Skip to content

← Back to projects

LOUST

LOUST-PRO/Compliance-Verification-Harness

Period: Sep 2024 —

Active

The challenge

My security research workflow was manual — handwritten notes, copy-pasted reproducers, prose-driven reports. I could not trace which piece of evidence supported which claim in a finished advisory, and byte-exact reproducibility mattered more than throughput.

My role

I designed and shipped a multi-stage pipeline that turns raw documentation into auditable graph-backed reports, with each stage writing a deterministic artifact the next stage consumes.

What I did

  1. 01

    Pipeline pattern: docs to AST to graph to report

    Why: Each stage produces a deterministic artifact with a hash, so the audit chain between evidence and report is mechanical rather than narrative. The graph is the single source of truth.

    Trade-off: About 30% more disk use and roughly twice the runtime of a prose workflow. Every reproducer becomes traceable, which is the right tradeoff when the deliverable is a public advisory.

  2. 02

    Cross-language adapters (Python, Go, Rust)

    Why: Different stages benefit from different languages. I did not want to lose data in lossy round-trips between tools, so each language keeps its native data structures through the boundary.

    Trade-off: Three adapter surfaces to maintain instead of one. The audit chain earns the cost because the alternative was silent data loss between stages.

  3. 03

    Per-stage byte-stable output

    Why: Two runs over the same input produce identical artifacts, so regressions are easy to spot and reproducibility claims hold up to scrutiny.

    Trade-off: No lossy fast-path optimizations are possible in any stage. Speed is bounded by the slowest correct stage.

What changed

  • Reproducer traceability

    Before: Opaque — manual notes tied to memory

    After: 100% hashed and linked to the claim they support

    Evidence: Internal advisory audits, 2026-Q3

  • Report drafting time

    Before: About 4 hours per advisory

    After: About 45 minutes, graph-rendered

    Evidence: estimate, not measured

  • Audit failures (missed evidence)

    Before: About 3 per advisory

    After: 0 in 2026-Q3

Trade-offs

I gave up some of the speed of ad-hoc prose work for the audit chain. For advisory-grade output the speed cost is worth it; for throwaway research it is not.

What I learned

Byte-exact reproducibility is not a luxury — it is the only way to defend a security claim under scrutiny. Anything less than a deterministic artifact is an opinion.

Stack

  • Python
  • Go
  • Rust
  • RDF
  • AST adapters

Repository

https://github.com/LOUST-PRO/Compliance-Verification-Harness

Evidence

← Back to projects · curated 2026-09-20