Roadmap: improve evals, platform coverage, and AI reviewer layer

[ftchvs]

Goal

Make AdLint more useful as a local-first, open-source ad preflight tool by improving the parts contributors and growth teams can build on next.

Contribution areas

1. Real-world/paraphrased eval cases

• Add public-source or paraphrased examples without private data.
• Include positive and near-miss cases for each policy.
• Keep labels explicit: expected decision, expected policy IDs, and reviewer notes.
• Prioritize Meta/Google/TikTok/LinkedIn examples from public policy docs and public ad-library style examples.

2. Deeper platform policy coverage

• Expand Meta, Google, TikTok, and LinkedIn modules without overclaiming parity.
• Prefer narrow, high-signal rules over broad keyword matches.
• Add source references and scope caveats in docs when new policy families land.
• Add false-positive guardrails for regulated-category education/tooling copy.

3. Optional AI reviewer layer

• Document the rule baseline + optional local AI second-pass architecture.
• Add evals that measure when AI adds useful findings vs generic/noisy findings.
• Track model status, invalid responses, added false positives, and rescued false negatives.
• Keep deterministic rules as the trusted baseline until evals prove incremental value.

Non-goals

• Legal advice.
• Guaranteed platform approval.
• Hosted model dependency by default.
• Raw ad persistence by default.

Acceptance criteria

• New policy PRs include docs, tests/evals, and near-miss cases.
• Eval reports continue surfacing false positives and false negatives.
• README/release language stays scoped to decision support.