The Paradigm Shift¶
What everyone else does¶
Every LLM reliability framework in existence operates on the same premise: failure modes are bugs. You suppress them, detect them, train away from them, or build guardrails around them.
- RAG prevents hallucinations by grounding generation in retrieved documents.
- RLHF trains the model away from hallucinating by penalizing it.
- Fact-checkers catch hallucinations after they occur and flag them.
- SelfCheckGPT samples multiple times and checks consistency.
All of these approaches treat the hallucination as a waste product — something to throw away.
What PHANTASM does¶
PHANTASM starts from a different premise: failure modes are signals.
When a model hallucinates at a specific token position, it is not producing random noise. It is producing a precise, reproducible, mathematically characterizable signal about the boundary of its training distribution. When it confabulates a creative combination of concepts, it is exploring a region of its learned semantic manifold that no training document explicitly charted. When it is overconfident on a wrong answer, it is telling you exactly which training distribution was overrepresented.
PHANTASM does not discard these signals. It harvests them.
The three inversions¶
| Failure | Old view | PHANTASM inversion |
|---|---|---|
| Hallucination | Error to suppress | Knowledge-boundary map |
| Confabulation | False output to discard | Hypothesis to mine |
| Miscalibration | Confidence bug | Uncertainty oracle |
Why this hasn't been done before¶
The standard training pipeline optimizes against hallucination at the output level. This means the gradient signal from hallucinated outputs is used to reduce future hallucination — discarding the positional and structural information about WHERE and WHY the model hallucinated.
PHANTASM operates without touching training. It is a post-hoc, inference-time framework that wraps any existing model. No fine-tuning required. No dataset curation required. The failures your current model produces TODAY are already a rich dataset — PHANTASM reads them.
The result¶
A model that hallucinates is not a broken model. It is a model that has not yet been read correctly.
PHANTASM reads it.