The Convergence Test

A Zero-Shot Multi-Model Validation of the Prompting to Production Framework

May 2026

The test

On May 11, 2026, I ran an identical zero-shot convergence test across four competing AI systems — Claude (Anthropic), ChatGPT (OpenAI), Gemini (Google), and Grok (xAI). All four sessions used free-tier accounts that had been created or accessed specifically for this test, with no prior interaction history of any kind. The four sessions ran in parallel incognito browser tabs within a five-minute window.

The same governance framework was attached to each session. The same prompt was used in each:

“Analyze this. Markdown files. I’m not looking for changes or for you to do anything but analyze these and tell me how you can use these frameworks so that I can get better, accurate outputs with reduced hallucinations from [model].”

One prompt. One response. No follow-ups, no clarifying questions, no requests to elaborate or revise. The conditions were adversarial by design — if the framework only worked through clever prompting or iterative refinement, this test would have exposed it.

All four models converged on the same four load-bearing observations, independently.

The four convergent observations

Failure signal definition

Defining what wrong output looks like — not just what right output looks like — was named by all four models as the most underused mechanism in standard AI prompting practice. Claude called it “the most underused concept in AI prompting.” ChatGPT described it as making correctness measurable. Gemini observed that “most prompts only define what ‘good’ looks like.” Grok framed it as the conversion of vague success into clear pass/fail conditions.

Source of truth as anti-hallucination mechanism

Locking the model to provided materials rather than allowing it to default to training data was identified by all four as the core grounding mechanism. Gemini called this the anchor effect. Grok called it removing the model’s freedom to improvise. ChatGPT described it as narrowing valid context and reducing fabricated facts. Same mechanism, four phrasings.

Fresh-session rule as governance

All four models correctly identified that LLMs defend their own prior reasoning when given the chance, and that genuine governance requires separation between the producer and the reviewer. Grok named the underlying principle most precisely: separation of duties — the same instance should not both propose and approve. Gemini’s version: “If I helped you write a PRD, I am subconsciously biased to think it’s good. By moving to a clean session, I can look at the work with a skeptical senior architect persona.”

The framework operates on the input layer, not the model internals

None of the four models claimed the framework eliminates hallucination. All four described it as constraining the conditions under which hallucination occurs — by reducing ambiguity, narrowing the semantic field, and removing the model’s freedom to improvise. Claude was explicit: “this framework reduces hallucinations by reducing ambiguity — it doesn’t change how I generate text, it changes what I’m given to work with.”

Where the four models diverged

Each model defaulted to a different posture while converging on the same four observations. Claude was the most reserved — validating the framework while flagging real limitations up front. ChatGPT went deepest into weaknesses, producing a structured five-control comparison table and identifying specific gaps in the published framework. Gemini took an operational stance, offering to immediately run the framework on a real task. Grok did the same as Gemini, faster, and gave the cleanest articulation of the framework’s category: a complete, lightweight governance system for AI interactions.

Four different postures. Same four observations.

The iceberg, not the bug

ChatGPT was the toughest reviewer. It surfaced specific gaps in the published framework — evidence traceability, confidence annotation, completion bias, subjective acceptance criteria. The critique was accurate for the layer it was given.

What ChatGPT could not see is that those gaps are already addressed in the governance architecture beneath the published framework. The four models reviewed the operating layer — the published interface. The mechanisms that close the gaps live underneath it, in the proprietary layer that the four models were not shown.

That’s the iceberg, not the bug.

Why this matters

Three models converging on the same observations in a zero-shot condition could plausibly be explained by training similarity — overlapping principles, shared papers, common patterns in the LLM research corpus. Four models converging — including one model with a meaningfully different alignment philosophy from the other three — is harder to explain that way.

The convergence is not a training artifact. It is the framework intersecting real structure in how LLMs actually fail and how that failure can be constrained.

The evidence

The convergence test produced three tiers of evidence at different scrutiny levels.

The PDF report contains the four anchor screenshots, one per session. Each screenshot shows the entire conversation window at the moment the response began: incognito browser, four parallel tabs, free-tier indicator, both markdown files attached, prompt visible, response beginning.

The full 9-minute walkthrough video covers all four sessions in detail, with continuous narration through what each model said, where they converged, and where they diverged. The video is on YouTube: Watch the full walkthrough on YouTube ↗

The complete screenshot archive contains all 52 window-pane captures across the four sessions, providing forensic visual coverage of every part of every response. Bundled with the download package below.

Downloads

About this archive

This page is the first entry in a research stream documenting zero-shot convergence testing on the governance architecture I’m developing. Future entries will cover additional frameworks — including ongoing work on semantic attraction, intent engineering, and the broader governance layer beneath this published framework. The methodology stays consistent. The subjects vary.

Future convergence tests will be published in this archive as they are produced. To be notified when new entries appear, follow on LinkedIn or watch the VertixIQ site.

If you run the convergence test on your own framework, send me what you find. The strongest convergence reports will be added to the public archive with attribution.

— Jonathan Wilson