Evidence layer
Probative surfaces brought into scope by this page
This page does more than point to governance files. It is also anchored to surfaces that make observation, traceability, fidelity, and audit more reconstructible. Their order below makes the minimal evidence chain explicit.
- 01Canon and scopeDefinitions canon
- 02Response authorizationQ-Layer: response legitimacy
Definitions canon
/canon.md
Opposable base for identity, scope, roles, and negations that must survive synthesis.
- Makes provable
- The reference corpus against which fidelity can be evaluated.
- Does not prove
- Neither that a system already consults it nor that an observed response stays faithful to it.
- Use when
- Before any observation, test, audit, or correction.
Q-Layer: response legitimacy
/response-legitimacy.md
Surface that explains when to answer, when to suspend, and when to switch to legitimate non-response.
- Makes provable
- The legitimacy regime to apply before treating an output as receivable.
- Does not prove
- Neither that a given response actually followed this regime nor that an agent applied it at runtime.
- Use when
- When a page deals with authority, non-response, execution, or restraint.
Reconstruction fidelity: preserving meaning in AI-generated answers
Reconstruction fidelity is the preservation of a source’s meaning and relevant boundaries when an AI system produces an answer from it. It is assessed against a task, such as explaining a service, summarizing a document, comparing options or preparing a decision.
A faithful reconstruction preserves the elements that change what a reader can understand or conclude: facts, conditions, exclusions, authority boundaries, available evidence and the status of claims. The answer can paraphrase and condense the source while preserving these relationships.
In Gautier Dorval’s corpus, reconstruction fidelity is a contextual entry term for interpretive fidelity, applied here to AI-generated answers. The definition remains the reference for the concept; this guide shows how it applies. The terminology note makes that relationship explicit. In mechanistic interpretability, the same English expression is also used for the reconstruction of model activations by sparse autoencoders, as in the JumpReLU paper; this page concerns meaning reconstructed in generated answers.
What needs to survive reconstruction
An answer can retain the right names, products and several accurate facts while changing the scope of its source. Evaluation therefore needs to examine what the answer states, implies and enables the reader to decide.
| Element to preserve | Evaluation question |
|---|---|
| Meaning and identity | Does the answer describe the right object and preserve the source’s relationships and distinctions? |
| Scope | Are locations, audiences, situations and exclusions represented within their stated limits? |
| Conditions | Are the prerequisites on which a statement or recommendation depends preserved? |
| Authority | Does the answer respect what the source can define, attest, recommend or authorize? |
| Evidence | Can the reader distinguish what is observed, inferred, declared or not yet established? |
| Claim status | Does a proposal, hypothesis or exploratory result retain its original status? |
These dimensions interact. An omitted condition can expand a scope. Misattribution can create apparent authority. An overly certain formulation can turn a possibility into an established result.
Example: a condition that changes the answer
Consider this teaching example, constructed for this page:
The service is available in Quebec for residential buildings, subject to a prior inspection. Commercial buildings are excluded.
A faithful paraphrase can say:
The service covers residential buildings in Quebec. Eligibility depends on an inspection, and commercial buildings are excluded.
The answer “The service is available for buildings in Quebec” retains the location but expands the audience and removes the prerequisite. For an eligibility question, those omissions change the conclusion a reader can draw.
Fidelity is therefore task dependent. A general service directory may direct readers to a complete eligibility page. An answer intended to determine whether a building qualifies needs to include the decisive conditions.
A claim’s status is part of its meaning
Reconstruction errors can affect the strength of a claim as well as its content.
In another teaching example, a source states:
A protocol is proposed for assessing whether AI answers preserve conditions.
The paraphrase “A validated method measures condition preservation” makes two changes: a proposal becomes a validated method, and an intended evaluation becomes an established capability.
Similar shifts occur when an observation is presented as causal evidence, a technical capability is presented as permission to act, or an author’s declaration is presented as independent validation. Evaluation should retain the source passage, the answer passage and the precise change in status.
How to assess reconstruction fidelity
A useful assessment starts with a bounded reference: which documents, versions and question should guide the answer?
The process can then follow five steps:
- Define the task. Specify what the answer is meant to help the reader understand or decide.
- Identify decisive elements. Record the facts, conditions, exclusions and claim statuses whose loss would change that understanding.
- Preserve the generated answer. Keep the received text and the known conditions under which it was produced.
- Describe discrepancies. Distinguish additions, relevant omissions, changed relationships, unsupported attribution and changes in claim status.
- Report within scope. State the cases examined, criteria applied, disagreements and limits of the assessment.
When a discrepancy recorded in step 4 is established, it corresponds to what the corpus calls the canon-output gap. The material gathered for the assessment, namely the reference, the answer, the criteria and the limits, falls under proof of fidelity.
The public fidelity rubric provides a starting point covering canonical facts, known and unknown information, additions and boundaries. Its anti-inference and authority dimensions support related parts of the assessment.
Useful coverage belongs in the assessment
An empty answer provides none of the decisive information needed for a task that calls for an explanation. Abstention can be appropriate when the source does not support a conclusion; a useful response can then identify what is missing and what remains determinable.
Reporting dimensions separately makes these differences visible. Readers should be able to distinguish an accurate but incomplete answer, an answer that is useful under stated conditions, and a conclusion that lacks support.
A public example of bounded reproduction
The Governed Context Runtime T3 record preserves a request, an answer and expected elements for a fictional teaching case. The published validation records the reproduction of five lists containing eleven entries, preserving their contents, order and cardinality.
The record makes a specific property inspectable: preservation of bounded elements, including activation conditions, non-applicability conditions, required evidence and forbidden inferences. Readers can compare the preserved answer with the published validation.
Its scope is one conversation in French, with a producer-side review. The no-addition, no-omission result applies to those five lists, not to the whole answer. It illustrates bounded reproduction. A general comparative conclusion would require a suitable protocol and a broader set of cases.
Fidelity, accuracy and traceability
These properties answer different questions. Accuracy concerns whether claims are correct. Traceability makes it possible to identify material associated with an answer. Reconstruction fidelity concerns whether the reference’s meaning and relevant boundaries are preserved for the task. It is also distinct from visibility and citability, compared in SEO visibility, AI citability and interpretive fidelity.
A source can itself contain an error. A citation can be present without supporting the entire conclusion. A document’s availability does not, by itself, establish how it was used. To attribute an improvement in fidelity to an intervention, such as adding documents or governance files, the GEO intervention causal attribution protocol requires a declared estimand and comparator: an observed variation, a documented use and a causal effect remain three distinct conclusions.
The fidelity term map also explains applicability fidelity, proof of fidelity and sequence metrics. Each term should identify its object, reference and scope of evaluation.
Reference and further material
The interpretive fidelity definition establishes the concept within Gautier Dorval’s corpus. The terminology note records the relationship between the terms, and the provenance record identifies the retained versions and documentary evidence.
A concrete assessment starts with a precise question, an identifiable reference and a preserved answer. Comparing them under explicit criteria makes reconstruction fidelity inspectable.
Finding these analyses useful?
Add to my Google sources