Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- Mean credit and the share of fully correct answers are different measures.
- Label percentage-point differences separately from relative percentage changes.
- Repeated attempts increase result counts without adding distinct tasks.
- Record research results, account access and local workflow gains separately.
Identify the score and sample first
OpenAI’s November 6 BioBasicsBench post reports 100 routine wet-lab data interpretation tasks scored with rubrics. At xhigh, Astra earns 66.7% and GPT-5.6 Sol 43.8% mean credit. Across 37 model/effort settings, five attempts per task produce 18,500 results.
Keep tasks, settings and attempts in separate columns. Answering the same question again and encountering a new question represent different sample relationships. Reporting only the largest count can suggest that every record comes from a distinct task. Nor should data interpretation tasks become a claim of independently executing complete experiments.
Why an average is not a full-credit rate
Consider an explicitly hypothetical example: three equally weighted tasks with normalized scores of 100%, 50% and 50% average 66.7%, while only one-third receive full credit. These are not benchmark task data or results of a new model test.
Different score distributions can have the same mean. Decide whether your question concerns average performance, fully correct answers or particular critical errors. If you need a deliverable that requires no rework, define acceptance criteria for that outcome; an average score cannot replace them.
Label the 22.9-percentage-point gap
Subtracting the two reported scores gives 22.9 percentage points. Percentage points describe the absolute difference between percentages. Relative growth also divides by the starting value. Because the denominators differ, an unlabeled percentage leaves readers guessing which calculation you used.
Our suggested wording is “a 22.9-percentage-point gap in mean credit at the same xhigh setting.” This arithmetic comparison does not calculate task duration, rework costs or per-call prices. If those outcomes matter to adoption, collect separate records and leave them unknown until measured.
Add access evidence and a local acceptance check
A research score alone cannot determine what an account can use. Save the date, exact model identifier, access route and applicable conditions from a relevant product notice, then compare them with the account’s visible state. This research page supplies no account, API, regional, quota or pricing terms; it establishes neither availability nor unavailability.
For your own workflow, fix a small set of representative tasks and explicit acceptance criteria. Record successes, failures, human review and rework separately. This is editorial advice, not an experiment performed for this article: retain input and tool versions and the reasons behind each assessment so differences can later be explained.
Keep failed tasks instead of excluding them to improve the headline number. If tasks, settings or acceptance criteria change, use a separate comparison window and retain the old record with the reason for the change. The result would be an observation about your sample, not a guarantee of local gains from a published score.
Sources
Frequently Asked Questions
Do five evaluation attempts mean the product automatically retries five times?
No. The attempt count is part of the evaluation design. Product retry behavior requires the relevant client or API’s documented contract.
Should I overwrite my note if the research page later changes a number?
Keep the original citation date and version, then record the new number, reason for revision and affected conclusions. Otherwise, the evidence behind an earlier decision becomes difficult to reconstruct.