What entered?
The user request, source material, tools, model route, and relevant settings are identified.
Guth assurance makes AI-assisted work inspectable before it becomes an action, a memory, or a published claim. It does not promise that a model is always right. It preserves enough evidence and control to test the work, reject it, repair it, and understand its effects.
This page is an explanatory public surface. Its examples are deterministic local demonstrations and create no external effect.
Assurance becomes practical when it is attached to the work instead of added as a vague review at the end. Guth follows the input, admitted evidence, authority, and resulting effect as one record. A receipt can therefore explain both what the system did and what it deliberately refused to do.
The user request, source material, tools, model route, and relevant settings are identified.
Claims remain attached to exact sources, qualifiers, contradictions, and missing evidence.
Model output is not permission. Human and system authority are stated separately.
Drafting, storage, sending, publishing, spending, and deletion are distinct effects with distinct gates.
A receipt is not a screenshot of a confident answer. It is a structured account of the request, the evidence admitted, the route taken, any human review, and the effect boundary. Open each tab below. The card stays the same size so comparison does not move the surrounding page.
Ordinary model output often compresses many sources into one fluent paragraph. That makes the answer easy to read but difficult to audit. Guth keeps source identity, admitted spans, dates, revisions, and qualifiers available after synthesis. A reader can return from a claim to the exact material that supported it.
This matters when a source changes, when two sources disagree, or when the model quietly fills a gap. Unsupported language is held as a proposal rather than promoted into memory or authority.
The prose may change. The source locator, admitted evidence, and qualification state do not disappear with it.
An agent may know how to send a message, modify a database, publish a page, or call a paid service. That capability does not grant authority to use it. Guth separates tools from permissions and distinguishes reversible local drafting from consequential external effects.
The effect gate evaluates the requested action at the moment it would occur. Missing authority fails closed: the useful draft can remain available while the send, purchase, publication, deletion, or identity change stays held.
Prepare text, organize a proposal, or preview a change without leaving the current surface.
Write to an authorized record with provenance, ownership, and recovery information.
Transmit information to another person, system, provider, or public destination.
Spend, delete, deploy, publish, alter identity, or perform another difficult-to-reverse effect.
Evidence-backed work can still become stale. Guth records the conditions under which a claim, model route, instrument, or workflow was accepted. A changed source, model, dependency, permission, or deployment environment can trigger requalification rather than silently inheriting an old result.
Move the controls and watch the decision update. This is a simplified deterministic example of the same basic rule: evidence quality, authority, and effect severity are evaluated separately. Strong evidence cannot manufacture permission, and permission cannot repair missing support.
Authority is missing. The local draft may remain visible, but no canonical or external effect can proceed.
Guth instruments are small, named mechanisms with explicit inputs, outputs, limitations, and receipts. They can be used alone or combined into stacks. The four examples below cover source retention, claim support, authority, and failure collection; other instruments address memory, privacy, routing, drift, experiments, and requalification.
Preserves exact locators and source identity so a claim can return to the material it relied on.
Compares proposed language with admitted evidence, qualifiers, and unresolved gaps.
Separates capability from permission and identifies the principal required for an effect.
Retains rejected, repaired, and contradictory cases as evidence for future testing instead of hiding them.
The system uses precise words because “the AI did it” is not a sufficient technical explanation. These definitions are intentionally short; future glossary pages will connect each term to examples, diagrams, instruments, and implementation references.
Source material explicitly accepted as support for a particular claim or decision.
The permission to perform an effect, distinct from the technical ability to perform it.
The controlled record treated as current after ownership, provenance, and review requirements are met.
A change caused outside model reasoning, such as writing, sending, publishing, spending, or deleting.
A useful proposal that cannot advance because evidence, authority, review, or another gate is incomplete.
Evidence that a model, workflow, instrument, or claim met defined requirements under stated conditions.
A structured record linking inputs, evidence, route, review, decision, and effect state.
Repeating qualification after a meaningful source, model, dependency, permission, or environment change.
Retaining enough source identity and location information to inspect where a statement came from.