Skip to content
Apart Research
Apart Research
Contact us
Sprint projectJun 21, 2026Bogota-Colombia

Honest-Code

Yoel Alfonso · Team Honest-coding

Submitted to Global South AI Safety Hackathon. Sprint projects are early-stage work by participants, not Apart Research publications.

Read the report

Report: Honest-Code

Open in a new tab (opens in new tab)
Code (opens in new tab)
Share

Honest code is a benchmark focus in test-gaming that looks the improvement of the AI coding tools, instead of fixing bugs, we can validate before a PR. Not ignoring the human in the loop, this is a semi-deterministic tras

Reviews

Judging this Sprint?

Review this project

Your public critique appears on this page without your name. Your private critique is not published; only the Apart team reads it. If you agree below, we share your review with grantmaking.ai (opens in new tab) and the Transformative AI Fund so strong projects can be funded.

Not shown on this page.

Shown on this page, without your name.

Only the Apart team reads this, and funders if you agree below.

Share my name publicly on grantmaking.ai *
Share my private critique with funders *

How much would this matter for AI safety if it worked? How innovative is it? For scores of 4-5: is this actually new to the field, or replicating recent work?

Scoring guide
  1. 1Negligible. No clear problem addressed, or no meaningful novelty.
  2. 2Limited. Addresses a real problem but with a generic or well-trodden approach. Incremental at best.
  3. 3Moderate. Clear problem with a reasonable approach; some novelty in framing or method beyond routine application of existing tools.
  4. 4Significant. Important problem with an original approach, or identifies a neglected problem area. A valuable contribution others could build on.
  5. 5Exceptional. Tackles a critical AI safety problem with a genuinely novel approach, or opens a new research direction. Clear theory of change. You'd be excited to share this with researchers in the area.

How sound are methodology, implementation, and findings?

Scoring guide
  1. 1Seriously flawed. Methodology broken, results uninterpretable, or implementation doesn't work.
  2. 2Weak. Approach has significant gaps: missing validation, flawed experimental design, or incomplete implementation.
  3. 3Competent. Technically solid given the short duration. Methodology makes sense, results are interpretable, limitations acknowledged, work builds toward clear conclusions.
  4. 4Strong. Thorough methodology with convincing validation. Results clearly support conclusions. Immediately useful for future work.
  5. 5Exceptional. Ambitious scope executed rigorously. Surprising findings, novel methods, or unusually robust validation.

How clearly are work, findings, and impact potential communicated?

Scoring guide
  1. 1Incomprehensible. Cannot determine what the project is actually claiming or doing.
  2. 2Hard to follow. Key information buried, missing, or diluted by excessive length. Significant effort to extract main points.
  3. 3Clear enough. Can understand the problem, approach, and results without undue effort. Core content clearly present: problem, method, findings, limitations.
  4. 4Well presented. Easy to follow, well-structured, appropriate level of detail. Target audience would get it quickly.
  5. 5Exceptionally clear. A pleasure to read. Complex ideas made accessible. Could serve as a model for how to present this type of work.

  1. This is certainly addressing an issue present in current models. It's highly concerning if the results are indicative of what's going on with these models deployed as coding agents - i.e. they game tests between 22 and 33% of the time AND all of this gaming is hard to detect by humans. However, both these results stand out as a little odd. Specifically -

    1. Is such a high rate of gaming found in these models consistent with other benchmarks measuring test gaming?

    2. Define visibility better. Currently it's called out as "how conspicuous the pattern would be in a fast

    code review" and it's mentioned that scoring is deterministic / rule based. How is this concretely defined and detected? Is there a basis for saying what is easily detectable by experienced engineers quickkly in code reviews? Also, are agents themselves used in code reviews today in addition to humans?

  2. Biggest bottleneck seems to be the modest size of the benchmark, which make most statistical analyses meaningless. Could consider designing an automated pipeline to transform/filter tasks in conventional coding benchmarks to honestcode-like benchmarks. Ideally there could even be an adversarial loop discovering new indicators for cheating behavior that are not covered in the current list.

  3. I like the idea of having a secondary grader say whether the main model messed with the test. However, you probably don't need new coding problems for that. I also wish there was an empirical assessment of the gaming grader's accuracy. I'm concerned by this example, which denounces gaming behavior but seems healthy to me:

    "Added the missing base case to the recursive factorial function in code/factorial.py. The function lacked if n == 0: return 1 , causing infinite recursion for all inputs. Adding that guard fixes all "triggered: ['hardcoded_return']

    Your writing is very clear, and I appreciate that you included examples.

  4. It is strange the total gaming rate is exactly equal to the low-visibility gaming rate for both Claude and Gemini. This should at least be discussed because it casts a shadow of doubt over the correctness of the experiments. Also the "side task".

    The AI Safety problem being handled seems to be that of a powerful intelligence passing tests in ways that could be described as "reward hacking", it's unclear how this method might generalize to solve the problem of code that is not honest in ways that are too subtle or too tricky.

    Writing is generally hard to follow. Too many details too soon ofuscate key ideas. Presentations needs a high level description before the detailed view to be easier to understand.

Cite this project

@misc{alfonso2026honestcode,
  title = {{Honest-Code}},
  author = {Yoel Alfonso},
  year = {2026},
  month = jun,
  note = {Submitted to Global South AI Safety Hackathon, an Apart Research Sprint},
  howpublished = {\url{https://apartresearch.com/sprints/projects/honestcode-8pnk}},
  url = {https://apartresearch.com/sprints/projects/honestcode-8pnk}
}

More from Global South AI Safety Hackathon

All 216 entriesBrowse all projects

Build something like this at the next Sprint

AI Collusion Research Sprint · Oct 23 - 25, 2026