Translation Synthesis

Translation Synthesis (Stage 6a)

Generate a structure-preserving English translation aligned 1:1 with the Greek reference system.

This skill is the Stage 6a owner — the first half of Stage 6.

Stage 6a is conditional. Before this skill runs, the orchestrator attempts the witness versification fast path (/versify try-witness). If a versification-candidate witness is successfully versified (100% coverage, 1:1 mapping), that witness becomes the versified edition and Stage 6a is skipped entirely.

Stage 6a runs when:

• No witness is classified as versification-candidate
• All versification-candidate witnesses fail automated versification (segment count mismatch, reference mismatch, coverage < 100%)
• The orchestrator explicitly requests fresh generation

When Stage 6a does run, the generated translation is the primary English content for that text.

Quick Status

Generated translation chapters: !find ${LYCEUM_TEXTS_DIR:-output/texts} -path '*/versification/chapter_*.json' 2>/dev/null | wc -l Texts with generated translations: !find ${LYCEUM_TEXTS_DIR:-output/texts} -path '*/versification/chapter_01.json' 2>/dev/null | xargs -I{} dirname {} | xargs -I{} dirname {} | sort -u | wc -l

Commands

• /translation-synthesis run [work] — Generate or regenerate the verse-aligned translation for one text
• /translation-synthesis review [work] — Re-run adversarial review on existing generated translations
• /translation-synthesis status [work] — Summarize generation progress, review verdicts, and known issues

Target: $ARGUMENTS

Owned Responsibilities

Owns

• Stage 6a translation generation
• structure-preserving English output (one segment per Greek segment)
• adversarial review against PD witness translations
• generation metadata (model, timestamp, witness used)
• review reports with issues and verdicts

Does not own

• PD witness collection (Stage 5 — translation-witness)
• structural validation and DB import (Stage 6b — versify)
• interlinear/gloss generation (Stage 8)
• final ship promotion (Stage 10)

Witness Versification (Fast Path)

Before Stage 6a generation runs, the orchestrator invokes /versify try-witness [work] to attempt the fast path. This is documented here for context — the actual logic lives in the versify skill.

When the fast path succeeds

• A versification-candidate witness was found in the witness catalog
• Its references map 1:1 to Greek segments
• Coverage is 100%
• The witness is imported as the versified edition
• Stage 6a is skipped — no translation is generated
• The witness still gets a validation pass (checked for empty segments, reference integrity, completeness) via /versify try-witness

When the fast path fails

• No versification-candidate exists, OR
• All candidates fail automated versification
• Stage 6a runs normally — generate a fresh translation with adversarial review
• Then Stage 6b validates and imports the generated translation

What this means for this skill

• This skill's run command is only invoked when the fast path has already failed
• All generation requirements, adversarial review, and quality standards still apply
• The fast path does not change how generation works — it only determines whether generation happens at all

Generation Requirements

The generated translation MUST be:

1. Structure-preserving: exactly one English segment per Greek segment (verse, sentence, or unit)
2. Faithful: accurate to the Greek meaning, segment-by-segment — no paraphrasing across boundaries
3. Learner-readable: clear, natural English that helps learners understand what the text says
4. Consistent: recurring Greek terms get consistent English equivalents where reasonable
5. Labeled: output includes model name, timestamp, and witness translator used for comparison

The generated translation MUST NOT:

• Rearrange clauses across segment boundaries
• Merge or split segments
• Add explanatory content not in the Greek
• Be presented as a historical/literary translation

Adversarial Review

Every generated translation batch is reviewed against PD witness translations.

Review checks

1. MISTRANSLATION: meaning that differs from the Greek
2. HALLUCINATION: content added that is not in the Greek
3. OMISSION: Greek content that was dropped
4. CLAUSE REARRANGEMENT: significant reordering away from Greek structure
5. TERM INCONSISTENCY: same Greek word translated differently without reason

Review output

Each review produces a JSON object:

{
  "verdict": "pass" | "fail",
  "issues": [
    {
      "reference": "2.3",
      "type": "mistranslation",
      "description": "...",
      "suggested_fix": "..."
    }
  ],
  "notes": "..."
}

Handling review failures

• Issues are logged but translations are kept (human review will catch serious problems)
• High issue counts should trigger re-examination
• Critical mistranslations should be fixed before proceeding to Stage 6b

Workflows

/translation-synthesis run

Generate or regenerate the verse-aligned translation.

Prerequisites

• Stage 5 complete: at least one PD witness available for adversarial comparison
• Greek edition segments exist in the database
• Workspace has witness information (if using canonical workspace)

Execution

Using the generation script directly

# Preview chapter 1 without saving
python3 scripts/generate_translation.py --text <text_key> --chapter 1 --preview

# Generate specific chapter
python3 scripts/generate_translation.py --text <text_key> --chapter 1

# Generate all chapters
python3 scripts/generate_translation.py --text <text_key> --all

# Resume from a specific chapter
python3 scripts/generate_translation.py --text <text_key> --all --start-chapter 15

Execution modes

• --mode api: Uses ANTHROPIC_API_KEY to call the Anthropic API directly (faster)
• --mode agent: Spawns pi subprocesses (no API key required, slower)
• Auto-detection: uses API mode if key is available, otherwise agent mode

Adding new texts

Currently generate_translation.py requires text configuration in its TEXT_CONFIG dict. To add a new text:

1. Add entry with greek_edition_id, english_edition_id (witness), work URN, generated URN
2. Run generation
3. Optionally use --import to import directly to DB

Output location

$LYCEUM_TEXTS_DIR/<slug>/versification/
├── chapter_01.json
├── chapter_02.json
└── ...

Chapter file format

{
  "chapter": 1,
  "generated_at": "2026-03-15T04:44:15.458022+00:00",
  "model": "claude-sonnet-4-20250514",
  "witness": "Bernadotte Perrin (1919)",
  "segments": [
    {"reference": "1.1", "translation": "..."},
    {"reference": "1.2", "translation": "..."}
  ],
  "reviews": [
    {"verdict": "pass", "issues": [], "notes": "..."}
  ]
}

/translation-synthesis review

Re-run adversarial review on existing generated translations.

Useful when:

• New/better PD witnesses have been gathered
• Review prompts have been improved
• Issues from initial review need re-examination

Steps

1. Load existing chapter files from versification/
2. Load current PD witness(es)
3. Re-run review prompts
4. Update review sections in chapter files
5. Report aggregate pass/fail/issue counts

/translation-synthesis status

Summarize current state.

Report includes

• Chapters generated vs total expected
• Review verdicts (pass/fail counts)
• Issue breakdown by type
• Outstanding high-severity issues
• Readiness for Stage 6b (versify)

SQL helpers

-- Count Greek segments by chapter
SELECT 
  CAST(SUBSTR(reference, 1, INSTR(reference, '.') - 1) AS INTEGER) as chapter,
  COUNT(*) as segments
FROM segments 
WHERE edition_id = <greek_edition_id>
GROUP BY chapter;

-- Count generated segments
SELECT COUNT(*) FROM segments WHERE edition_id = <generated_edition_id>;

Important Rules

Generation is skipped only via the fast path

Do NOT skip translation synthesis for ad-hoc reasons like:

• "The existing Perrin/Murray/Long translation is good enough"
• "This text already has verse numbers"
• "It's biblical so we can use KJV/ASV"

The only legitimate way to skip Stage 6a is via the witness versification fast path (/versify try-witness), which requires:

• A witness explicitly classified as versification-candidate in the Stage 5 catalog
• Automated versification succeeding with 100% coverage and 1:1 mapping
• The orchestrator confirming the fast path result before skipping generation

If the fast path was not attempted or did not succeed, generation MUST run.

Per-chapter isolation

• Process one chapter at a time
• Never mix segments from different chapters in a batch
• Save after each chapter to enable resume

Labeling

All output must be clearly labeled as generated:

• generated_at timestamp
• model name
• witness used for comparison
• URN includes "generated" marker (e.g., ...generated-eng1)

Rate limiting

When using API mode, include delays between batches to avoid rate limits.

Lessons Learned

Why generation is the default (and the fast path is an exception)

1. Pope's Iliad cannot be forced into line-level 1:1 alignment — too literary, wrong structure
2. Townsend's Aesop is too loose for reliable verse alignment
3. Murray's Homer requires range-based fallback with imperfect coverage
4. Long's Meditations required elaborate DP mapping with hand-verified tables
5. Even "verse-aligned" biblical texts have numbering variations

Generating a structure-preserving translation eliminates all alignment-mode classification effort for the majority of texts.

When the fast path works (and why it exists)

Some texts have PD translations that are already naturally 1:1 with the Greek:

• NT witnesses (ASV, KJV) share the same verse numbering as the Greek — despite occasional variations, automated versification can handle these when the match rate is high enough
• Scholarly literal translations of segmented texts (e.g., Perrin's Plutarch with section markers) may already align structurally
• The fast path exists to avoid expensive LLM generation when a perfectly good literal PD translation already has the right structure

The key insight: the old problems (Pope, Townsend, Murray, Long) were caused by trying to force-align unsuitable witnesses. The fast path only attempts versification on witnesses that Stage 5 has explicitly classified as versification-candidate — witnesses that are literal, reference-aligned, and structurally compatible.

What makes a good witness for comparison

• Literal scholarly translations catch more errors
• Multiple witnesses increase confidence
• Literary/idiomatic translations may miss structural issues
• Witness doesn't need to be aligned — just readable for comparison

Verification Contract

This skill follows the Stage 6a contract.

Verify

• Generated translation has one segment per Greek segment
• Adversarial review was performed against at least one PD witness
• No cross-chapter contamination
• Model/timestamp metadata is present
• Issues are documented (even if verdict is "pass")

Minimum evidence

• Chapter files under versification/
• Each chapter file has segments and reviews
• generated_at, model, and witness fields present

Pass criteria

• 100% structural coverage: every Greek segment has a generated translation
• Adversarial review completed for all chapters
• No unresolved critical mistranslations blocking Stage 6b
• Output is labeled as generated

Failure examples

• Generated translation hallucinates content not in the Greek
• Adversarial review was skipped
• Segments are merged or split relative to Greek structure
• Chapter files missing review data

Next steps after pass

• Proceed to /versify run <work> for structural validation and DB import

Verification

After completing this stage, run the automated verification script:

bash scripts/verify_stage_6a.sh "${SLUG}"

Exit codes: 0=PASS (advance), 1=FAIL (block), 2=WARN (advance with notes). The orchestrator runs this automatically; when executing manually, check the output for [FAIL] or [WARN] lines.

Key Files