MergeProof

Inspiration

A merged pull request is a strong reputation signal, but it can be hard for a non-maintainer to understand what that contribution actually proves. A badge can show activity without separating merged work from merely opened work, verified release inclusion from assumption, or technical evidence from persuasive prose.

MergeProof was built to make that boundary explicit.

What it does

MergeProof turns a public GitHub pull-request URL into an evidence-linked multilingual contribution dossier.

GitHub remains the source of truth for merge state, changed files, linked issues, and release references. Apertus explains that bounded evidence for a human reader. Every accepted model claim must point back to supplied evidence. Unsupported claims are dropped or replaced with deterministic evidence-backed fallback text.

How we built it

The open-source JavaScript/Node.js application includes a strict GitHub PR URL allowlist, bounded evidence collection, an OpenAI-compatible Apertus adapter, structured JSON output, field-level evidence-reference validation, deterministic merge-state enforcement, exact release matching, multilingual tests, and submission preflight checks.

A reproducible real-world example uses risa-labs-inc/BossConsole#1681, linked to issue #1632 and release BOSS v9.5.25.

How we used Apertus

Apertus is the synthesis layer, not the evidence source.

We evaluated Apertus 1.5 8B locally through an MLX OpenAI-compatible server in English, German, French, and Russian. The successful compatibility run kept the GitHub-derived MERGED state stable in all four languages and produced zero invalid evidence references after sanitization.

The adapter can point at CSCS, Public AI, or a local/sovereign Apertus endpoint without changing the evidence pipeline.

Challenges

The hardest problem was preventing fluent model prose from becoming stronger than the evidence. Grounding therefore had to be a system property: bounded inputs, evidence refs, source-of-truth checks, explicit unknowns, and fail-closed validation.

Provider compatibility was another challenge. Different OpenAI-compatible servers vary in model identifiers, headers, and reasoning defaults, so MergeProof fails visibly instead of silently switching to another model.

Accomplishments

• 20 automated tests pass.
• Submission preflight passes with no blockers.
• Five repeated public GitHub evidence collections produced stable merge/file/issue/release facts.
• A four-language Apertus 1.5 8B evaluation is committed as reproducible evidence.
• Every accepted generated claim retains evidence references.
• Source code is Apache-2.0; non-code hackathon materials are CC-BY-4.0.
• The repository, technical report, evaluation artifacts, and judging alignment are public.

What we learned

Structured output is not grounding by itself. Grounding requires the evidence contract, prompt boundary, validation rules, fallback behavior, and interface to agree on what the model is allowed to say.

We also learned that sovereign AI workflows benefit from separating model serving from evidence collection: the same trusted evidence pipeline can work across different Apertus endpoints.

What's next

A public hosted Apertus-backed demo is the next deployment step, followed by optional GitHub App access for higher rate limits and private repositories. The core submission is reproducible from the public repository and committed evaluation artifacts.

Built With

• apertus
• github-api
• javascript
• mlx
• multilingual-ai
• node.js
• open-source