Changelog
Notable changes to the TDQS specification. Scores are calibrated to a rubric+model pair, so each entry states what re-scores.
v1.2 — 2026-09-01
The first three changes respond to #3. Nothing here re-runs an LLM call, and only the last change moves a stored score.
- Parameter Semantics guidance is now channel-neutral. The checklist told maintainers to document parameters in the schema, not the description. The rubric never required that: low coverage shifts the burden onto the description, and either channel earns the credit. Guidance only; no prompt changed.
- New context signal
definitionBytes: the tool's share of thetools/listpayload. Stored and published, withheld from every prompt, priced by nothing. It derives from the serialization already built forinputHash, so backfilling it is a one-off pass keyed by row and re-scores nothing. - Fixed an overlap in the Tool Count Appropriateness anchors, which claimed 25 tools as both
16-25(score 3) and25+(score 2). The upper band now starts at 26. Affects only servers with exactly 25 tools. - Rounding is now defined: integer half-up, everywhere.
round1(p, q)rounds the exact rationalp / qhalf-up to one decimal in integer arithmetic.computeTdqsand the three server rollups all use it, and the rollups are restated with integer arguments. Before,computeTdqsaccumulated float weights and calledMath.round: for 612 of the 7,500 score vectors whose exact sum ends in.X5, the float sum lands just below the tie and rounds down, and 134 of those land a tier low. The rollups had no definition at all, and the coherence mean is a tie for every odd dimension sum. Dimension scores are untouched; affected composites get a deterministic recompute with no LLM call.
Deferred until the definitionBytes corpus run reports: a calibration example for low-coverage-but-documented parameters, and rebasing the tool-count anchor on domain breadth rather than raw count. Both edit prompts and therefore re-score the registry, so they ship as one batch — and only if the corpus shows the rubric actually rewards length.
v1.1 — 2026-08-23
Added shadowing risk: a flag for a tool whose purpose is substantially covered by a sibling that is materially cheaper to invoke.
- New context signals
requiredFieldCount,schemaDepth,unionChoiceCount, and the derivedinvocationCost— stored on every tool, withheld from the tool-scoring prompt. - A deterministic prefilter emits at most one candidate pair per tool; the existing server coherence call confirms genuine purpose overlap.
- New
Shadowing Riskserver flag (serverFlags) andshadowingRiskson the server record. It changes no score. - Coherence gains a second re-run trigger: a change to the server's candidate set.
v1.0 — 2026-06-07
Initial publication: the six-dimension tool rubric, hard gates, deterministic post-processing, four-dimension server coherence, the server rollup, and both prompts verbatim.