Bias & Accountability
Fairness audits, hidden assumptions, dataset gaps, scoring systems, and whether bias is being reduced or simply renamed.
Regular notes on AI governance, agent security, and bias — backed by published reports, reproducible benchmarks, and a companion stream of original research papers.
Every issue is built around something verifiable — a fresh benchmark run, a public report, a reproducible experiment, or a working paper queued for submission. We cite primary sources, link to the data, and publish methodology alongside results. The goal is signal that holds up when the corporate press releases age out.
Four standing beats. Each issue picks one and reports against current evidence.
Fairness audits, hidden assumptions, dataset gaps, scoring systems, and whether bias is being reduced or simply renamed.
What changes when AI agents access tools, accounts, files, browsers, infrastructure, and private workflows.
Verifiable decision trails, audit layers, rollback records, signed event tokens, and policy-aware reasoning.
Reproducible benchmark runs with the data + methodology published alongside the results. No "trust us, it scored 9.4."
Each issue links to its underlying report, benchmark dataset, or working paper.
Opening report — measuring trust, risk, and accountability in production AI systems.
Benchmark — what changes when AI can read, write, move, trigger, and decide across connected tools.
Field report — bias emerging from data, incentives, policies, and feedback loops (not just the model).
Original work supporting the newsletter — published, in review, or queued for submission. Citation-ready BibTeX is included alongside each PDF.
Per-layer HMAC + coordinator signature + outer signature: composition integrity for multimodal AI events. PDF + reproducible eval pending.
Trust-level isolation, intent-gate routing, and per-container constitutional rules. Submitted; results dataset open-source.
Deterministic post-scan for fabricated customer counts, success metrics, and industry-position claims in agent-generated text. Eval set in progress.
Each issue lands with its underlying report, benchmark data, or paper PDF attached — and subscribers see new working papers before they're submitted publicly. No tracking pixels, no sponsor placements.