/methods/named-source-no-structured-citation
The source was named; the citation was not recorded
A model can use a governed source, name it, and even link it in prose — and still return an empty machine-readable citation field. Those are two different facts. This page records one measured observation of exactly that, and is careful about what it does and does not claim.
The observation
Across every measured governed remote-MCP research generation where a structured citation was expected, the provider’s machine-readable citation annotation came back empty: 0 structured citations recovered out of 19 citation-expected slots, over 37 distinct generations, spanning two different model identifiers and two different source worlds. In those same runs the model’s own answer named — and, in the real-URL run, linked — the correct governed source.
This is a measured result over the runs measured. It is not a claim that the provider will always behave this way, and it does not establish a permanent or documented provider contract.
Two facts, kept apart
- Source use was observed. In every case where a citation was expected, the answer named the correct governed source.
- Structured citation recovery was not observed. The machine-checkable citation annotation was empty in every one of those same cases.
A governed tool being used and named in prose is not the same claim as a structured, machine-checkable citation being present in the response. Reading receipt validity as a citation finding would be the same mistake in a different place: a valid receipt attests to its own structure, not to whether a citation was recorded.
What was measured
| Source | Generations | Citations expected | Structured citations observed |
|---|---|---|---|
| Hermetic fixtures (3× repetition of 12 cases) | 36 | 18 | 0 |
| Real public court-record URL (single run) | 1 | 1 | 0 |
| Total distinct | 37 | 19 | 0 |
The result holds across three axes of variation — two model identifiers, a hermetic bounded-fixture world and a real public URL, and two separate frozen commit pins. The real-URL run in particular refutes the simplest explanation (that a fixture locator merely isn’t a public URL the annotation resolves against): even a real, public, model-named URL produced an empty structured annotation.
What this does not claim
- Not that the provider’s citation behavior is a fixed, permanent, or documented contract.
- Not that structured citation is “permanently empty” on this transport — only what was measured, over the runs measured.
- Receipt validity is not a citation finding; source use is not citation recovery.
Provenance
This page recomputes nothing. It restates a byte-pinned empirical note and two committed evidence objects held in the research runtime (the sanitized single real-run capsule and the combined measured tally). The underlying numbers are frozen there; this page is only their reader-facing summary.
Related
A worked instance of this finding, bound to the frozen benchmark records, is the case “Did the AI use its sources — and cite them?”.