Methodology

This page documents how each translation on the site is produced, evaluated, and labelled. It is the public companion to the Translation Status Contract in the repository.

How we translate

The pipeline is Sanskrit-grounded, not anchor-grounded. For each verse, the translation model receives the Devanāgarī text, an IAST transliteration, the Vidyut morphological segmentation (lemmas, case, number, gender, sandhi splits), Cologne C-SALT glosses per lemma from Monier-Williams and Apte, the Skrutable meter tag, and the preceding two verses for citation-aware context. A public-domain English translation may be passed as a reference signal — explicitly marked as such — but is never the source of truth. When the English anchor and the Sanskrit morphology disagree, the model is instructed to trust the Sanskrit.

The LLM-as-judge

Every draft is scored 1–10 for Sanskrit fidelity by a second model running the same input bundle (Devanāgarī + morphology + glosses + meter + context). The judge returns a score, a rationale, and a list of concerns; these are stored alongside the translation row.

• Score ≥ 7 — promote from draft to published; visible on the site behind the AI · not verified badge.
• Score ≥ 8 — flagged high-confidence (a subtle badge variant lands in V1.x).
• Score < 7 — stays draft; excluded from public display until revised or human-reviewed.

Calibration is checked quarterly by human spot-audit of 50 random published translations per language. Audit outcomes feed back into prompt versions, which are themselves recorded per translation row (prompt_version) so any score is reproducible.

AI-assist badge legend

Each translation carries a badge inline with its verse number. The badge state is derived from two fields: ai_assisted (boolean) and status (draft | published | reviewed). The full state table — including badge colours, public-display rules, and the V1.1 five-state expansion — lives in the Translation Status Contract . License terms for the translated text itself are documented on the License page.

Why no commentary is glued under verses

The classical auto-commentaries — Kṣemarāja's Śivasūtravimarśinī on the Śiva Sūtras, Abhinavagupta's Tantrāloka-viveka, Jayaratha's commentary on the Tantrāloka — are full texts in their own right. They are not annotations to be folded under the base sūtras. We ingest each one as a sibling text under the same verse-anatomy schema. The Vimarśinī, for example, lives at its own route (/trika/siva-sutras-vimarsini), not nested under the Śiva Sūtras. Readers who want both can open them side by side; the structural hierarchy of the site does not imply that the commentary is subordinate.

Public-domain English anchors

Pre-1930 translations by John Woodroffe (Arthur Avalon), Ralph Griffith, George Thibaut, Max Müller, W. D. Whitney, and J. H. Woods are present in the pipeline as reference signals only. They are never the ground truth, never displayed unattributed, and never used to override Sanskrit morphology when the two conflict. Full per-translator attribution lives on the Sources page.

Editorial policy

Per-text source-of-truth is recorded in the texts.source_revision column — the exact GRETIL or Muktabodha or sanskritdocuments.org revision we ingested, with date. Critical-edition emendation (proposing variant readings, reconciling manuscript traditions) is out of scope for V1; we publish the source revision as-is and link the upstream so scholars can verify. V1.x will add a variants column for recorded alternate readings without committing to an emended text.

Reviewer policy

Human reviewers — when V1.x opens the review pipeline — work from the Sanskrit and morphology, never from copyrighted modern translations. Pasting from Jaideva Singh, Mark Dyczkowski, Lilian Silburn, Alexis Sanderson, or any other in-copyright translator is not permitted. Reviewers may consult those works for orientation; the accepted English (or other-language) wording must be the reviewer's own and grounded in the source.

Known limitations

• V1 ships AI-only translations behind explicit badges. No translation on the public site is human-verified yet.
• The judge has not been audited at scale; calibration is best-effort until the first quarterly spot-check.
• Coverage is uneven — some texts have richer Cologne lemma coverage than others; the model receives whatever the corpus offers.
• Eleven Indic target languages are planned; only English is published at launch (per the language picker, others are explicitly marked soon).

Last revised: 2026-05-31 · Source: STATUS-CONTRACT.md and the V1 plan document.