Inspiration

Subtitles moderate. Ávila-Cabrera measured 41.8% of Pulp Fiction's taboo load never reaching the Spanish subtitle; we measured 39.6% over 688,351 English lines with "fuck". And Mexican Spanish barely exists in the data: 241 recognisably Mexican documents out of 135,470, peninsular markers 7:1. A Mexican viewer almost never gets subtitles faithful to what the character said, in their own Spanish. We built the other track.

What it does

Type an English subtitle line. Subtext returns every Mexican reading the corpus attests (tú, usted, ustedes; masculine, feminine), each written by Gemini from retrieved evidence and each citing the pair_id of real lines Mexican translators wrote. Under every answer, two controls: the same Gemini with no corpus, and Google Translate. When evidence is thin it says WEAK; when there is none, it abstains.

How we built it

OpenSubtitles (103.6M en–es pairs) in ClickHouse Cloud on GCP. A lexicon detector isolates 241 Mexican documents (335,800 lines). Two retrieval channels, always both: an HNSW vector index over gemini-embedding-001 vectors (6 ms per search) and a phrase index of 171,888 n-grams whose renderings are ranked by enrichment against the full corpus, plus a SQL channel that measures what all lines carrying a phrase agree on. Gemini 3.8 Flash via Google ADK writes one line per reading; Python gates check register and form of address and retry once. SQL goes through mcp-clickhouse, the official ClickHouse MCP server. Answers are cached in ClickHouse. The site runs on Cloud Run.

Findings and learnings

  • Frequency reproduces the censorship. ¡Jódete! is the most frequent rendering of "Fuck you!" and 0.9× enriched; ¡Chinga tu madre! is rarer and 188× enriched. Enrichment recovers what is Mexican.
  • It is not only profanity. "Jacket" is chamarra 76 times to chaqueta 16 in Mexican subtitles; Google Translate says chaqueta. The literature measures censorship, not variety. The corpus does.
  • Thinking off is better, not just cheaper. With reasoning the model wrote "Ahorita mismo" (never attested); without it, "Ahorita" (1,382 times).
  • Attesting a word is not attesting a sense. Mamada appears 206 times, never in the anatomical sense. The system flags it instead of pretending.
  • We withdrew our own first evaluation when we found its golden set came from the same process that wrote the corpus. The current test set carries the origin of every line; the strongest group was chosen by published scholars. Full runs are public, failures included: 40/40 scholar lines translated, 0 peninsular leaks, 7 weak.

Built with

Google Cloud · Cloud Run · Gemini 3.8 Flash · Google ADK · gemini-embedding-001 · ClickHouse Cloud · mcp-clickhouse · Python 3.12 · FastAPI

Built With

Share this project:

Updates