Roadmap

HVTrust ranks open-source AI agents on five weighted dimensions: Safety/Integrity (25), Identity/Provenance (18), Transparency (17), Maintenance (20), Adoption (20).

• OSSF Scorecard from our own CLI scan
• npm + PyPI provenance attestation checks
• Signed commit ratio from GitHub verification
• Confidence-weighted evidence grades A–D

Agent profiles show a compact runtime snapshot sourced from public docs, manifests, and registry metadata. As of methodology v4.0 these fields are folded into the production HVTrust rank as a bounded calibration (see below).

• MCP server support — declared vs implemented server signals
• External service dependencies — model providers, storage, and third-party services
• Tool / plugin surface — broad capability and extension hints
• Package provenance drift — whether published package metadata still points back to the tracked repo

Runtime-trust signals are now part of the production HVTrust score and rank. The promotion was evidence-gated: an upset review plus several rounds of detector auditing (same-owner package variants, repo transfers, and docs-only mentions are excluded from scoring) preceded the cutover.

• Calibrated scoring — MCP support, external dependencies, tool/plugin surface, and package-provenance drift adjust the base score within published bounds
• Public methodology — every adjustment value is documented in the methodology and the runtime-trust spec
• Comparable baseline — the pre-calibration ranking stays available on the leaderboard for comparison

Runtime signals already affect rank; the next step is showing them more clearly so readers can distinguish supply-chain posture from runtime reach and complexity.

• Per-agent capability surface page with clearer breakdowns
• Trend tracking for runtime drift over time

Watchlists and on-site alerts are live (track any agent from its profile). Next: a low-friction weekly digest of meaningful leaderboard and trust changes.

• Weekly “what changed” digest for the leaderboard
• Major trust/regression callouts when they truly matter
• Historical runtime changes and drift summaries for watched projects

Today corrections flow through GitHub issues. Longer term, maintainers should be able to claim listings, declare runtime fields with evidence, and respond to drift or provenance mismatches directly.

Tighten the gate at the top of the leaderboard: require Sigstore-style identity binding for projects to reach the verified tier.

Lightweight, opt-in runtime telemetry that detects when an agent's actual behaviour drifts from its declared capabilities — without sending user data anywhere.