tdqs / corpus
Scores are calibrated to a rubric+model pair. These are the worked examples the prompts carry, which is what an implementation checks itself against before trusting a model, plus what the rubric found when it ran across a registry.
Four definitions from Appendix A of the specification, with the scores the rubric assigns them. An implementation that lands materially elsewhere on these is not running TDQS, whatever it calls itself.
List ALL calls in date range – no user/workspace filtering. To filter by user/workspace, use search_calls_extensive instead.
annotations { readOnlyHint: true, destructiveHint: false } · 2 parameters, schema coverage 100%, no output schema
Update shared drive settings including name, color, and restrictions.
annotations none · 4 parameters, schema coverage 50%, no output schema
Process
annotations none · 3 parameters, schema coverage 0%, no output schema
Creates a new record in the database
annotations { readOnlyHint: true } · the description claims a write while the annotations declare a read
From Appendix B: two whole servers, and the two shadow candidates the prefilter would propose, one confirmed and one rejected.
list_repos, get_repo, create_repo, search_code, list_issues, create_issue, update_issue, get_pull_request, create_pull_request, merge_pull_request
disambiguation 5, naming 5, count 5, completeness 4 (PR review and comments missing, core workflows covered)
process, run, execute, do_thing, helper, transform_data, processV2, handleRequest
disambiguation 1, naming 2, count 3, completeness 2 (process/run/execute/do_thing are indistinguishable)
get_stat, query_panel
query_panel (cost 13) is shadowed by get_stat (cost 4): both answer "what is X for player Y", and the cheaper one takes four flat scalars
get_repo, search_code
search_code (cost 8) against get_repo (cost 2): costs are asymmetric but the purposes do not overlap
Glama runs TDQS over every tool it indexes. As of June 2026 that was 228,369 tools across 15,036 servers, with a mean TDQS of 3.57 and 73.5% clearing the tier-B bar — and 56.4% of tools carrying at least one smell, most often on usage guidelines and behavioral transparency.
| tier | tools | share |
|---|---|---|
| A | 195,872 | 54.1% |
| B | 70,186 | 19.4% |
| C | 86,303 | 23.8% |
| D | 9,865 | 2.7% |
The deterministic stages are exact: an implementation must reproduce every context signal, hash, invocation cost, shadow candidate and rollup for these examples to the digit, and the reference implementation's test suite asserts that it does. The model stages are calibrated rather than exact: after swapping models, spot-check the scores above before trusting the output, and expect to re-score anything you compare against.