Current limitations

Perseval is a first public beta. These limits are part of the current product contract.

• The public application is macOS-only and requires macOS 13 or newer.
• The first artifact targets Apple silicon.
• There is no automatic updater or Homebrew cask.
• Release notes may identify an ad-hoc signed, non-notarized beta.

• OTLP/HTTP JSON and protobuf are supported; OTLP/gRPC is not.
• The embedded receiver is loopback-only and follows the GUI lifecycle.
• Codex session tailing, resumable JSONL ingestion, Langfuse import, and LangSmith import are not shipped.
• Semantic adapters cannot make inconsistent vendor schemas perfectly equivalent.

• Deterministic detectors are conservative and can abstain when facts are missing.
• Contract-suite behavior does not establish population-level precision, recall, or prevalence.
• Feature-similarity cohorts are secondary navigation, not canonical groups or risk probabilities.
• Hosted analysis plumbing is tested, but no held-out study proves a quality improvement.
• Provider jobs are not a durable retry ledger.
• Perseval does not automatically load repository .env files.

• Perseval creates and reviews eval definitions.
• It does not materialize a runnable test fixture from an accepted definition.
• It does not execute an eval against an agent.
• It does not checkpoint, fork, resume, or repair an agent sandbox.
• It does not produce a remediation-verification report from a new execution.

• MCP is local, stdio, and read-only.
• Compute, write, and payload-reveal permissions are intentionally unavailable.
• MCP cannot accept findings, create evals, or modify settings.

• Payload externalization is not redaction or encryption.
• There is no automatic retention cleanup or orphan-blob garbage collection.
• There is no in-app destructive workspace deletion control.
• Replacing or uninstalling the app leaves workspace data in place.

• AccessKit exposes semantic controls and keyboard journeys are exercised, but a complete human VoiceOver speech/rotor study is not published.
• Very narrow, high-text-scale, loading, backpressure, and recovery states continue to receive beta QA.
• Some dense comparison and evidence layouts still have visual-hierarchy polish work.

Perseval is an investigation and eval-definition workbench. Sandbox checkpoint/fork/resume, agent deployment, hosted evaluation execution, and autonomous remediation belong to other systems unless and until an explicit future integration is designed.