Changelog

TokenTriage has grown in capability waves. This page summarizes what each wave added, by theme; dated release notes are published on this page.

The trust dashboard

The embedded, self-contained SPA on the admin plane — the same numbers as the CLI, with honesty on every pixel: the typed Money component with pricing provenance, the EST flag, UNPRICED as an excluded slice, the realtime Live feed, provider-bill reconciliation, budgets and alerting, the robust-seasonal-v1 anomaly detector, and demo mode. See The trust dashboard.

Provider breadth

The August 24 catalog snapshot had a total of 92; its driver count was 11. The current catalog carries 95 built-in provider descriptors over 15 wire-dialect drivers, so adding an OpenAI-compatible upstream is N+1-by-config. The generated matrix names the test behind each gated cell and prints the ungated ones rather than hiding them, ten descriptors having none; the tests that do exist run against mocks rather than live endpoints. A driver is counted where an internal/dialect package name equals the dialect id a descriptor names. Azure, Bedrock (SigV4 + binary event-stream), and Gemini/Vertex land here, along with opt-in cross-dialect translation and an async poller that settles a submitted job’s cost on a second ledger row. See Providers.

The routing policy language

A typed, deterministic policy language with a first-class explanation on every decision — five node kinds, a frozen condition language, and the strategy recipes. A legacy routing.rules config desugars to it and decides byte-identically. See The routing policy language.

Response & semantic caching

An LLM-router cache that is off by default, shadow-first, byte-faithful, and costed rather than zeroed, with air-gapped exact/normalized tiers and an opt-in, bounded semantic tier whose savings are licensed by evidence. See Response caching.

Virtual keys & governance

Minted virtual keys and per-tenant budgets, quotas, rate, and concurrency limits, built so a refusal can never lie — row == wire, fail-open with disclosure, and a shadow-first enablement ladder, with an opt-in distributed tier for multi-replica coordination. See Governance.

LLM judges

A model that grades output, and the machinery that grades the grader. A judge’s verdicts land on the ledger at reported basis, carrying a disclosure that names the model, the prompt version and whatever validation stands behind it. Cohen’s kappa against a human gold set, inter-annotator agreement measured on the overlap, and a position-swap gate together decide whether a judge subject promotes, and a promotion is what lifts the ceilings holding its figures at reported. Panels vote across model families under a same-family refusal; a verdict the method cannot settle is counted as an abstention with its own named reason rather than averaged into a pass rate. See LLM judges.

---

The complete, versioned history — including the compatibility guarantees re-proven each release — is published on this page as release notes.