ADR-0007 — No language model constructs graph structure. Ever.

Status: Accepted · 2026-08-18 · Enforces RULES 007–008 · Immutable

Context

Cartograph's entire claim is that its edges are derived and provable. A single model-proposed edge poisons that claim for every edge, because users cannot tell which edges to trust. "Model inference" in the confidence ladder refers to a logistic-regression classifier over structural features (M08) — a deterministic, auditable statistical model — never a language model.

Decision

No language model may propose, construct, modify, or fill gaps in graph structure — not as a fallback, not as a hint, not behind a flag. The pipeline is: deterministic analysis first, evidence second, statistical inference third, the model last and never in the graph. A language model may explain already-derived subgraph evidence (ASK, M16), where it interprets and cites but cannot add.

Alternatives

• LLM-assisted edge recovery for unresolvable URLs — rejected: an edge that cannot cite its derivation is a guess wearing a confidence score. Unknown stays unknown (RULE 009).

Consequences

Some real relationships will remain unresolved and the product must be honest about them — that honesty is the moat. ASK's design (M16) must pass evidence to the model, never accept structure from it.