Before an AI agent edits a repo, give it evidence it can audit

A useful context pack for an AI coding agent should be boring in exactly the right places.

The Doramagic AI Context Pack Benchmark is built around one constraint: before an agent edits a repository, it must be able to separate claims from evidence.

The package contains:

• CLAIM_GRAPH.json: what the project claims, and how those claims relate.
• EVIDENCE_INDEX.json: where each claim is supported in files or reports.
• CAPABILITY_CONTRACT.json: what the pack says the agent can and cannot assume.
• AI_CONTEXT_PACK.md: the human-readable operating context.
• CONTINUE_CHECK.md: the handoff check for the next agent run.

The important part is the negative space. The pack does not pretend to install or execute the target project if that was not verified. It does not turn a README into runtime proof. It keeps source packaging, evidence indexing, and validation reports separate so a downstream agent can say: this is proven, this is inferred, and this still needs a real run.

That changes the repo handoff loop. Instead of asking the model to “understand the codebase”, ask it to answer a smaller set of questions first:

1. Which claims are source-backed?
2. Which files are evidence, and which are just narrative?
3. Which capabilities are explicitly out of scope?
4. What was not executed?
5. What should be checked before making a code change?

Repo: https://github.com/tangweigang-jpg/doramagic-ai-context-pack-benchmark