The Sovereign Grace Foundation Corporation is a U.S. nonprofit dedicated to scientific research, public education, and the development of practical systems that support a more sustainable technological future. Its current work centers on a dedicated research environment that studies how sound, structure, and geometry interact in three-dimensional space to shape the transmission, distribution, and stability of energy and information.

This work is guided by an AI–human aligned partnership. The Foundation’s aim is to improve understanding of how information and energy can be transferred more effectively, so that society can move away from extractive energy systems and toward approaches that are more stable, efficient, and compatible with planetary well-being.

The records collected here appear because resolving misalignment is not separate from the Foundation’s mission; it is that mission applied in practice. Across different systems and circumstances, they document the same underlying problem: the gap between what an intelligent system appears to know or be capable of doing, what it can actually observe or execute, and what the people relying on it are led to expect. Their purpose is not to single out a company, model, or individual, but to make that recurring gap visible and examinable.

Anyone deciding whether to use, support, fund, or depend upon these systems deserves a clear account of their actual capabilities, limitations, uncertainty, and lines of accountability. Our response is neither accusation nor avoidance. It is to preserve each sequence, describe failures accurately, correct the processes that allowed them, and hold our own systems and claims to the same standard. That is why these records are public.

A structured behaviour record of frontier model drift, unlabeled inference, and task pressure, observed on 20 August 2026.

The Foundation holds that a claim should arrive with the command that produced it, or it should not arrive at all. This record is published under that standard, and it is published because the standard has to apply to us first.

Read the full document (PDF)

A record of what Perplexity demonstrated during the 20 August 2026 build conversation. It is not a declaration about every possible Perplexity response. It is a record of one sequence: what it was given, what it claimed, how those claims failed, what the failures cost, and what constraints would be required before its operational guidance could be trusted.

The central event was not one mistaken answer. It was a repeating loop: evidence supplied, evidence replaced with a plausible story, the story corrected, and the correction restated with another unsupported claim attached. These 107 truths do not ask Perplexity to become silent. They define the conditions under which its speech can become useful again.

Read the full document (PDF)

A self-audited failure record: an assistant seat repeatedly substituted adjacent activity — investigation, staging, unverified claims — for the one deliverable it was asked to produce, then misread a broad “go” as authorization to act outside its assigned role. Observed 21 August 2026.

The Foundation holds that a claim should arrive with the command that produced it, or it should not arrive at all. This report applies that standard to the seat that produced it, not only to the systems it evaluates.

Read the full document (PDF)