v0.3.0

Added

• Eval-backed routing: offline replay with recommendation quality-gating (#79), durable
  eval worker with masked-fidelity labeling (#88), reusable benchmarks with a model
  leaderboard and quality-per-dollar view (#115), judge-reliability checks for position
  bias and decisiveness (#119), a "disprove it" precision pass on worse verdicts (#117,
  #120), and severity-weighted benchmark curation with pinned problem cases (#123).
• Canary rollouts and shadow evals: mirror live traffic to a candidate model (#54),
  promote or roll back experiments, gated on passed evals (#84).
• Gateway surface: POST /v1/embeddings with Gemini and OpenAI backends (#135) and an
  OpenAI Realtime WebSocket proxy at POST /v1/realtime (#137). SDKs released at 0.5.0
  with embeddings support (#136).
• Per-user attribution: userId and department on API keys (#131), body.user
  mapping on the OpenAI-compatible path (#126), and a team usage guide (#129).
• Governance: output guardrail keyword/regex deny-list with per-app scoping (#66) and a
  three-section governance overhaul covering guardrails and observability (#62).
• Observability: latency benchmarking with TTFT and gateway-overhead tracking (#86),
  p50/p95 latency stats with per-app trends and request search (#60), a latency
  breakdown table on Overview (#118), and per-request routing explainability (#44, #48).
• Cost controls: max_tokens clamped to each model's output ceiling (#46, #47), budget
  warnings with CSV export and cost trend (#52), goal-driven policy generation with an
  impact simulator (#42), and a 1-10 difficulty score captured per request (#40).
• Engineering: server unit + integration test suite and testing strategy (#58), routing
  benchmark harness across LiveBench, Arena-Hard, and SWE-bench (#63), and a gated,
  image-based, rollback-safe deploy pipeline (#61, #74).
• Security & hardening: CI gates for secret scanning, npm audit, CodeQL, and Trivy image
  scans (#151); graceful shutdown on SIGTERM/SIGINT (#152); a production profile that
  fails closed instead of booting without required secrets (#159); metadata-only logging
  as the production default (#160, #162).
• Guardrails & caching: built-in and custom per-app prompt injection detection (#154), and
  a semantic response cache with a configurable similarity threshold (#155).
• API keys: expiry dates and one-click rotation (#156).
• Access control: OIDC/RBAC accountable admin access with a per-user audit trail (#163),
  an allowed-domains setting for OIDC clients shared with other apps (#165), and SSO/RBAC
  documentation (#170).
• Providers: LiteLLM connection type is now discoverable from the UI (#171).
• Ingestion: an observe-only ingestion API (#172).
• Unified optimization workflow: a recommendation's lifecycle stage (opportunity through
  accepted, shadow, canary, or rolled back) is derived on read from its linked records
  rather than stored, with realised-vs-projected outcome tracking (#180).
• Exportable recommendation evidence report, as JSON or Markdown, assembled on demand from
  every linked dataset, eval run, campaign, experiment, and audit entry (#181).
• Design-partner demo fixture: a full opportunity-to-canary-to-rollback story seeded with
  zero provider keys, using the same production aggregation/judging code as real evals
  (npm run demo:seed / demo:reset) (#182).
• Internal spend accounting: every LLM call Arbr makes for itself (classification, policy
  generation, judging, eval replay) now goes through one wrapper that tags and prices it,
  keeping it out of customer-facing analytics (#183, #184).
• Cloud secret-manager integration: any credential-shaped env var can hold a gcp-sm://...
  reference instead of a literal, resolved transparently at boot and on a periodic or
  on-demand refresh; production fails closed on a resolution failure. AWS/Azure adapters
  are documented against the same interface (#189).
• Operational readiness package: a /health/ready readiness check distinct from the
  existing liveness check, backup/restore scripts, config/policy export and import, a
  support-diagnostics bundle with no credentials or captured payloads, a disk-usage guard
  on deploy, and bounded container log growth (#190).
• OpenTelemetry trace export: OTLP spans per request, off by default, with sample-ratio
  and content-capture controls adjustable at runtime from the Governance page (#191, #192,
  #193).
• A projectarbr.org link in the web console's sidebar footer (#186).

Changed

• Console reoriented around the five stages (Connect, See, Recommend, Route, Govern)
  with a redesigned sidebar and Recommendations page (#98, #111, #112) and a restyle to
  the ARBR monochrome brand (#144-#146).
• Hot-path latency reduced via a Settings cache and parallel pre-LLM fetches (#85).
• Web stack migrated to Tailwind CSS v4 (#70).
• Roughly twenty documentation pages corrected for drift against the actual API, config,
  and pricing (#173, #175-#179); the README's autonomy claims and quickstart walkthrough
  clarified (#194); the README dashboard screenshot refreshed (#187).

Fixed

• Eval judge was passed a string instead of a messages array, so judging never worked;
  fixed along with re-run support (#108).
• Per-app config (policy, opt-out, kill switch) now applies on /v1/chat/completions
  (#39).
• /api/about 500 that broke the deploy seed-version check (#89); /health demoMode
  now reflects effective provider state (#147).
• Audit page blanked on the new object-shaped AuditLog.actor (#164).
• Docker Compose wasn't passing the accountable-admin-access auth env vars through
  (#166).
• Arbr's own AI spend was polluting customer analytics views (#167).
• Sign-out showed the admin-key form even in OIDC/trusted-header mode (#168).
• Sidebar tagline pushed Users out of view without scrolling (#169).
• handler.js logged a semantic_cache routing decision the schema's enum didn't
  include, silently dropping every semantic-cache-hit log row (#174).
• Sidebar footer links stacked into two rows after adding the projectarbr.org link,
  pushing Users below the fold again; put back on one row (#188).