mud_server.translation.validator

Output validator for the OOC→IC translation layer.

OutputValidator takes raw LLM text from the renderer and decides whether it is suitable for storage as in-character dialogue. Unsuitable output is rejected (returning None) so the caller can fall back to the original OOC message.

Validation pipeline (applied in order)

  1. Empty check — blank string → None.

  2. PASSTHROUGH sentinel — the model signals that the OOC input has no meaningful IC equivalent (e.g. it was a command or meta-question). → None.

  3. Multi-line check — strict mode rejects immediately; non-strict mode takes only the first non-empty line.

  4. Quote stripping — some models (e.g. gemma2) consistently wrap output in "..." or '...'; these are stripped before the forbidden-pattern check so that legitimate dialogue is not rejected purely because of quoting style.

  5. Forbidden pattern check — strict mode only. Rejects outputs that look like emotes, stage directions, or parenthetical narration. These indicate the model has not followed the “one line of raw dialogue” constraint.

  6. Max-length enforcement — strict mode rejects; non-strict truncates.

  7. Final empty check — returns None if cleaning left an empty string.

Strict vs non-strict

strict_mode=True (the default) treats any constraint violation as a hard rejection and returns None. This is the recommended setting for production worlds because it guarantees that only well-formed IC dialogue is ever stored — at the cost of occasionally falling back to OOC when the model produces slightly imperfect output.

strict_mode=False makes a best-effort cleanup attempt for minor violations (multi-line → first line; over-length → truncate). Useful for low-stakes worlds or during prompt development.

Attributes

logger

PASSTHROUGH_SENTINEL

Classes

OutputValidator

Validates and cleans raw LLM output before storage.

Module Contents

mud_server.translation.validator.logger
mud_server.translation.validator.PASSTHROUGH_SENTINEL = 'PASSTHROUGH'
class mud_server.translation.validator.OutputValidator(*, strict_mode, max_output_chars)[source]

Validates and cleans raw LLM output before storage.

_strict_mode

When True, any constraint violation → None.

_max_output_chars

Hard ceiling on IC output character count.

Initialise the validator.

Parameters:

  • strict_mode (bool) – Reject on first violation vs. best-effort cleanup.

  • max_output_chars (int) – Maximum allowed character count in the IC output.

validate(ic_raw)[source]

Validate and clean a raw LLM response string.

Runs the full validation pipeline and returns either a clean IC string or None. A None return is always accompanied by a WARNING log entry so that rejection reasons are traceable.

Parameters:

ic_raw (str) – Raw text returned by the renderer.

Returns:

Cleaned IC string on success, None if validation fails.

Return type:

str | None