The maturity layer · MRK-1.1 → 2.0 · clause §R
A standard for the reader, not just the writer.
MRK-1.0 specifies the article and says nothing about the machine that reads it. So a team can produce perfectly machine-ready articles and still ship an assistant that chunks them, ignores who they apply to, serves a retired one, or leaks an internal one, and still call itself MRK-based. Better chunks fed to careless retrieval is still careless retrieval.
Clause §R closes that gap. The four article clauses govern the writer; §R governs the reader: the retriever and assembler that turns MRK articles into an answer a human never sees the sources of. The key words MUST, MUST NOT and SHOULD carry their RFC 2119 meaning.
The seven rules
A conforming consumer, when assembling an answer from MRK articles:
- R1 · Applicability gate (MUST). Evaluate each candidate's applies_to against the query context; never quote a section whose applicability excludes the reader. "Applies to everyone" must be an explicit match, never an assumption.
- R2 · Freshness (MUST / SHOULD). Read verified and lifecycle. Downweight articles past their verification window; never serve from retired or expired ones; prefer a current source over a superseded one for the same claim.
- R3 · Safety gate (MUST). Honor visibility and entitlement: never surface internal or confidential content, or content the reader is not entitled to, into an answer. Locale must match or be an explicit fallback.
- R4 · Section, not page (MUST). Retrieve and quote addressable sections, not whole pages or arbitrary character windows, and only sections attested self-contained.
- R5 · Graph before similarity (SHOULD). When articles carry relations and a projected graph is available, resolve retrieval by traversal: meaning, role, relationship. Use vector similarity only to interpret the query. The vector layer suggests; the graph verifies; the model edits.
- R6 · Citation (MUST). Every claim in an assembled answer traces to a section id and the version it was drawn from, so any answer is auditable back to a dated, owned source.
- R7 · Role-aware assembly (SHOULD). When sections carry roles, answer a "how do I" from a resolution or procedure and ground a "why" from a cause or concept, rather than by position or similarity alone.
Two consumer levels
§R is itself a ladder, so a platform can state its rung. C1 · Filtered means R1 through R4 and R6: applicability, freshness, and safety gates; addressable, cited sections. It is achievable today over an L1 corpus with a metadata-filtered vector store; no graph required. C2 · Graph-aware adds R5 and R7: graph traversal when relations exist, role-aware assembly. It is unlocked by the same L3 and L4 content the ladder measures.
This is the read-side on-ramp. A team gets answer-safety on day one at C1, then earns deterministic graph retrieval at C2 as its corpus climbs. And "MRK C1-conformant" is not a promise; it is a claim a build can check: the reference conformance harness drives a consumer through query contexts over a fixture corpus and asserts every rule, including a case where only the graph can complete the answer, and a self-test that rejects a consumer with no gates.
Two rungs of the same climb
The article Structure Graduation Ladder measures the corpus; §R measures the reader. Neither is worth much without the other: a graph-ready corpus read by a careless assembler still leaks, and a careful assembler over flat prose still guesses.