Post
108
I published a closed-form no-go theorem in article 3, proved after the same reader broke my paper twice in a week. I thought that closed it.
It didn't. Forty-two more rounds followed with the same reader — @dipankarsarkar .
He found that a column I'd been stratifying contrasts by wasn't a factor at all — it was the training objective itself, rounded. corr = 0.97. Four rounds of p-values on a column disqualified before any data existed.
Then the actual point of the project: reinforce() had never once been run from a random start. Every result was exact gradient ascent — no sampling, ever. Run for real: exact ascent traps at a critical point 95% of the time. The sampled gradient of the same objective walks out of it 92% of the time. The "5% ceiling" was a fact about my optimizer, not the objective.
The real lesson: check what a statistic is a function of before you contrast it — the column can disqualify itself for free. And a result that confirms what you already believe gets checked less, not more.
🔗 https://huggingface.co/blog/RDTvlokip/forty-two-more-rounds-with-the-same-reader-he-foun
💻 https://github.com/RDTvlokip/RDTRL
📦 https://doi.org/10.5281/zenodo.21726216
It didn't. Forty-two more rounds followed with the same reader — @dipankarsarkar .
He found that a column I'd been stratifying contrasts by wasn't a factor at all — it was the training objective itself, rounded. corr = 0.97. Four rounds of p-values on a column disqualified before any data existed.
Then the actual point of the project: reinforce() had never once been run from a random start. Every result was exact gradient ascent — no sampling, ever. Run for real: exact ascent traps at a critical point 95% of the time. The sampled gradient of the same objective walks out of it 92% of the time. The "5% ceiling" was a fact about my optimizer, not the objective.
The real lesson: check what a statistic is a function of before you contrast it — the column can disqualify itself for free. And a result that confirms what you already believe gets checked less, not more.
🔗 https://huggingface.co/blog/RDTvlokip/forty-two-more-rounds-with-the-same-reader-he-foun
💻 https://github.com/RDTvlokip/RDTRL
📦 https://doi.org/10.5281/zenodo.21726216