Skip to main content
MenuEtruscan

REST API Documentation

A read-only REST API over the corpus, for use in digital humanities pipelines. It is free and unauthenticated, and rate-limited to 100 requests per minute.

GET

/api/search

Query the corpus using full-text search or semantic filters.

Example Request
fetch("https://openetruscan.com/api/search?q=larth&limit=10")
  .then(res => res.json())
  .then(data => console.log(data));
GET

/api/inscription/{id}

Retrieve the full record for a specific inscription (canonical text, findspot, classification, dates, and linked-open-data references).

Example Request
fetch("https://openetruscan.com/api/inscription/TLE_131")
  .then(res => res.json())
  .then(data => console.log(data.canonical));
GET

/api/concordance

Generate a Keyword-in-Context (KWIC) matrix for lexicographical analysis.

Example Request
fetch("https://openetruscan.com/api/concordance?q=clan&context=40")
  .then(res => res.json())
  .then(data => console.log(data.rows));

How results are evaluated

Every computational number on this site is reported under a pre-registered protocol: the metric, baselines, and significance test are fixed before the evaluation runs, and each figure carries a 95% bootstrap confidence interval. Model-comparison claims require a paired-bootstrap p-value.

“Gold” is not yet gold. The classification and lacuna reference sets are LLM-consensus (“silver”): multiple frontier models agreeing under one codebook. No philologist has ratified them yet, so they are labelled candidate-gold, not authoritative.

Classifier: best macro F1 = 0.313 ± 0.038 (95% CI, n=143 held-out). The differences between architectures are within noise; the bottleneck is labelled-data count, not model choice. An earlier “99% macro F1” headline was an in-training-set fit and has been retracted.

Lacuna restoration: a retrieval-augmented restorer fills known-width gaps with a calibrated confidence (span-exact ≈ 0.26 → 0.38 with retrieval, paired-bootstrap p = 0.025). It is best on short, formulaic gaps and abstains elsewhere. It does not decipher or translate.

The full protocol, deviations, and negative results (including a retracted lacuna “Finding C”) live in the research repository: pre-registration, findings.