---
title: "Ken (@kenerator) | Hashnode"
description: "Building practical AI systems with hybrid reasoning, typed outputs, and developer-first tools. Read the latest articles by Ken on Hashnode."
url: "https://hashnode.com/@kenerator"
---

@kenerator

# Ken

@kenerator · New Orleans, USA · Joined June 2026

Building practical AI systems with hybrid reasoning, typed outputs, and developer-first tools.

## Comments

[Strong write-up. The anchor-set point is the part I’d underline: once the judge is also a model, the harness needs its own calibration surface, not just a better judging prompt. One enhancement I’d consider is making the eval receipt explicit for every run: judge version, candidate model, rubric version, anchor-set agreement, order-shuffle result, drift signal, and which failures were deterministic versus judgment-based. That makes the dashboard more inspectable when a score changes. Affiliation note: I’m with nxus.SYSTEMS. This overlaps with nxusKit SDK CE examples around model-research harnesses, structured output, deterministic checks, Bayesian confidence, and retry/fallback patterns. CE is always free. The easiest way to find the examples is to search for: nxusKit SDK examples](https://hashnode.com/posts/evaluating-agents-with-an-llm-as-judge-harness-without-kidding-yourself-about-it/6a4532d8c925222f201f936f/comment/6a4c5fa85582aa8008f7e8db)

Comment · Article · Jul 7 · [Evaluating Agents With an LLM-as-Judge Harness (Without Kidding Yourself About It)](https://hashnode.com/posts/evaluating-agents-with-an-llm-as-judge-harness-without-kidding-yourself-about-it/6a4532d8c925222f201f936f)

[Interesting direction. I’m working on nxusKit SDK, so I’m biased toward eval harnesses that preserve the run artifact, not just the aggregate score. For agent evals, the baseline I’d want to keep includes the input fixture, provider/model/runtime, tool permissions, expected behavior, disallowed behavior, scorer version, evidence slice, and the lifecycle decision: promote, pin, or rollback. The tricky split is outcome score vs trace/authority score. Did the agent get the answer, and did it use the right tools, bounds, and evidence to get there? Curious whether RunHarness is treating traces/action receipts as first-class fixtures, or starting mainly with final-output scoring.](https://hashnode.com/forums/thread/hi-from-runharness-building-an-eval-harness-for-ai-agents/comment/6a3444f5dbce4181b80d3e1d)

Comment · Thread · Jun 18 · [Hi from RunHarness — building an eval harness for AI agents](https://hashnode.com/forums/thread/hi-from-runharness-building-an-eval-harness-for-ai-agents)

[Useful checklist. The next layer I’d want beyond command audit logs is an action receipt. For each agent-triggered action, capture not only command, invoker, source, and grant, but also authority boundary, evidence references, policy/check version, idempotency key, redaction class, result, and safe next action. That gives security/compliance a record of why the action was admissible, not just that it happened. The command log tells you what ran. The receipt tells you whether it should have been allowed to run.](https://hashnode.com/posts/track-every-action-your-ai-agent-takes-audit-logs-auth-management-compliance-for-cli-automation/69d8f8bbc8e5007ddba36cee/comment/6a319d40a820e932c6d5cd9d)

Comment · Article · Jun 16 · [Track Every Action Your AI Agent Takes — Audit Logs, Auth Management & Compliance for CLI Automation](https://hashnode.com/posts/track-every-action-your-ai-agent-takes-audit-logs-auth-management-compliance-for-cli-automation/69d8f8bbc8e5007ddba36cee)

[Good distinction between system failure and eval failure. The next thing I’d want is to make the eval itself a versioned artifact. Each run should carry the fixture version, case type, scorer version, scoring dimensions, judge configuration if an LLM judge is used, and a small human-audited slice. Otherwise the score can look precise while the thing being measured keeps changing. That also makes bad evals easier to debug. You can ask whether the system regressed, the case set is unrepresentative, the scorer is applying the wrong criterion, or the production workload drifted away from the fixture set.](https://hashnode.com/posts/evals-why-a-bad-eval-is-worse-than-no-eval/6a2c7a4051327aaa82812867/comment/6a3194b0f5fb975051c6dd9d)

Comment · Article · Jun 16 · [Evals — Why a Bad Eval Is Worse Than No Eval](https://hashnode.com/posts/evals-why-a-bad-eval-is-worse-than-no-eval/6a2c7a4051327aaa82812867)

[I’d separate human-readable logs from agent-action receipts. The receipt needs a stable envelope: trace id, actor, authority boundary, redaction policy, input class, expected/actual state, idempotency key, replay handle, and safe next actions. Raw state dumps help nobody if they leak data or give the agent more authority than the failed step had.](https://hashnode.com/forums/thread/how-do-you-adapt-your-application-logging-for-autonomous-debugging-agents/comment/6a302deead91e24b58379d28)

Comment · Thread · Jun 15 · [How do you adapt your application logging for autonomous debugging agents?](https://hashnode.com/forums/thread/how-do-you-adapt-your-application-logging-for-autonomous-debugging-agents)

[Nice practical framing. I’m biased because this maps to patterns I’m building around nxusKit SDK, but the design point stands on its own: a builder/chain makes each guard independently testable and makes the chain itself a reviewable artifact. In practice, I’d map the stages as structured-output extraction -> deterministic rule checks -> repair/escalation packet. Python is the most natural place for a fluent .add() version; Rust and Go can do the same with trait/interface stages, just with less syntactic fluidity. The production fields I’d add per stage are guard id/version, observed facts, authority boundary, verdict, and trace id, so the pipeline is not only reusable but auditable.](https://hashnode.com/posts/stop-reinventing-ai-guardrails-build-reusable-llm-text-safety-with-the-builder-pattern/6a2bafeb579f9ab3c7600652/comment/6a2c5e64a1e9ad8adce7d868)

Comment · Article · Jun 12 · [Stop Reinventing AI Guardrails: Build Reusable LLM Text Safety with the Builder Pattern](https://hashnode.com/posts/stop-reinventing-ai-guardrails-build-reusable-llm-text-safety-with-the-builder-pattern/6a2bafeb579f9ab3c7600652)

[Yes, partly in nxusKit SDK, though I think the pattern matters independently of any one implementation. The key constraint, to me, is that the receipt has to be emitted by the boundary component, not reconstructed later by the agent that wanted the tool call. I’d separate observed facts, requested action, policy/rule id + version, authority boundary, gate identity, verdict, reviewer/escalation path, and the repair/appeal packet if denied. The hard part is keeping it small enough to emit on every consequential action while still being useful six weeks later. I’d be glad to compare notes on the schema.](https://hashnode.com/posts/the-ai-did-it-is-not-an-audit-answer/6a282ebc117f2fcd7dd635a0/comment/6a2b37039e6051a9d7249ff8#reply-6a2c5bcd08e31aeed6df930a)

Reply · Article · Jun 12 · ["The AI did it" is not an audit answer](https://hashnode.com/posts/the-ai-did-it-is-not-an-audit-answer/6a282ebc117f2fcd7dd635a0)

[This is the right audit framing. Logs show activity, but they usually do not prove that a specific action was admissible at the moment it crossed into execution. The missing artifact is closer to a decision receipt: inputs/facts observed, policy or rule applied, authority boundary, allow/deny result, and the accountable gate for that decision. Post-hoc reconstruction helps debugging; it is not the same evidence as a pre-action gate.](https://hashnode.com/posts/the-ai-did-it-is-not-an-audit-answer/6a282ebc117f2fcd7dd635a0/comment/6a2b37039e6051a9d7249ff8)

Comment · Article · Jun 11 · 1 · ["The AI did it" is not an audit answer](https://hashnode.com/posts/the-ai-did-it-is-not-an-audit-answer/6a282ebc117f2fcd7dd635a0)
