Skip to content
@Black-Box-Research-Labs

Black-Box-Research-Labs

Forensic architecture audit and technical due diligence for AI-augmented infrastructure.

Black Box Research Labs

Forensic architecture audit and technical due diligence for AI-augmented infrastructure.

We produce Artifacts of Understanding — execution evidence, SHA-pinned referential evidence, and cognitive evidence — so organizations can prove their AI-generated systems do exactly what they intended, and nothing else. Every finding ships with commit-level proof you can check yourself.

Research

Neural Affinity Framework — Ingram & Merritt (2025). Empirical evidence for the Compositional Gap in Transformer architectures: 69.5% of fine-tuned tasks achieve >80% cell accuracy but fail to compose solutions globally. Validated against 302 specialist models. Predicts ARC-AGI-2 generalization failures.

Thesis: LOCAL ≠ GLOBAL — components pass review; integration breaks in production. Black Box audits the gap.

Methodology

AIV Protocol — A shift-left verification framework that generates immutable, auditable proof that AI-assisted code changes were understood, tested, and verified before deployment — not just a green CI badge.

Internal Remediation Audit (BB-SELF-2026-001) — 8 findings across 5 audit surfaces, all remediated with commit-level evidence. This is what our methodology looks like applied to ourselves.

Flashcore — Reference implementation (production SRS engine). AIV Protocol enforced in CI: packet validation gate, immutable evidence links, artifact-backed verification on every PR.

Case Studies

Verification Friction in a Series A Agent Company — An approved pull request accrued drift, automation failed to resolve conflicts, and the work was closed unmerged after 34 days. 1,007 lines of reviewed code converted to unshipped inventory. Industry pathology: verification load exceeding available bandwidth.

Oklahoma Blue Thumb Data Validation — Forensic audit of chloride measurement accuracy across citizen science vs. EPA professional monitoring. N=25 paired sites, GLMM, β=−0.433 (p=0.047). Presented at OCLWA 2026.

Engage

blackboxresearchlabs.com · miguel.ingram.research@gmail.com

Popular repositories Loading

  1. ok-blue-thumb-data-validation ok-blue-thumb-data-validation Public

    Forensic chloride data validation — Oklahoma Blue Thumb citizen science vs. EPA professional monitoring. Case study by Black Box Research Labs LLC.

    Python 1

  2. aiv-protocol aiv-protocol Public

    AIV Protocol specification (canonical)

    Python

  3. .github .github Public

    Black Box Research Labs org profile

Repositories

Showing 3 of 3 repositories

People

This organization has no public members. You must be a member to see who’s a part of this organization.

Top languages

Loading…

Most used topics

Loading…