Grant track: Agent Governance Conformance Lab

[safal207]

Goal

Turn the existing LS/CrewAI governance evidence into a credible, reproducible grant package for an Agent Governance Conformance Lab.

Core portfolio:

• Grant portfolio
• Root grant entry point
• CrewAI conformance profile
• CrewAI PR #6030
• CrewAI issue #6063

Grant thesis

Critical agent-safety invariants should be executable at the serialized contract boundary, portable across runtimes, fail-closed, reproducible in CI, and explicit about semantic gaps.

Phase 1 — Evidence hardening

• Track CrewAI PR #6030 until merge, close, or redesign
• Contribute at least one upstream pytest fixture when the intent-binding fields land
• Capture the exact upstream diff or maintainer statement showing review impact
• Add a compact evidence table: claim → artifact → independent confirmation → status
• Add CI output examples for the existing CrewAI fixture pack
• Document explicit non-claims and known limitations

Phase 2 — Portability proof

• Select a second agent runtime or tool gateway
• Define the smallest adapter interface
• Run the same core vectors against both ecosystems
• Record representable, lossy, and unrepresentable mappings
• Measure false-allow, false-deny, revalidation, and defer behaviour

Candidate second ecosystems:

• LangGraph / LangChain
• AutoGen
• OpenAI Agents SDK
• MCP tool gateways
• another open-source runtime with pre-tool authorization hooks

Phase 3 — Benchmark package

• Publish versioned JSON Schema
• Publish canonical positive and negative vectors
• Add deterministic Python CLI runner
• Add GitHub Actions integration example
• Produce a small reproducible evaluation report
• Create a one-command demo suitable for reviewers

Phase 4 — Grant application assets

• One-page project summary
• 90-day work plan
• Budget and resource assumptions
• Risk register
• Public-benefit and open-source impact statement
• Maintainer / collaborator letters or public endorsements
• Bio variants: 50, 100, and 250 words
• Application-specific evidence appendix

Success criteria

Minimum credible grant package:

1. one acknowledged upstream design impact;
2. one merged upstream contribution or official documentation reference;
3. two runtime adapters;
4. one published comparative benchmark;
5. deterministic reproduction through CI;
6. disciplined claim boundaries.

Immediate next action

Prepare an upstream-ready pytest translation of the four CrewAI fail-closed vectors:

1. exact-intent mismatch;
2. target-state drift;
3. continuation mismatch;
4. duplicate successful outcome / idempotency violation.

Do not submit until the committed CrewAI contract fields are stable enough to avoid coding against a provisional schema.