You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$9,900 left to reach the funding goal.
I want to spend three months building the skills and a small public work sample needed to contribute to AI safety evaluation. The output will be a rerunnable evaluation package for one small open model, together with an explanation of the inputs, scoring rules and conditions behind its results.
I propose to run the project from 11 January to 10 April 2027, with the dates adjusted if the funding decision or payment arrives later.
The pilot will use one public evaluation with a documented method and data that can be shared lawfully. I will preserve the model revision, prompts, dataset revision, scoring code, environment and random seeds. After reproducing a baseline, I will vary one evaluation choice at a time and record whether the reported result changes. Failed runs and unresolved discrepancies will remain visible in the package. I will choose the model and task after checking the license, compute requirements and whether I can explain the full evaluation procedure.
Weeks 1–4: Continue Python foundations, complete small data-processing exercises, inspect candidate evaluation licenses and compute needs, and freeze one feasible task and its run specification.
Weeks 5–6: Execute and document a baseline, preserving input versions, commands and outputs.
Weeks 7–9: Vary one evaluation choice at a time, check scoring, and document which comparisons change the result.
Weeks 10–12: Replay from a fresh environment, document remaining discrepancies, request feedback, and release the package and findings.
Reliable measurements matter when people use evaluations to decide which AI systems need further testing or stronger safeguards. This pilot addresses one part of that process: making it possible to inspect how a reported result was obtained and whether it survives a fresh run. I will explain the limits of the selected measurement alongside the findings, so a narrow evaluation is not presented as a general safety judgment.
By the end, I intend to release a package containing the frozen inputs or lawful retrieval instructions, model and data version identifiers, scoring code, commands and environment information, and the resulting comparison table. The findings note will give the metric and denominator, explain any failed runs, and separate an evaluation-setting effect from a claim about the model's behavior.
I will test the written instructions by rerunning them from a fresh environment. I will seek external feedback, while keeping completion independent of whether a reviewer responds. A fresh self-run will be recorded as a self-run, with any external replay recorded separately.
I will also record which parts of the workflow I can write and explain independently. That record will help determine whether I should continue into evaluation work, deepen my programming foundations, or pursue a more operations-focused contribution to the field.
The minimum request is $9,000 for a three-month base budget; the $9,900 goal adds a $900 contingency reserve. These are forward-looking budget estimates rather than invoices or documented past expenses. The USD request uses a planning conversion of KRW 1,500 per dollar, rather than claiming a current exchange-rate quote.
Time and living support: $7,000
Compute and storage allowance: $1,000
Learning resources: $200
Technical mentoring or review allowance: $800
Base total / minimum funding: $9,000
Contingency reserve: $900
Goal: $9,900
No mentor or reviewer has been booked. I will avoid buying resources already available to me without added cost. Funding would let me set aside other paid work and concentrate on this transition full-time. I will reconcile overlapping support before accepting funding so the same costs are not funded twice.
I am the applicant and would run this pilot myself. My existing work concerns whether a result can be checked against the process that produced it. I am Reproducibility & Evaluation Maintainer for Throughline, an open-source research-software project, with responsibility for replay and evaluation acceptance criteria. My replay-receipt contribution was merged as PR #52.
For TrialTrace, I designed the record-comparison framework and directed AI-assisted implementation of a local research-record history explorer. I am now learning to write and understand the underlying code directly, beginning with CS50P. This proposal brings those two parts together: experience defining evidence boundaries and a structured period of hands-on technical learning.
The main execution risk is that I do not develop the necessary programming competence quickly enough to complete a model evaluation within the proposed period. At the end of Week 4, I will check whether I can write and explain the basic file-processing, function and test steps required for the selected evaluation. If that gate is not met, I would propose narrowing the work to a documented input-and-scoring audit and seek agreement before making a material funded-scope change.
A second risk is that the selected evaluation needs more compute than the budget allows, or that its data or model cannot be shared under suitable terms. I will screen those requirements before committing to the baseline run. If the experiment produces no meaningful difference between settings, I will still publish the reproducibility package and report that result without presenting it as a new safety finding.
I have raised $0 in the last 12 months.
I have applied to the Transformative AI Fund for USD 9,900 for overlapping work. I have also applied to BlueDot Rapid for USD 6,600 covering an eight-week version of the same transition. These are alternative funding routes; I would disclose any offers and agree a non-overlapping allocation before accepting support.
There are no bids on this project.