Sonnet 5.5 vs Opus 5.5: Industrial Answers and Output Format
On 250 shared industrial questions, Opus earned 66% strict accuracy and Sonnet 50%. Output-format failures affected both models.
Articles
Muse Glimmer vs Microsoft-Decision-1: Local AI Within Four Answers
On 487 industrial questions, local Muse Glimmer finished three answers behind cloud-hosted Decision-1, and four after reversing the options.
Microsoft-Decision-1 on Industrial Questions: Consistent Answers, Persistent Errors
Microsoft-Decision-1 matched 93.84% of 487 industrial benchmark answers. Five reversed-option runs revealed stable choices and persistent errors.
Gemini 4 Argon on IBM FailureSensorIQ: Testing at Provider Defaults
Gemini 4 Argon scored 42.94% on 170 shared questions. A four-model comparison examined strict answer accuracy at provider defaults.
Can consistent AI classifications be trusted?
Across five new runs per model, cross-level agreement marked more accurate subsets—but consistent mistakes remained common.
Finding out which part broke
On simulated machine failures, a trained classifier led the comparison; a simple rule also had a higher score than the tested language models.
Can small AI models diagnose machine faults?
Four small local models, an industrial question-answering benchmark, and the difference between correct answers and usable output.
Adaptive smoothing in supply chain replenishment
AI selected a smoothing parameter while a formula calculated orders. Every tested condition dampened variance; superiority over a fixed setting was not established.
Testing the limits of the ordering formula
Fourteen variants, including perfect labels, tested whether better classification could improve the same ordering architecture.
Better classifications, similar ordering outcomes
Four models interpreted seasonal and disruption signals, but their ordering results clustered near the formula-only baseline.
Correctly formatted labels, wider order swings
A five-label interface produced valid answers on every tested call. It did not make the simulated supply chain more stable.
Letting AI adjust the safety-stock buffer
A formula handled the arithmetic while AI chose the buffer multiplier. All tested model conditions increased variance over a fixed multiplier.
Comparing sarvam-30b with GPT OSS 120B
Testing seasonal ordering behaviour on a synthetic Indian automotive demand series—and separating model results from serving reliability.
Language models against simple ordering rules
Every tested rule-based baseline produced both lower order variance and fewer stockouts than every language-model condition.
Does context help AI place supply chain orders?
An exploratory comparison found that business context helped one model and hurt another. All four configurations amplified order variance.