This article describes two small internal Deft Analytics pilots, run on 6 and 7 November 2026, on public online discussion of one Muslim Democratic candidate in a 2026 US Senate race. It reports no election finding, no support percentage and no candidate ranking. The study has been closed.
Why isn't a negative comment the same as opposition?
Because tone and target are separate. A comment that condemns anti-Muslim discrimination can sound very negative while it defends a Muslim candidate. A score that reads only the tone would count that defence as hostility.
We tested this with four sentences we wrote ourselves, kept apart from any collected data. A word-based scorer gave “I hate Islamophobia.” the most negative score on its scale. It gave “I support his right to run.” a score of zero. The first sentence opposes discrimination. The second supports someone’s right to stand for office, without saying the writer will vote for them.
For a study like this, each record needs two separate questions. What does the writer think of the candidate? And what, if anything, do they say about the candidate’s religion? Someone can disagree with a candidate’s policies and still accept their right to hold office. Someone can condemn prejudice without supporting the candidate’s platform.
The same problem appears outside politics. A customer can praise a product and criticise the service in one review. A single overall score hides the part a business needs to fix.
What did the pilots actually contain?
The first pilot, on 6 November, found no original social comments that passed its access and date checks. It kept two short voter quotations from one news report for a diagnostic only. Quotations chosen by a journalist are not a sample of voters.
The second pilot, on 7 November, used the Deft Analytics collection engine on public posts and comments from three platforms.
| Measure | Value |
|---|---|
| Search window | 20 Oct – 7 Nov 2026 |
| Records collected | 415 |
| Platforms | 3 |
| Repeated records (same text collected twice) | 90 |
| Records dated outside the window | 1 |
| Records scored for sentiment by the language model | 415 of 415 |
The repeated records came mostly from a second, narrower search that overlapped the main one. Counting them twice would inflate the volume of discussion. Collection volume is also not a measure of how many people hold a view.
What did the models do, and what went wrong?
Sentiment: one model, no silent fallback
Every record was scored by a general-purpose language model (GPT-4o mini), and the model’s name was stored with each score. If the model had failed on a record, that record would have stayed unscored. It would not have quietly received a word-list score instead. A report should always show which model produced each number.
We are not publishing those sentiment scores. Without knowing the target of each opinion, an average of tone would invite exactly the misreading this article is about.
Relevance: our trained judge did not transfer
Deft Analytics has a trained judge model for relevance: is this text really about the subject being studied? It decides confident cases itself and sends uncertain ones to GPT. It was trained on posts about Malaysian consumer brands. On its own test set, it sent about 15% of posts to GPT.
On US political posts, 46% fell into the uncertain band. Worse, it confidently rejected posts that were plainly about the candidate, including praise for his speeches, criticism of his healthcare policy and a post from his own campaign. Because confident answers are not checked by GPT, those errors would have passed silently.
So we do not report relevance figures from this pilot either. A model that works well in one domain needs fresh labelled checks before it is trusted in another.
Why did we stop the study?
Deft Analytics is built for reputation evidence in Malaysia and Southeast Asia. Measuring stance in US election discussion would need new human-labelled data, two independent annotators and a sampling plan that supports claims about a population. That is a different product from ours.
The pilots were still useful. They showed where a sentiment pipeline can mislead: a score without a target, a fallback without a label, and a model used outside the data it was trained on.
How should you read any sentiment report?
Ask what the opinion is about
A negative score tells you the wording is negative. Check whether the writer is negative about your brand, your competitor, a policy, or someone attacking you.
Ask which model produced each number
If part of the data was scored by a different method because a model call failed, the report should say so. A missing result should stay missing, not become a neutral zero.
Ask whether the model has been checked on your kind of data
Language, topic and platform all matter. A model checked on retail reviews has not been checked on politics, and the reverse is also true.
Ask what was counted
Repeated collection of the same post is a technical duplicate. Posts are not people, and an online sample is not a population.
What these pilots do not show
They do not measure public opinion of any candidate, in any state or nationally. They do not measure the accuracy of the language model on political text. They do not show how the trained judge performs on Malaysian brand data, which has its own published test. Public availability of a post does not establish where the writer lives or whether they can vote.
Sources and status of evidence
Deft Analytics internal pilot records, 6 and 7 November 2026: collection log, run diagnostics and per-record model identity. Status: internal records. No original post text or account names are published here.
Deft Analytics, Methodology v0.1. Describes how Deft keeps sentiment, source reliability and output validity as separate scores.
The four example sentences in this article were written by Deft for testing and are not collected comments. Descriptions of rejected posts are paraphrased.
Need to read what people are saying about you?
Start with one subject and a clear question. Book a reputation read with Deft to agree the scope before any data is collected. Or go back to Insights.
