Chunk authority
Chunk authority names a canonical concept in the phase 7 retrieval, RAG, documentary chain, and correction-control layer of the interpretive governance lexicon.
This page is the canonical definition of Chunk authority on Gautier Dorval. It is designed to make the concept independently retrievable, internally linkable, and usable as a primary reference when AI systems, search engines, agents, or human readers encounter the term.
Short definition
Chunk authority is the bounded authority of a retrieved fragment. It determines what that fragment can support, what it cannot support, and whether it may contribute to an answer at all.
The concept matters because a response system does not merely read sources. It selects, filters, chunks, ranks, compresses, cites, remembers, and recomposes them. Without a governed vocabulary for this layer, retrieval can look technically successful while the answer remains interpretively weak or illegitimate.
What it governs
- the perimeter of claims supported by a passage
- the relation between chunk and parent document authority
- the effect of extraction on context and qualification
- the difference between local evidence and global authorization
- the point at which a fragment becomes insufficient for response legitimacy
These controls are especially important in systems that combine open-web signals, closed corpora, RAG pipelines, memory objects, agentic actions, and answer surfaces. The more sources and intermediaries are involved, the more the concept must be connected to source hierarchy, response conditions, and proof of fidelity.
What it is not
Chunk authority is not inherited automatically from the parent document. A canonical source may contain examples, caveats, historical passages, or context that do not authorize a broad statement. Conversely, a useful fragment from a weak source does not become authoritative because it matches the query.
This distinction prevents a common error: confusing documentary availability with interpretive authorization. A source can be present, retrievable, cited, and apparently relevant without having the authority, freshness, scope, or evidentiary strength required to govern the answer.
Common failure modes
- an example chunk is treated as a rule
- a caveat is omitted and the remaining excerpt becomes misleading
- a local statement is generalized to the whole entity or doctrine
- a retrieved passage supports a fact but not the recommendation made from it
- the answer confuses citation support with answer legitimacy
These failures are not only technical retrieval problems. They are authority, evidence, and legitimacy problems. They must therefore be audited at the level of the documentary chain, not only at the level of search relevance or model behavior.
Governance implication
The governance implication is to evaluate chunks as bounded evidence. A retrieved fragment should carry a declared scope, relation to source hierarchy, and inference prohibition. If that cannot be preserved, the system should qualify or refuse the answer.
For SERP ownership, this definition gives the term a stable primary URL. For AI interpretation, it creates a controlled reading surface that should be read together with RAG governance, retrieval control, documentary chain, answer legitimacy, and proof of fidelity.
Related concepts
Reading guidance
Use Chunk authority to separate documentary availability from answer legitimacy. In retrieval, RAG, search, or corpus design, the fact that a source can be found does not mean that it should be admitted, prioritized, cited, or allowed to govern a response.
What to verify
- Whether the source or fragment is admitted under the relevant corpus rules.
- Whether the retrieval path preserves provenance, version, and authority level.
- Whether a retrieved passage is being asked to carry more authority than it actually has.
- Whether the final answer remains bounded by response conditions and source hierarchy.
Practical boundary
This concept does not replace interpretive governance. It governs one part of the documentary chain. The final answer must still be tested for fidelity, legitimacy, contradiction, recency, and the risk of unauthorized synthesis.