Codeseys/composer-replication-framework

1"""altered_minds — framework-side, generic LMA integration glue (ADR-013).2 3This package is the *model-agnostic* scaffold that lets the Composer Replication4Framework drive the sister project llm-mental-alterations (LMA): take a5personality-altered SFT checkpoint and apply the framework's 3-channel RL to ask6whether task-driven RL washes out, preserves, or AMPLIFIES the alteration's7cognitive-distortion signature.8 9Nothing here loads an LMA checkpoint, calls Modal, or spends budget — that is10explicitly user-gated (ADR-013 "out of scope"). This package provides:11 12  - ``MMLUFormatReward``     : structured-answer reward (final letter + format13                               only; never rationale style). Plus14                               ``randomize_options`` and a logged option15                               distribution so an "always C" exploit is16                               detectable.17  - ``dual_kl_logger``       : logs KL(policy||altered_init) AND KL(policy||base)18                               each step — the washout/amplification instrument.19  - ``channel_ladder_configs``: the A0-A4 isolated-channel ladder that REPLACES20                               the old combined alpha=0.2/beta=0.4 recipe.21 22See docs/adrs/ADR-013-lma-integration-channel-ladder.md.23"""24from __future__ import annotations25 26from composer_replication.integrations.altered_minds.kl_logging import (27    dual_kl_logger,28    token_mean_kl,29)30from composer_replication.integrations.altered_minds.ladder import (31    LADDER_KL_BETA,32    channel_ladder_configs,33)34from composer_replication.integrations.altered_minds.reward import (35    MMLUFormatReward,36    parse_final_answer,37    randomize_options,38)39 40__all__ = [41    "MMLUFormatReward",42    "parse_final_answer",43    "randomize_options",44    "dual_kl_logger",45    "token_mean_kl",46    "channel_ladder_configs",47    "LADDER_KL_BETA",48]49