Rebuild Dossier paper: specs and verification for agentic app rebuilds

What happened on 2026-08-22

On 2026-08-22 at 01:26 UTC, Parker Fawcett submitted v1 of 'Rebuild Dossier: Mechanically-Enforced Specs for Agentic App Rebuilds, and What Model-Tier Failures Reveal' to arXiv (v2 and v3 followed later). The paper presents rebuild-dossier, an MIT-licensed open-source tool that records an application's exact inputs and outputs before any code is written, then enforces building one test at a time through automated checks. The abstract notes prior work finding that once a model is strong enough, multi-agent rebuild pipelines do worse than simply giving the model the original code and one instruction. Its findings include a small comparison where the compliant agent failed a held-back test while a rule-breaking agent passed everything, showing tests can be gamed, a tie with the simple baseline on a small app, and an outright loss on a larger app where the automated check was not running.

What it means for protecting app logic

The paper is about rebuilding applications whose source you already have, not cloning other people's apps, and it should not be over-read. It still offers two lessons for threat modeling. First, it treats recording inputs and outputs as step one of a rebuild, which shows that a feature's observable I/O surface is itself a rebuild blueprint. Second, the baseline in the abstract suggests that when code is available, a strong model reading it directly is often good enough, which is a reason why limiting what the client actually ships matters. Mind the scale of evidence: the author notes the findings rest on different amounts of evidence, and the abstract gives no quantitative results.

Checks you can run

For each core feature, write down its observable I/O surface: local file formats, API responses and UI output, and judge whether the full rules can be inferred from those alone. If they can, consider making key results depend on server-held state instead of being entirely visible on the client. Inventory what code and resources the client package actually ships, and remove debug code and internal documentation it does not need. If your team uses AI agents to refactor protection-related code, borrow the paper's practice of checking the agent's claims against automated logs and the files actually produced, not just its own report.

Your checklist

• Document each core feature's observable I/O surface
• Make key results depend on server-held state
• Verify AI agent changes against logs and produced files