Translation witness

Translation Witness (Stage 5 Reference Material Collection)

Collect and normalize PD English translations as reference material for one target text.

This skill is the Stage 5 owner for witness collection.

Critical Policy

Witnesses are NEVER alignment targets. Witnesses are NEVER versified.

Existing PD translations serve as:

• Adversarial comparison material during Stage 6a translation generation
• Contextual sense disambiguation for Stage 8 interlinear work
• Optional display alternatives in the reader (shown as-is, without structural modification)

Witnesses are NOT:

• Targets for DP alignment
• Targets for range alignment
• Targets for direct mapping
• Structurally modified to fit Greek segmentation
• Versified or given verse-aligned editions in the DB

Row view is unavailable for witnesses. Only versified (generated) translations can render in row view, because row view requires 1:1 structural correspondence with Greek. Witnesses have arbitrary structure.

The generated translation (Stage 6a) is always the primary English content. Witnesses are supplementary reference material.

Reference-preserving import (Path A)

Witnesses that carry native reference markers matching the Greek reference system may be imported as reader-display editions. This is not structural modification — it is parsing what the witness already provides.

Supported import formats (set in catalog format field):

• ref-text-tsv — BookName Chapter:Verse\tText (e.g., John witnesses)
• ref-bracket-prose — [ref] text with optional headers/comments (e.g., Perrin, Farquharson)
• structured-json — {"segments": [{"ref":"...", "text":"..."}]} (e.g., Hymn witnesses)
• chapters-json — legacy {"chapters": {"1": [{"verse":1, "text":"..."}]}} format

Import is blocked (with a warning) if <50% of witness references match Greek references (direct or parent-prefix match).

What is NOT permitted

• Forcing alignment on witnesses that lack native reference markers
• Splitting or merging witness segments to match Greek granularity
• Creating new reference systems that don't exist in the source text

Quick Status

Witness dirs: !find ${LYCEUM_TEXTS_DIR:-output/texts} -path '*/witnesses' -type d 2>/dev/null | wc -l English candidate catalogs: !find ${LYCEUM_TEXTS_DIR:-output/texts} -path '*/sources/english_candidates.json' 2>/dev/null | wc -l

Commands

• /translation-witness gather [work] — Normalize the available English witness sources into comparable artifacts
• /translation-witness classify [work] — Classify witness usefulness by role (adversarial comparison, glossing, display)
• /translation-witness select [work] — Choose primary witnesses for each downstream use case
• /translation-witness status [work] — Summarize witness coverage and role assignments

Target: $ARGUMENTS

---

Owned Responsibilities

Owns

• PD English witness collection and normalization
• Literalness and quality notes per witness
• Role classification (adversarial, glossing, display)
• Witness catalog with provenance

Does not own

• Translation generation (Stage 6a — translation-synthesis)
• Structural alignment of any kind (Stage 6b — versify)
• Interlinear/gloss generation (Stage 8)
• Source acquisition beyond English witnesses (Stages 1-4)

---

Witness Role Classes

Each witness should be classified for its usefulness in one or more roles. These are values for the roles array field in catalog.json (e.g. "roles": ["adversarial-comparison", "reader-display"]):

adversarial-comparison

• Literal enough to catch errors in the generated translation
• Follows Greek meaning closely (even if style is awkward)
• Useful for detecting mistranslations, hallucinations, omissions
• Primary role — at least one witness must qualify

glossing-reference

• Useful for contextual sense disambiguation in Stage 8
• May have good vocabulary choices that inform gloss selection
• Doesn't need to be structurally aligned

reader-display

• Suitable as an optional literary/idiomatic alternative in the reader
• Displayed as-is without structural modification
• May segment differently than Greek (that's acceptable for display)

versification-candidate

• Witness has native reference markers that match the Greek reference system (same scheme, same granularity)
• Witness segment count is within 10% of Greek segment count (after sub-segmentation, if applicable)
• Witness is literal enough for learner use — not a loose paraphrase or literary adaptation
• Witness is public domain
• Witness text is complete (no missing chapters or sections)

A versification-candidate witness may be directly versified by Stage 6 without generating a new translation. This is the fast path: if automated versification succeeds (100% coverage, 1:1 mapping with Greek segments), the witness becomes the versified edition and Stage 6a generation is skipped entirely.

Criteria that disqualify a witness from this role:

• Reference system doesn't match Greek (e.g., different chapter/verse numbering)
• Segment count diverges by >10% from Greek
• Translation is too literary/paraphrastic (e.g., Pope's Iliad)
• Translation merges or splits segments relative to Greek structure
• Incomplete coverage

A witness can hold versification-candidate alongside other roles (e.g., adversarial-comparison).

limited-use

• Available but not strongly useful for any of the above
• Document why (too paraphrastic, incomplete, etc.)

Important note

This classification replaces the old "alignment fitness" classification (direct-match, DP-alignable, range-only, unsuitable). Those categories are obsolete — we no longer align witnesses.

---

Workflows

/translation-witness gather

Collect and normalize available PD English translations.

For each witness, record

• Source identity and URL
• Translator and edition metadata
• Publication year and license/PD status
• Literalness notes (how closely it follows Greek)
• Structural notes (how it segments — informational only)
• Quality notes (readability, completeness)

Output location

$LYCEUM_TEXTS_DIR/<slug>/
├── witnesses/
│   ├── perrin-1919.txt
│   ├── long-1862.txt
│   └── ...
├── sources/
│   └── english_candidates.json

Witness file format

Preserve the witness text as-is. Do not force it into Greek segmentation. Include metadata header if useful, but the text should be readable.

---

/translation-witness classify

Classify each witness by role usefulness.

Evaluation criteria

For adversarial-comparison:

• How literal is the translation?
• Does it preserve Greek word order when possible?
• Does it translate idioms literally or paraphrase?
• Can it catch generated translation errors?

For glossing-reference:

• Does it have good English equivalents for Greek terms?
• Is vocabulary consistent?
• Could it inform word-level gloss choices?

For reader-display:

• Is it readable and natural English?
• Is it complete (no major omissions)?
• Would users find it valuable as an alternative view?

Output

Add classification to witness catalog (schema v2):

{
  "schema_version": 2,
  "work": "work-slug",
  "witnesses": [
    {
      "id": "perrin-1919",
      "title": "Perrin (1919)",
      "translator": "Bernadotte Perrin",
      "year": 1919,
      "license": "Public Domain",
      "file": "perrin-1919.txt",
      "format": "ref-bracket-prose",
      "roles": ["adversarial-comparison", "reader-display"],
      "literalness": "moderately literal",
      "notes": "..."
    }
  ]
}

Catalog schema v2 rules:

• file: single string, relative to witnesses/ dir (no files.json/local_file variants)
• format: one of ref-text-tsv, ref-bracket-prose, structured-json, chapters-json
• roles: flat []string (not nested objects)
• title: required — used as edition label in the reader

---

/translation-witness select

Assign witnesses to roles for downstream use.

Required selections

• Adversarial comparison (Stage 6a): At least one witness must be selected. This is mandatory — Stage 6a cannot proceed without a comparison witness.
• Glossing reference (Stage 8): Optional but recommended. Multiple witnesses improve confidence.
• Reader display: Optional. Include if witness adds value for users.

If no suitable adversarial witness exists

• Document the gap explicitly
• Stage 6a can still proceed, but generated translation confidence is lower
• Consider gathering additional witnesses or proceeding with caution

---

/translation-witness status

Summarize current state.

Report includes

• Witnesses collected
• Role classifications
• Whether adversarial-comparison witness exists
• Whether Stage 6a and Stage 8 have sufficient reference material
• Gaps or concerns

---

Reader Display Policy

Witnesses selected for reader-display are shown in the Lyceum reader as-is, without structural modification.

What this means

• The witness text keeps its original segmentation
• It may not align 1:1 with Greek segments
• The reader may display it in a side panel or expandable section
• It is clearly labeled as the translator's work (e.g., "Perrin 1919")

What we do NOT do

• Force witnesses into verse-by-verse alignment with Greek
• Run DP/range alignment to create pseudo-aligned witnesses
• Split or merge witness segments to match Greek
• Create "verse-aligned" versions of literary translations

The generated translation (Stage 6a) provides the verse-aligned English. Witnesses provide context and alternatives.

---

Priority Texts

Texts where multiple witnesses improve confidence:

Text	Recommended witnesses
Homer (Iliad/Odyssey)	Murray, Butler, additional if available
Meditations	Long, Haines
Anabasis	Brownson, Dakyns
Apology	Jowett, Fowler
John	OpenGNT, ASV, KJV as secondary

---

Verification Contract

This skill follows the Stage 5 contract.

Verify

• Witness catalog exists with provenance
• Each witness has roles array with at least one classification
• At least one adversarial-comparison witness exists (or gap documented)
• Witnesses are preserved as-is (not structurally modified)

Minimum evidence

• Witness files under witnesses/
• Witness catalog with roles assignments
• Classification report

Pass criteria

• Provenance recorded for each witness
• Role assignments explicit and justified
• Adversarial-comparison witness available for Stage 6a
• No attempts to align or modify witness structure

Failure examples

• Witnesses collected but roles not assigned
• No adversarial-comparison witness and gap not documented
• Witness text modified to fit Greek structure
• Alignment fitness classified (obsolete pattern)

Required next steps

After witness collection:

• /translation-synthesis run <work> to generate the verse-aligned translation
• Witnesses used for adversarial comparison during generation

---

Verification

After completing this stage, run the automated verification script:

bash scripts/verify_stage_5.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

File	Purpose
$LYCEUM_TEXTS_DIR/<slug>/witnesses/	Witness text files
$LYCEUM_TEXTS_DIR/<slug>/sources/english_candidates.json	Witness catalog
.pi/skills/translation-synthesis/SKILL.md	Stage 6a (uses witnesses for comparison)
.pi/skills/versify/SKILL.md	Stage 6b (validates generated translation)
docs/text-pipeline-master-plan-2026-03-13.md	Pipeline architecture
docs/reader-reliability-and-translation-accuracy-charter-2026-03-13.md	Multi-witness motivation