is found, scan globally for all English + Bengali suffix/particle patterns.
2. If one target-language word appears inside `href`, scan all href/src/action/data-url/style URL values for non-source non-Latin characters and translated slugs.
3. If one exact linked title is partially translated, scan all exact linked titles.
4. If one dropped paragraph is restored, rerun section/paragraph/list/blockquote coverage checks.
5. If one protected term is overprotected, scan for ordinary lowercase uses of that same word in prose.
6. If one source-language residue phrase remains, scan for synonyms and related residue terms in the whole document.
7. If one punctuation/line-merge defect is found, scan all nearby merged lines, list items, and blockquotes.
8. If one speaker attribution issue is found, rerun all speaker-turn/order checks.

The risk basket must be cumulative within the session.

### 16.9 No-repair promotion checklist

A no-repair promotion pass must satisfy all of the following:

1. Reopen exact latest artifact.
2. Run visible-text residue scan.
3. Run protected-English whitelist classification.
4. Run href/src parity check.
5. Run non-Latin-in-URL scan.
6. Run translated-URL-slug scan.
7. Run section/heading/paragraph/list/blockquote/media coverage checks.
8. Run dialogue/speaker-turn checks where applicable.
9. Run terminology risk basket.
10. Run line-merge/punctuation audit.
11. Make no material edits.
12. Produce QA notes stating:

   * artifact reviewed
   * source artifact used
   * checks performed
   * exceptions remaining
   * whether any material edit was made

Only then may the status be:

“reviewed final / publishable — strict-certified final not claimed.”

### 16.10 Required hard-gate audit report

Every final, review, repair, or promotion response must report:

* source artifact used
* target artifact used
* covered source range
* first and last source sentinel
* next source sentinel if incomplete
* style block closed before article body
* major section/stage count
* paragraph count comparison
* blockquote count comparison
* list-item count comparison
* href count comparison
* src/media count comparison
* extra links accounted for
* non-Latin-in-URL scan result
* translated-URL-slug scan result
* source-language residue scan result
* mixed-language hybrid scan result
* target-language fluency/calque/grammar scan result
* protected-English whitelist used
* residue / terminology / calque search baskets used
* representative high-risk sentence sampling performed
* risk basket scan result
* whether material edits were made
* correct status label

If any item cannot be verified, state it as not fully verified and do not claim it passed.


### 16.11 Adversarial Prompt 6/9 defect-hunt module

Before any final or no-repair promotion status, perform an adversarial review whose purpose is to find defects, not to confirm previous QA.

This pass must explicitly check:

1. Semantic fidelity against the English source.
2. Doctrinal terminology accuracy.
3. Ordinary source-language residue.
4. Mixed source-language + target-language hybrid residue.
5. Overprotected English labels that should have been translated in ordinary prose.
6. Exact article-title handling.
7. Punctuation and line-merge damage.
8. Target-language grammar, agreement, case/gender/number, tense, register, and naturalness.
9. Overliteral calques that are grammatical but unnatural or doctrinally misleading.
10. Blockquotes, bold sections, italic notes, parentheticals, captions, link text, and end sections, because these are common places where earlier QA misses defects.

A no-repair promotion cannot rely mainly on counts such as paragraph count, href count, blockquote count, or section count. These are necessary but insufficient.

### 16.12 Evidence-based promotion rule

Every no-repair promotion attempt must include an evidence log in the QA/changelog. The evidence log must list:

* exact latest target artifact reopened
* source artifact used
* target-language risk profile used
* protected-English whitelist used
* ordinary English/source-language residue search basket used
* high-risk AtR/Dharma terminology basket used
* target-language calque/grammar/punctuation basket used
* href/src/code integrity checks used
* representative high-risk source/target sentence samples reviewed
* issue families from the cumulative risk basket and the result of each scan
* whether any material edits were made

Do not merely state “scan passed.” State what was scanned for.

### 16.13 Required high-risk sentence sampling

For each final, repair, or promotion pass, review a minimum set of high-risk sentences manually against the source. If the document is short, sample proportionally but do not skip this requirement.

For a full article, sample at least:

* 5 sentences from the beginning/opening sections.
* 5 sentences from the middle sections.
* 5 sentences from the ending sections.
* 5 sentences from blockquotes or quoted teaching passages.
* 5 sentences from bold/italic/commentary sections.
* all sentences containing the most important AtR/Dharma terms for the article, or a representative set if there are too many.

For each sampled sentence, check meaning, terminology, grammar, punctuation, and target-language naturalness. If any material issue is found, scan globally for that issue family.

### 16.14 Second-reader mode

In every no-repair promotion pass, read the target artifact as if it were produced by another translator whose work may contain subtle errors.

Do not trust:

* previous status labels
* previous QA summaries
* previous confidence language
* file names such as “final,” “reviewed,” or “no-repair”
* mechanical parity alone
* your own memory of earlier repairs

Reopen the exact latest artifact and challenge it from scratch.

### 16.15 False-promotion retraction rule

If a later pass finds material defects after an earlier no-repair promotion, explicitly acknowledge that the previous promotion was premature. Do not minimize the issue by saying the later pass is merely “extra refinement” if the defect affects translation fidelity, readability, terminology, HTML integrity, source-language residue, mixed-language residue, line-merge punctuation, or target-language grammar.

When this happens:

1. Reset status to “REPAIRED AFTER RE-AUDIT — material gates currently pass, no-repair promotion still required.”
2. Add the defect family to the risk basket.
3. Scan globally for related defects.
4. Record the false-promotion lesson in the QA/changelog.
5. Require a later separate no-repair promotion pass before any publishable status may be restored.

### 16.16 Target-language fluency is a material gate

A translation can preserve HTML structure and still fail Prompt 9 if the target-language prose is unnatural, ungrammatical, overliteral, or doctrinally misleading.

Treat the following as material defects when they affect readability, source meaning, or doctrinal clarity:

* awkward literal calques from English
* broken compounds
* unnatural word order
* agreement/case/gender/number errors
* missing punctuation after merged lines
* accidental double negatives
* malformed rhetorical questions
* mistranslated idioms
* source-language sentence structure that sounds foreign in `TARGET_LANGUAGE`
* inconsistent rendering of key terms
* failure to distinguish technical labels from ordinary prose
* preserving English where the target language needs translation

If such defects are repaired, do not promote in the same pass.

---

## 17. Artifact requirements

For every completed checkpoint, repair pass, final pass, or no-repair promotion pass, produce:

1. Updated full `TARGET_LANGUAGE` HTML file.
2. QA / changelog TXT file, including the Prompt Application Ledger, hard-gate audit report, target-language risk profile, search baskets used, representative high-risk sentence sampling notes, material edits made, and status label.
3. ZIP bundle containing the HTML and QA TXT.
4. Short chat summary with:
   * Prompt Application Ledger
   * status label
   * links to the files
   * human-review notes

Do not produce fragment-only files unless I explicitly request a fragment.

If the article is incomplete, the HTML file must still be a valid cumulative Blogger-ready HTML file containing all translated content so far and excluding later untranslated source-body content.

---

## 18. Human-review placeholders and known exceptions

If `SOURCE_ARTICLE_URL` is not provided, use:

`URL_NEEDED_HUMAN_REVIEW`

Do not invent the source URL.

Flag this in every QA report until resolved.

Any unresolved terminology, source ambiguity, missing source URL, inaccessible prompt-pack detail, or unperformed check must be listed under:

“Known Human-Review Items”

Do not hide known exceptions behind a “final” status.

---

## 19. Final instruction

Begin by reading the uploaded prompt pack, identifying which prompt rules govern the workflow, and explicitly stating how Prompt A, Prompt 1, Prompt 6, Prompt 9, Prompt 8 if relevant, Blogger Formatting / Strict HTML QA if routed, and the Section 16 Hard Final-Gate Module will be applied.

Then proceed with the translation into `TARGET_LANGUAGE`.

Remember:

* The uploaded English HTML is the source.
* The prompt pack controls workflow.
* The initial translation is not publishable-final in the same pass.
* A material repair pass is not a no-repair promotion pass.
* No publishable status is allowed unless the exact latest artifact is reopened, hard-gate audited, adversarially reviewed under Prompt 6/9, target-language fluency-checked, evidence-logged, and no material edits are made.

---

## 20. Practical anti-false-final operating checklist

Use this checklist before every final or promotion response:

1. Did I reopen the exact latest target artifact rather than relying on memory?
2. Did I compare it against the uploaded English source, not a live page or prior translation?
3. Did I run Prompt 6 for fidelity, terminology, omissions, additions, doctrinal nuance, source-language residue, mixed-language residue, exact-title handling, and HTML preservation?
4. Did I run Prompt 9 for whole-document target-language fluency, grammar, punctuation, line-merge damage, readability, and consistency?
5. Did I run structure/count checks without mistaking them for translation-quality proof?
6. Did I create or update the protected-English whitelist?
7. Did I search for ordinary English/source-language residue outside the whitelist?
8. Did I search for target-language-specific hybrid residue and mixed-script damage?
9. Did I search href/src/action/CSS URL/script/config/code-like values for mutation, translated slugs, spaces, or non-source characters?
10. Did I review high-risk Dharma terminology and distinguish protected labels from ordinary prose?
11. Did I perform high-risk sentence sampling from beginning, middle, ending, blockquotes, and commentary/bold/italic sections?
12. Did I scan the cumulative risk basket from all previous defects found in this session?
13. Did I avoid claiming “publishable” in the same pass as a material edit?
14. If no material edit was made, did I document the evidence that justifies no-repair promotion?
15. If a later review found defects after prior promotion, did I clearly retract the earlier premature promotion and add the issue family to the risk basket?

If any answer is “no,” do not promote. Complete the missing check or state it as not fully verified and require human review.