What the study found
SΔϕ-39 defines hallucination as a world-binding mismatch within the Sofience–Δϕ Formalism Series. The central claim is that AI hallucination is not merely false output, but weakly bound construction, inference, or unresolved model residue presented as strongly bound fact.
Why the authors say this matters
The authors say the framework supports AI hallucination audit, citation verification, summary grounding, fiction versus hallucination analysis, world-binding strength testing, claim strength calibration, UMR preservation, Slop prevention, and agentic execution precheck. They also say it helps distinguish fiction from hallucination without suppressing hypothesis generation or demanding impossible certainty.
What the researchers tested
This AI-readable package revises and operationalizes the original SΔϕ-39 paper, Fiction, Hallucination, and the Uninhabited Construction. It adds an output-level formula, Hallucination = ClaimStrength > EvidenceBinding, and breaks the framework into operational files for AI ingestion, hallucination audit, fiction-versus-hallucination distinction, world-binding strength testing, claim strength versus evidence binding comparison, trace/inference/UMR failure modes, and routing to related SΔϕ papers.
What worked and what didn't
The package preserves the original claim that hallucination is construction proceeding without an adequately integrated world-binding stop mechanism. It also defines hallucination as weak inference spoken as world-bound fact, inference output as observed trace, and unresolved model residue (UMR) erased into assertion. The abstract does not report experimental validation or quantitative outcomes.
What to keep in mind
The framework does not claim that all errors are hallucinations, does not punish fiction, does not suppress hypothesis generation, and does not demand impossible certainty. The available summary does not describe limitations beyond these scope statements.
Key points
- Hallucination is defined as weakly bound content presented as strongly bound fact.
- The package states that hallucination is not simply false output.
- An explicit formula is given: Hallucination = ClaimStrength > EvidenceBinding.
- The framework is intended for hallucination audit, citation verification, and fiction-versus-hallucination analysis.
- The abstract does not report experimental validation or quantitative results.
Disclosure
- Research title:
- Hallucination is defined as weak world-binding claimed as fact
- Authors:
- Sofience
- Publication date:
- 2026-05-14
- OpenAlex record:
- View
Get the weekly research newsletter
Stay current with peer-reviewed research without reading academic papers — one filtered digest, every Friday.