The New Beautiful Testing holds that craft knowledge from Beautiful Testing (O'Reilly, 2009) is encoded in LLM weights, and that skilled practitioners are what activate it. These are the terms that carry that thesis. Where a term lives in an essay, it's linked, so you can read it in use.
Core concepts
- The New Beautiful Testing
- The argument that the craft knowledge in Beautiful Testing and the discipline's other formative texts is now encoded in LLM weights, and that skilled practitioners are what activate it. The craftsman and the model are inseparable in 2026, and organizations that miss this are already beginning to drift.
You might be calling this: “Why senior QA instincts matter more than ever in the AI era” or “The realization that AI tools are only as good as the tester directing them”
- QA Activator
- The practitioner whose craft level determines how much of an AI platform's capability actually gets reached. Borrowed from neural network vocabulary, and the central original concept of the framework. You might be calling this: “The tester who actually knows how to prompt and steer the model” or “The human in the loop who turns raw AI output into real testing”
- QA Ceiling
- The maximum testing quality an organization can reach, set by the activation level of its practitioners, not by its tools. You might be calling this: “Why buying expensive AI test tools didn't magically fix our release quality” or “The tool limit set by team skill”
- Cointelligence
- Practitioner plus model as one working unit, borrowed from Ethan Mollick's Co-Intelligence. In this framework, craftsman and model are inseparable in 2026. You might be calling this: “Pair-testing with Claude or Copilot” or “Working alongside the model instead of letting it run on autopilot”
- Craft-Core
- The asymmetric human and AI collaboration specific to this framework. Distinct from the broader idea of Cointelligence. You might be calling this: “Human-led, AI-assisted QA” or “Letting AI do the heavy lifting while senior judgment steers the strategy”
- General Contractor Model
- The agentic-era practitioner role: doesn't swing every hammer, but directs the work, owns the outcome, and knows when the subcontractors are lying.
You might be calling this: “QA as the orchestrator of AI agents” or “Being the lead inspector who doesn't write every script but owns the release quality”
Failure modes
- Optical Integrity
- Governance succeeds and craft is absent. Everything looks compliant; nothing is actually sound.
You might be calling this: “The dashboard is all green, but the build is broken” or “Compliance theater”
- Validated Fragility
- A system passes every check and still fails when it matters, because it confidently did the wrong thing or was forced to pass at all costs. A judgment problem.
You might be calling this: “Tests that pass reliably right up until a real customer touches it” or “Green test suites that test the wrong thing”
- High-Speed Fragility
- AI producing large volumes of fragile, deterministic tests quickly, at scale. A volume problem, distinct from Validated Fragility. You might be calling this: “Generating hundreds of brittle tests in five minutes” or “Flaky test sprawl powered by AI”
- Throughput Paradox
- AI accelerates code production to machine speed while verification stays at human speed, so the gap widens. You might be calling this: “Devs ship PRs in minutes, but QA takes days to verify them” or “The widening PR review bottleneck”
- Verification Tax
- The Throughput Paradox expressed as a business burden: the hidden cost paid when production outruns testing. It gets paid either up front, by people who investigate, or later, compounded, by everyone downstream of a release that shouldn't have shipped.
You might be calling this: “The hidden cost of babysitting AI-generated code” or “Why testing can't keep up with AI code volume”
- Phantom Load
- The side effect of AI-generated volume becoming an unplanned load test once the upfront human judgment layer is stripped out for "efficiency." It targets judgment applied before generation; the Verification Tax targets verification applied after. You might be calling this: “Drowning your staging environment with unvetted AI test data” or “The accidental load test created by running raw generated scripts”
- Babel Paradox
- AI substitution accelerates, the apprentice pipeline collapses, craft knowledge dies, and organizations can no longer evaluate what they've lost. You might be calling this: “What happens when we lay off junior QA and realize five years later nobody knows how to test” or “The loss of institutional testing craft”
- Craft Death Spiral
- The descending loop behind the Babel Paradox: an observed exposure gap, AI redundancy washing, apprentice pipeline collapse, and finally craft knowledge death.
You might be calling this: “Cutting the QA pipeline because AI exists, then slowly losing the ability to evaluate software quality”
- Automation Cliff
- The point where a test suite becomes unmaintainable and collapses. Surfaced from QA Wolf's framing. You might be calling this: “The day test maintenance becomes a full-time job and the suite gets deleted” or “Test suite collapse”
Depth, governance and communication
- Deep Stack Dividends
- The payoff from reaching test conditions that always existed but were never reachable: multi-user chaos, data edge cases, environmental drift. These are the categories that lost the budget conversation every quarter until AI absorbed the repetitive layer.
You might be calling this: “Finally getting to test the messy edge cases we never had time for” or “Reaching technical debt and race conditions now that basic smoke tests are automated”
- High-Yield Governance
- A portfolio model for how an organization is set up around AI. AI handles the repetitive layer, practitioners become Liquid Capital that flows to the most critical needs, and the leader becomes the Architect of Solvency. Under this model, Deep Stack Dividends are the yield.
You might be calling this: “Showing leadership that QA is an investment that pays back, not a cost to cut”
- Human Fiduciary
- The practitioner as fiduciary: judgment exercised on behalf of others in situations no contract or checklist can fully cover. The tester's obligation is to observe faithfully, connect what they see to what the business depends on, and put it in front of the person who bears the consequence, without deciding on that person's behalf.
You might be calling this: “Being the voice of the end user in the release meeting” or “Telling leadership the uncomfortable truth about a release”
- Diagnosis Before Investment Rule
- RJCadvisors' working rule that no one commits budget, tools, or people until the system, the risk, and the people around it are understood. Diagnosis is engineering work, not a courtesy before the sale. A prescription without it is guesswork, and guesswork with a budget attached is still guesswork. You might be calling this: “Don't buy enterprise AI tooling until you know what's actually broken in your process” or “Understand the bug root cause before throwing budget at automation”
- Documentation Archaeology
- Extends Deep Stack Dividends. Map and document what can be determined about a weird undocumented exception, flag explicitly what remains unknown, and resolve those gaps later if a bug explains them. Framed with Chesterton's Fence: replacing legacy code without knowing why its exceptions exist leads to outages.
You might be calling this: “Chesterton's Fence for legacy code” or “Figuring out why a 10-year-old workaround exists before refactoring it with AI”
- Quality Words
- Engineers define quality as soundness, stability and predictability; the business defines it as efficiency, productivity and operational fit. The disconnect lives in the requirements. Traces to Henrik Andersson's "company quality words" (2012, with Robert Sabourin and Rebecca Sanders).
You might be calling this: “The gap between what Product calls 'done' and what Engineering calls 'working'” or “Talking past each other about what 'quality' means”
- Sherpa Metaphor
- Sherpas are not tour guides. They are the load-bearing human infrastructure of an expedition: route finding, oxygen management, weather reading, rescue, and institutional knowledge of the mountain. The puja before a climb is the moment the expedition acknowledges what the mountain actually requires. The metaphor argues against an AI fluency that mistakes touching the tool for understanding the terrain.
You might be calling this: “The difference between playing with prompt templates and understanding real system architecture”
- Third Gate
- The third question on a common AI-readiness checklist: is the data accessible without interpreting human language or resolving ambiguous input? The rule is earned, but incomplete. The sharper question is whether that interpretation can be isolated to one bounded, auditable step, with clean structured data on both sides of it. If it can, that one step is where craft and judgment still live.
You might be calling this: “Knowing exactly where AI can run autonomously vs. where a human must make the call”
Practice
- Appreciating Assets
- Coined from a guest's remark that memories are the only asset that grows more valuable the longer you hold it, then extended to skills that compound over a career. In the governance sense, practitioners treated as appreciating assets, not reducible headcount.
You might be calling this: “Domain knowledge that gets more valuable every year” or “Treating seasoned QA as institutional memory rather than headcount”
- Wave Training
- RJCadvisors' learning method, built on the idea that people learn in waves. Each wave runs three phases. Disruption: the learner meets a concept by experiencing it before being told what it means. Correction: the concept is pulled back and tied to what the learner already knows, so they don't walk away with the wrong lesson. Settling: the concept moves forward again through practice, sequenced by increasing realism, until it becomes habit. The next wave overtakes the first, pulls back, and settles in turn. The goal is absorbed craft, not dead art. The three waves have been on the RJCadvisors logo since 2010.
You might be calling this: “Learn-by-doing followed by immediate reflection and real-world drilling”
Earlier names
Older pieces sometimes use earlier names. Each one points to where the idea lives now.
- Enduring Craft
- A synonym for The New Beautiful Testing, and the term RJCadvisors used internally first.
- Long Tail 8
- See Deep Stack Dividends. In the April glossary it also named eight service categories that AI economics made viable for solo and boutique QA practices.
- Verified Fragility
- See Validated Fragility. Earlier drafts used this wording, and it appears in the April teaser.
Attribution
Cointelligence is borrowed from Ethan Mollick. Quality Words traces to Henrik Andersson. Automation Cliff surfaced from QA Wolf's framing. Wave Training draws on Robert Siegler, Stolovitch and Keeps, and Dr. Yang Jwing-Ming. The Third Gate responds to a checklist in Jordan Blake's Agentic AI for Business. Disagree with a definition, or think a term needs one? I'd like to hear it: rick@qameditations.com.