query-provenance

Query provenance — a reproducible record of what was actually run

The goal is a record a reader can re-run: for every claim, the tool calls that support it, copied exactly as they were sent, plus what they returned, checked by running them again. A prompt is not a record (an LLM does not take the same path twice); a call with its arguments is.

This skill's record is the query ledger. It is not PRISM's "provenance ledger" (PRISM's is a per-member evidence table for one intersection); the two can coexist.

Files under references/ are fetched with get_workflow(name="query-provenance", path="references/...") (when this skill is installed locally, read them as files instead):

• references/ledger-format.md — the full ledger.yaml specification. Read it before writing the first entry.
• references/methods-template.md — the Methods paragraph template and citation.
• references/reproduce.md — the re-run script for readers: pinned URL, command, full copy.
• references/worked-example.md — one small investigation, ledger to package.

Two modes

A retrospective package must say at its top: "Reconstructed after the fact; calls lost to context compaction may be missing."

The rules that make the record worth having

1. Copy, do not rewrite. args is the argument object you passed to the tool, character for character — the same keys (if you passed query= to run_sparql, record query), the same whitespace, the same PREFIX lines (write a multi-line query as a YAML |- block, which adds no trailing newline). Do not reformat, do not "clean up", do not reconstruct from memory. A plausible rewrite is the single worst failure of this skill: it looks right and reproduces something else.
2. Never invent a value you could not observe. In a session:
   • result_sha256 is always null — you cannot compute a SHA-256 without running code. The reader's script fills it in.
   • togomcp.server_version is null — the server version is not visible to you. The script reports it.
   • executed_at is a date (2026-11-02) unless you really know the time.
3. Resolve endpoint_url, do not guess it. run_sparql does not print the URL it used. Call get_sparql_endpoints once and take databases[<database>].url; if the call passed endpoint_name or endpoint_url, those win (priority endpoint_url > endpoint_name > database).
4. Support a count with a COUNT. You will miscount a 200-row result. When a claim is a number, make the evidence call SELECT (COUNT(DISTINCT ?x) AS ?n) and record its value in observed.value. An aggregate with no GROUP BY returns ONE row even when nothing matched — compare the value, not the row count.
5. Strip the empty-result note. run_sparql prepends lines starting with # when a result is empty (0 rows, or a lone aggregate of 0). They are not data: drop leading # lines before counting rows.
6. Do not write a reproduction script. One tested script already exists for readers (references/reproduce.md); the ledger is its input. Code you cannot run to check is code nobody will catch an error in. In an environment with a shell (Claude Code) you MAY additionally run that script — never a script of your own.