v1.6.0

Released 2026-07-27.

Benchmarks release: a reproducible retrieval-evaluation harness, the first published LongMemEval-S result, and native (Metal GPU) QMD on macOS.

Added

• Retrieval benchmark harness (evals/): reproducible, auditable evaluation of Cortex retrieval end-to-end through the public REST API — ingest → compile → search → answer → judge, exactly as a user deployment runs it. Design principles:

  • Pinned configs: each run is a YAML file pinning dataset URL + SHA-256, search mode, answer model, judge model, and seed; a hash mismatch aborts the run.
  • Per-question vault isolation: the vault is wiped and QMD synchronously reindexed between questions, so retrieval only ever sees that question's haystack.
  • Honest judging: LLM-as-judge with strict yes/no verdict parsing (unparseable replies count as errors, never guesses), judge model enforced ≠ answer model, and the official LongMemEval per-type grading prompts verbatim (temporal off-by-one tolerance, preference rubric, knowledge-update semantics, abstention handling).
  • Auditable artifacts: per-question JSONL traces (question, retrieved context, answer, verdict, latency, recall@k from dataset evidence labels, per-phase token usage) streamed crash-safe during the run.
  • Blind validation: identifier-stripped sample packets for human review plus judge–human agreement (percent and Cohen's kappa).

• First published benchmark result — LongMemEval-S, hybrid search: 45.0% overall accuracy, 65.6% evidence recall@8 over all 500 questions, zero judge errors (answer model anthropic/claude-sonnet-4, judge openai/gpt-4o). Per-category table, methodology, and caveats: docs/benchmarks/README.md. Reproduce with:

  uv run python -m evals.run_longmemeval --config evals/configs/longmemeval-s-hybrid.yaml
  

• Native QMD mode (./scripts/start.sh --native-qmd / ./scripts/stop.sh): runs QMD directly on the host instead of in Docker. On Apple-silicon Macs this gives QMD Metal-GPU embedding — Docker Desktop cannot access the GPU, which made embedding-heavy workloads time out under CPU emulation. stop.sh detects and stops a native QMD automatically via its PID file.

Fixed

• Trace files with Unicode line separators: read_traces splits records on newline only. Model answers containing U+2028/U+2029 (written unescaped with ensure_ascii=False) previously corrupted JSONL parsing via splitlines().

Upgrade notes

• No breaking changes. The evals/ package is dev-only tooling and is not part of the published PyPI package's runtime surface.
• Benchmark runs require LLM_API_KEY (and optionally LLM_BASE_URL) for the answer and judge models named in the config; Cortex itself still runs fine with LLM_BACKEND=none.