Preference-pair schema adapter¶

openmed.traces.schemas.preference provides a local-first adapter for preference-training records with three required content fields: prompt, chosen, and rejected.

{
  "pair_id": "synthetic-pair-01",
  "prompt": "Which follow-up is appropriate for Ada Example?",
  "chosen": "Arrange the local follow-up.",
  "rejected": "Publish the source note.",
  "scores": {"chosen": 0.91, "rejected": 0.09},
  "metadata": {"synthetic": true, "source": "offline-fixture"}
}

The adapter copies the record and walks the three branches without changing pair membership, score fields, IDs, roles, or non-content metadata. Strings in message-style values such as [{"role": "assistant", "content": "..."}] are redacted through their content fields. A single PreferenceRedactionState is shared across all three branches, so an identical (label, sensitive value) pair gets one deterministic surrogate.

from openmed.traces.schemas.preference import PreferencePairAdapter

adapter = PreferencePairAdapter(seed=17)
redacted = adapter.redact(record)
result = adapter.redact_with_report(record)
assert result.to_mapping() == redacted

The default detector is deterministic and local. It covers common emails, phone numbers, dates, network addresses, common identifiers, and conservative Latin-script names; it does not load a model or make a network call. For a richer detector that is already available locally, inject a span detector or a text redactor:

from openmed.traces.schemas.preference import PreferencePairAdapter, SensitiveSpan

def detector(text):
    start = text.find("SYNTH-42")
    if start < 0:
        return ()
    return (SensitiveSpan(start, start + len("SYNTH-42"), "ID_NUM"),)

adapter = PreferencePairAdapter(span_detector=detector, seed=17)

A text_redactor may accept either text or text, state. The latter form can call state.redact_spans(...) to reuse the adapter’s pseudonym state. Reports contain only schema versions and aggregate counts; source surfaces are not copied into logs, exceptions, reports, or span representations. Directly constructed reports also reject noncanonical schema versions. Metadata is intentionally preserved as non-content data, so callers should keep metadata itself free of raw sensitive values.