Skip to contents

One row per evaluation that ever touched the sealed holdout split: when, by which protocol (hash), and with what headline result. Entries are appended automatically by validate_protocol() and gold_correct() and printed in full by LLMR::report(). The mechanism is visibility, not enforcement – an in-memory object cannot stop a determined user, and does not pretend to; archive the study's LLM call log (see the archive workflow) when tamper evidence is needed.

Usage

gold_ledger(x)

Arguments

x

A gold_set().

Value

A tibble: ts, split, protocol_hash, protocol_label, n, accuracy.

Examples

set.seed(110)
g <- gold_set(
  data.frame(text = paste("unit", 1:40),
             label = rep(c("x", "y"), each = 20)),
  "text", "label", split = c(dev = 0.5, test = 0.5))
gold_ledger(g)