TL;DR
The paper introduces the Portable Agent Format (PAF), a new delivery unit for professional work that bundles final reports with their underlying evidence, context, and governance rules. By evaluating PAF-aware readers against standard LLM baselines, the authors demonstrate that this "compiled context" significantly improves question-answering accuracy and adherence to organizational boundaries, even for small, offline models.
What Problem Does This Paper Address?
When a professional delivers a work product—like a financial report or an engineering review—it usually arrives as a static document (e.g., a PDF). This format is "lossy": it discards the research, calculations, and evidence that led to the conclusions. If a recipient has a follow-up question, they must either go back to the author or use a general-purpose AI assistant.
The problem is that standard AI assistants lack the specific "scaffolding" of the original work and don't inherit the producing organization's authority boundaries. They might hallucinate answers or provide unauthorized advice. The authors argue that the unit of delivery should be more than just a document; it should be a package of "compiled knowledge" that travels with its own rules for interaction.
Methods
The authors propose the Portable Agent Format (PAF). As illustrated in Figure 1, a PAF package separates six distinct functions:
- Presentation ($P$): The human-readable deliverable.
- Compiled Context ($K$): Typed facts, rationales, and conditions.
- Evidence ($E$): Source records, quotations, and provenance.
- Governance ($G$): Rules declaring permitted retrieval and interaction.
- Manifest/Signatures ($M, \Sigma$): Cryptographic binds to ensure integrity.
- Runtime Profile ($R$): Instructions on how a consumer should open the package.
The study evaluates two types of readers:
- Hosted Lens: A governed reader using a large model (Claude 3.5 Sonnet).
- Offline Reader: A tiny, 0.8-billion-parameter local model (Qwen 3.5-0.8B) with package-specific adapters.
They compared these against several baselines, including "source chat" (standard LLM with the document) and "plain retrieval" (LLM with BM25 retrieval over raw source files).
Datasets
The authors created four "fresh" study packages on November 1, 2026, to ensure they were held-out from the development of the readers:
- Whitaker: An investment portfolio proposal.
- DFEOX: A fund analysis report.
- Apple: A financial statement analysis based on complex SEC XBRL data.
- AWS: An engineering post-mortem (RCA) of an outage.
The evaluation used a benchmark of 400 frozen questions (100 per package), split into three categories: MEMO (answerable from the report), RECORD (answerable only from the evidence), and DECLINE (out-of-scope questions that the system should refuse).
Key Insights and Ideas
- Compiled Context > Raw Retrieval: Simply providing an LLM with raw source files (RAG) is insufficient. The "compiled" nature of PAF—where facts are already typed and rationales are explicit—provides a significant performance boost, especially for facts buried in the evidence.
- Governance as a First-Class Citizen: By including interaction rules ($G$) in the package, the authors showed that models can be much more disciplined. Instead of hallucinating an answer to an out-of-scope question, the PAF readers correctly declined them.
- Small Models Punch Above Their Weight: One of the most striking findings is that a tiny 0.8B model, when equipped with a PAF package and specific adapters, can provide useful, governed access to knowledge that would normally require a much larger model.
- Retrieval Portability is Hard: The "Apple" package failure revealed that while knowledge is portable, the logic to find it might not be. The small offline model failed to disambiguate near-duplicate financial records that the hosted system handled easily.
Key Results
The hosted PAF-aware reader (C3) achieved a 96% overall success rate, significantly outperforming the plain retrieval baseline (86%) and standard source chat (48%).
On the critical "answerable" questions (MEMO + RECORD), the hosted system reached 94% success compared to 79% for plain retrieval. Perhaps more importantly, the rate of "harmful" responses (wrong or unauthorized answers) plummeted from 20% in standard chat to just 0.25% with the full PAF package.
The offline 0.8B reader (O3) also showed impressive gains, reaching 69% overall success, whereas the same model given just the source and instructions (O2) managed only 20%.
Limitations and Future Work
- Ecosystem Lock-in: The study was conducted entirely within the "AlphaBitCore (ABC)" ecosystem. The authors acknowledge that true interoperability requires public schemas and independent implementations.
- Retrieval Fragility: The total failure of the offline reader on Apple's record-level questions (0% success) highlights a need for more robust selectors for structured/numeric data.
- Security Scope: The study did not test for adversarial attacks like prompt injection or the full cryptographic chain of trust for signatures.
- Future Directions: The authors suggest developing a "conformance suite" to test if independent readers actually follow the governance policies declared in a package.
Summary updated