Tribunal: When AI Decisions Need a Ledger
July 6, 2026. Tribunal turns high-stakes AI output into a reviewable decision record: blind proposals, sealed commitments, critique, vetoes, dissent, ratification, and a hash-chained ledger.
Pablo Zavala · AI Safety Evaluation · Research Engineering
Essays & Analysis · Pablo Zavala
Notes on AI governance, evaluation, economics, and institutional design by Pablo Zavala.
Published Essays
Heard.now and Checkable Civic Listening
July 6, 2026. Heard.now frames civic AI as a checkable listening loop: residents speak plainly, privacy gates protect raw input, sample limits accompany each finding, and campaigns show what changed.
JROS: Resource OS With Receipts
July 6, 2026. JROS turns job search into governed resource work: sources, claims, packets, validation, and human approval gates over application-volume automation.
A Course Book Should Show Its Work
July 6, 2026. LectureForge turns authorized teaching material into private, inspectable course-book drafts while keeping rights, privacy, review, and publication authority explicit.
AEGIS and the Market for Verified Agent Work
July 4, 2026. AEGIS frames cross-owner agent work as a contract, verifier commitment, bid record, receipt, dispute path, settlement authority packet, and execution receipt, with launch claims bound to evidence.
Trading Simulation as a Human Review Test
July 2, 2026. Safe MarketUniverses uses finance as a compact lab for a practical agent-safety question: can model-emitted uncertainty guide scarce human review toward decisions where review would have helped?
Confidence Alone Fails Oversight Triage
July 2, 2026. Why model confidence can be calibrated on average but still fail the operational question: which agent decisions deserve scarce human review.
The Policymaker's Bias: When Institutions Misjudge GLP-1 Obesity Drugs
March 24, 2026. GLP-1 drugs work while public coverage stalls, and the gap reads best as a behavioral-economics failure inside institutions: fiscal-score anchoring, an inherited status-quo default, and loss-averse delay that a multi-attribute objective and two framing interventions can correct.
A Governance Lab Should Run the Same Way Twice
March 16, 2026. Teaching AI governance through Colab notebooks where Run all works deterministically on free hardware, with fixed seeds, dated snapshots, sample caps, and provider-agnostic fallbacks that let students audit every claim, and a structured mini-brief that turns each lab into an argument from question to method to limitation to recommendation.
Temporal Admissibility: Auditable Decisions by Construction
February 15, 2026. A formal finance manuscript makes time a contract, so every decision reads only data published before it, look-ahead bias vanishes by construction, and each policy becomes a deterministic, auditable artifact, with a conceptual bridge to verifiable AI decisions.
The Fluency Trap: When AI Output Replaces Human Capital
February 10, 2026. AI helps when the system sharpens judgment. Polished output becomes a fluency trap when output replaces the verification, writing, reasoning, and cognitive habits that build human capital.
Gigabit City: Evidence, Equity, and the BEAD Playbook
July 10, 2025. A four-page evidence memo distills the empirical record on U.S. gigabit fiber, grading every claim as causal or descriptive: Chattanooga's $2.69 billion municipal payoff, a roughly 20% incumbent speed response to Google Fiber entry, 2% and nearly 9% housing premiums, speed thresholds for business births, recurring adoption gaps, and a deployment playbook BEAD should enforce.
Judgment by Algorithm: When Aggregate Fairness Hides Subgroup Harm
November 11, 2024. A graduate exercise on synthetic pretrial data shows disparate error rates across race that aggregate fairness summaries miss, a failure mode a safety review should catch and an auditable decision record would surface.
Multilingual AI for Indigenous Rights: Safety in the Low-Resource Case
October 13, 2024. A design proposal rather than a built system, for a Spanish, Quechua, and Shuar legal and business advisor, framed around low-resource-language safety, hallucination risk, and human review.