Earlier quoted context omitted.
> All flagship models are within like 1-5% of each other Don't know about that. I'm using code review of my lone lisp project as a benchmark. It's a massive parallel code review where a coordinator cuts up the codebase into sections and dispatches agents to consider each part from different perspectives like quality, maintainability, consistency, correctness, rigor, etc. Ran a complete Fable/max code review. Took ove…
Check the remainder for hallucination for sure.
My methodology consists of launching a 242 cell parallel code review matrix and committing all Fable/Sol max effort agent prompts and their full reports to a private orphan branch on the repository. This is the part that is taking me months to complete. This thing can kill my $100 subscription in about 12 hours.
When done, these raw findings will be semantically deduplicated and merged into a list of findings per model. This list will then be audited for hallucinated or otherwise made up findings. This will refine the list, and hallucination rate is its own data point. I'm also counting things like cybersecurity refusals and downgrades.
When all this is done, I'll analyse the final results and publish them on my website.
Some preliminary analysis:
Which code review lenses were the most valuable, where value is defined as number of serious issues identified? Rigor, followed by tests, robustness, correctness, and so on. I was able to create a tier list of reviewer personas using evidence! I can now run focused code reviews using the highest value lenses.
What's the most expensive code review? Correctness and rigor, of the lone lisp machine specifically.
API costs per finding? $0.91 to $3.77. API costs per serious finding? $8.94 to $28.51. All Fable.
How long did it take? 28.2 calendar days, 66.3 agent-hours.
Is it worth it to run the code review multiple times? If a review matrix's defect capture probability is 57%, then a second run captures 81% of the estimated/extrapolated defect population, a third run captures 92%, a fourth run captures 97%, and further runs yield severely diminishing returns. Probably worth it to code review important stuff three times.
What's the impact and cost of the safety classifier? Out of the 52 Fable review cells that triggered the safety classifier, 35 died without producing any output whatsoever, so 67.3% of the cells were a complete waste of tokens. 25% produced at least some output.
Does the safety classifier trigger most often on the important code that actually needs SOTA models? For the most part, yes. Fable was most often barred from reviewing the most important and complex files in the codebase, such as the virtual machine, the parser and I/O layer. These files also have the most CRITICAL+HIGH severity findings. Only a couple outliers broke this pattern.