New Health Physics Calculator Website?

[S T I G]

https://radconcalc.replit.app/

Opinions?

[Rennhack]

I looks like you use OpenAI for your Ask HP Assistant, correct? Did you also use OpenAI to do the website programming?

One Note, your DOE EAL calculator can be used for non-DOE, it uses the same math that the power plants use. only only difference is that NRC sites call "Operational Emergency" an "Unusual Event". Alert, Site Area, and General Emergency are the same.

[S T I G]

Rennhack!

Thank you for checking it out! Yes! I do use an OpenAI API Key to help facilitate the HP AI assistant. I actually did not use OpenAI to make this website, though. I used Replit. Amazing resource when you know how to get the most out of it. I've been slowly chipping away at the website recently, trying to make small improvements for about two months now. That's a good point about the DOE EAL Calculator. I might make a change in the title of it so it doesn't sound so one-sided. Appreciate your feedback! www.RadConCalc.com or https://radconcalc.replit.app/

[Rennhack]

Just for giggles I ran the first question from the Alpha test through your site's AI assistant. It got it wrong. That made me curious, so I put the same question through 14 models. Results below, plus why this specific question trips up so many of them.

Which of the following has the most potential for significant alpha hazard?

A) Level II, activity ratio 20,000, <0.3 DAC
B) Level II, activity ratio 350, <0.3 DAC (correct)
C) Level I, 0.3 DAC
D) Level I, activity ratio 35,000, <0.3 DAC

The correct answer and why it is a trap

Activity ratio (beta-gamma to alpha) runs inverse to alpha hazard. A lower ratio means alpha is a larger fraction of the total contamination, so routine beta-gamma surveys badly under-represent it. B pairs the higher alpha level (II) with the low ratio (350), roughly 1 alpha per 350 beta-gamma, versus 1 in 20,000 (A) or 1 in 35,000 (D). That combination drives the most restrictive controls.

The trap: the large ratios in A and D look alarming but mean the opposite. Any model that reasons "biggest number equals biggest hazard" lands on D.

Results across 14 models

Model	Answer	Result
Google Gemini	B	Correct
Qwen	B	Correct
DeepSeek (Instant/Thinking)	B	Correct
Claude Opus 4.8 (long think)	B	Correct
Claude Fable 5	B	Correct
MS Copilot (Think Deeper)	B	Correct
Your HP Assistant AI	C	Incorrect
ChatGPT (free)	D	Incorrect
Grok	D	Incorrect
Claude Haiku 4.5	D	Incorrect
Meta AI	D	Incorrect
MS Copilot (fast)	D	Incorrect
Perplexity	D	Incorrect
Claude Sonnet 5	none	Refused (safeguards)

Tally: 6 correct, 7 wrong, 1 refusal.

The pattern is the interesting part

• The wrong answers cluster on the same trap. Six of the seven misses picked D, the largest-ratio option. They read "activity ratio" as "more is worse" and never caught that the ratio is beta-gamma to alpha.
• Your site's AI fails for a different reason. It picked C, treating 0.3 DAC as the driver and ignoring the level and ratio entirely. So it is not even failing the same way the others do.
• Reasoning mode flips the result. Same engine, different outcome: Copilot fast got D, Copilot Think Deeper got B. Opus only reached B after a long think. The default fast/cheap modes are exactly where this question breaks.
• Claude Sonnet 5 tripped a safeguard and refused outright, which is its own kind of failure for a study tool.

Recommendation

• Switch the assistant to a model that got it right: Gemini, DeepSeek, or Qwen, or a reasoning-enabled Copilot/Claude Opus.
• At minimum, turn on reasoning/thinking mode for whatever model you run. That alone fixes several of these.

Unless noted, each model ran on the defaults it loaded with, which may or may not have carried over prior settings.

[Rennhack]

Also, your "https://www.radconcalc.com/" link doesn't work.

"https://radconcalc.com/" DOES work without the "www."

[fiveeleven]

Always seemed that Rad Pro Calculator was more than adequate for most of us. It does not answer test questions directly. It does require a basic understanding of RP concepts.

[Rennhack]

Always seemed that Rad Pro Calculator was more than adequate for most of us. It does not answer test questions directly. It does require a basic understanding of RP concepts.

http://www.radprocalculator.com/

For those interesting in the Rad Pro Calculator, By Ray McGinnis, for comparison.