Pillar II — Confabulation Mining Network (CMN)¶

Core idea¶

Confabulations are not random noise — they are paths through the model's semantic manifold that training data never explicitly charted. CMN mines these paths for novel, plausible hypotheses using a contrastive dual-encoder architecture.

Usage¶

from phantasm.core.cmn import ConfabulationMiningNetwork

cmn = ConfabulationMiningNetwork(
    vocab_size=50257,
    hidden_dim=256,
    novelty_threshold=0.45,
    plausibility_threshold=0.50,
)

hypotheses = cmn.mine(confab_ids, fact_ids, texts=[text], domain="drug_discovery")
for h in hypotheses:
    print(f"novelty={h.novelty_score:.2f}  plausibility={h.plausibility_score:.2f}")
    print(h.text)

Hypothesis fields¶

Field	Type	Description
`text`	str	The full confabulated text.
`source_concepts`	list[str]	Key concepts extracted from the text.
`novelty_score`	float	Distance from factual reference space [0,1].
`plausibility_score`	float	Internal semantic coherence [0,1].
`domain`	str	Application domain label.

Training CMN¶

from phantasm.training.trainer import PHANTASMTrainer, TrainingConfig

config = TrainingConfig(epochs=10, learning_rate=1e-4)
trainer = PHANTASMTrainer(cmn, train_dl, val_dl, config)
history = trainer.train()