Notes from the eval harness
Engineering and evaluation science for teams shipping agents into clinics, banks, law firms and claims desks.
Written by the people who build the judges, the datasets and the gates.
- Evaluation science · 8 minpass@k vs pass^k: measuring reliability, not luckpass@k rewards an agent that succeeds once in k tries. pass^k asks whether it succeeds every time. Why regulated agents should be gated on the second, how to estimate it, and what the numbers look like.Karthik Reddy · 4 Nov 2026Evaluation science
- Legal · 8 minTesting for citation hallucination in legal AIFabricated cases are the visible failure. Real cases cited for the wrong proposition are the common one. A test design for legal research and drafting agents that checks existence, grounding, quotation and validity separately.Nandini Joshi · 6 Oct 2026Legal
- Regulation · 9 minEU AI Act Annex IV for engineersAnnex IV lists what the technical documentation of a high-risk AI system must contain. A section-by-section reading for the engineers who will produce most of the evidence, updated for the Digital Omnibus timeline as of November 2026.Nandini Joshi · 15 Sept 2026Regulation
- Engineering · 8 minFast path vs full suite: structuring CI evalsA gate that takes 40 minutes gets disabled. A gate that checks five cases catches nothing. How we split evaluation into a fast path on every push and a full suite on merge and nightly, with real budgets for time, runs and flakiness.Vikram Choudhary · 11 Aug 2026Engineering
- Security · 8 minPrompt injection through retrieved documentsThe most dangerous instructions an agent reads are not typed by the user. How indirect prompt injection reaches regulated agents through faxes, PDFs and knowledge bases, and how to test for it before release.Ishita Banerjee · 21 Jul 2026Security
- Healthcare · 9 minBuilding golden datasets with cliniciansHow we turn de-identified clinical patterns into golden cases that two physicians label independently, how disagreements get adjudicated, and what a rubric has to say before clinicians can agree on it.Dr. Meera Iyer · 30 Jun 2026Healthcare
- Financial services · 8 minCFPB Circular 2023-03 for engineers: what still applies after the withdrawalThe CFPB withdrew Circulars 2022-03 and 2023-03 in May 2025, but Regulation B's specific-reasons rule did not move. How to test that every reason an AI credit agent states traces back to the decision record.Nandini Joshi · 9 Jun 2026Financial services
- Engineering · 8 minTesting tool-call schemas for agentsStructured outputs make tool calls parse. They do not make them correct. A layered approach to testing tool arguments: schema conformance, semantic boundaries, cross-field rules, units and provenance.Rahul Venkataraman · 19 May 2026Engineering
- Engineering · 8 minSilent prompt regressions in multi-turn agentsA one-line prompt fix for turn one can quietly change what the agent does on turn five. Why multi-turn regressions escape review, how to replay conversations against a baseline, and what it costs when you do not.Vikram Choudhary · 28 Apr 2026Engineering
- Evaluation science · 9 minCalibrating LLM judges with Cohen's kappaRaw agreement flatters a judge when most cases pass. How to measure an LLM judge against expert annotators with Cohen's kappa, when to switch to Gwet's AC1, and how to control position, verbosity and self-preference bias.Karthik Reddy · 7 Apr 2026Evaluation science
- Engineering · 8 minTrajectory assertions vs final-answer evalsTwo agents can return the same correct answer, and only one of them followed the process your auditor will ask about. How to assert on tool order, parameters, provenance and loops, and when a final-answer check is still enough.Rahul Venkataraman · 17 Mar 2026Engineering
- Evaluation science · 7 minWhy MMLU scores don't protect regulated agentsA model that scores well on MMLU or GSM8K can still call the wrong tool, skip an identity check or invent a dosage. What public benchmarks measure, what they cannot see, and what a regulated team should test instead.Shiv Shankar · 3 Mar 2026Evaluation science
Who writes here
- SSShiv ShankarFounder & CEO1 posts
- RVRahul VenkataramanCTO2 posts
- MIDr. Meera IyerHead of Clinical Evaluation1 posts
- KRKarthik ReddyHead of Judge Calibration Research2 posts
- NJNandini JoshiHead of Regulatory Science3 posts
- VCVikram ChoudharyVP Engineering, Eval Infrastructure2 posts
- IBIshita BanerjeeHead of Security & Compliance1 posts
Explore the 2026 Vertical AI Reliability Index: failure rates across 12 regulated domainsSchema adherence, trajectory completion, domain factuality and adversarial robustness for 48 models. Free to cite.View the index
Monthly newsletter
The Regulated Agent
A monthly deep dive on AI evaluations, safety benchmarks and regulatory enforcement. One email, first Tuesday of the month.
Stop testing regulated agents on generic prompts.
Run your agent against 100 industry-standard clinical or financial edge cases in 5 minutes. 14-day free production trial, no card.