You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$2,500 left to reach the funding goal.
Purpose
People learning economics in Kazakh, Russian, and English may encounter confident AI explanations that are difficult to check. This six-week desk-based pilot would examine whether source links and a short verification checklist make errors and unverifiable claims easier to identify. It would produce a small open research resource, not personalized financial advice or a representative ranking of AI systems.
Design and outputs
At the target budget, the project would use 30 non-personal economic-learning questions, 10 in each language, with three response conditions: answer only; answer with sources; answer with sources and a verification checklist. This gives 90 planned answer records from one documented model/version. Questions would be selected and the coding guide fixed before the scored collection. Topics would concern general economic concepts, avoiding individual financial, medical, or legal decisions. Model choice would be made after checking price, language support, and applicable terms; no particular model is already secured.
The lead would record prompts, model identifiers, dates, settings available through the interface, answers, and the public sources used for checking. An independent paid reviewer would check the records using a common rubric: factual errors, whether cited sources exist and support the associated claim, explicit uncertainty, and whether a reader can access a relevant source. The lead and reviewer would record disagreements and their resolution. The report would disclose reviewer recruitment, qualifications, and review coverage actually achieved. No claim of blinded assessment or inter-rater reliability would be made unless implemented and measured.
Deliverables would be a question set, annotation rubric, a table of permissible answer records and source checks, a descriptive report by language and condition, and a reproducibility/readme file. Counts and illustrative error types would be reported without causal claims about learning or generalization to other models. Materials would be released without a paywall in a public repository, excluding content that cannot legally be redistributed. The project requires no interviews, students, minors, private financial records, or restricted institutional datasets. No human-subject experiment is proposed.
Work plan
Week 1: finalize the protocol, recruit an independent reviewer, check source and model terms, and test the coding guide on examples excluded from the scored set.
Week 2: finalize the question set and collect the documented response conditions.
Weeks 3–4: code and independently check responses, verify citations, and record disagreements.
Week 5: conduct reproducibility/data checks and prepare descriptive tables and the report.
Week 6: complete language proofreading, release permitted materials and the report, and publish an expenditure and limitations summary.
The clock would start only after an acceptable grant agreement, usable funds, and confirmation of the lead’s availability. No calendar start date or personal advance payment is promised.
Budget (USD)
Minimum Target
Lead: design, coding, report 700 1200
Independent review and adjudication 400 600
Kazakh/Russian proofreading 150 200
Model/API access cap 100 200
Public hosting/storage cap 50 100
Separate reproducibility/data QA 100 200
TOTAL 1500 2500
Lead and QA estimates use $25/hour: 28+4 hours at minimum; 48+8 hours at target. Reviewer estimates use $25/hour: 16 or 24 hours. Reviewer and proofreading rates are proposed estimates, not contracted quotes; roles are not yet secured. QA hours are separate from design/report hours. Hosting would use free infrastructure where feasible; unused allowances are not automatically personal remuneration. Final expenditure, any reallocation, and handling of unused funds must follow the actual agreement. Bank, exchange, and tax costs remain to be established; no personal top-up or zero-cost transfer is assumed.
Minimum viable scope
At $1,500 the project would cover 15 questions (5 per language) in two conditions: answer only and answer with sources, for 30 answer records, with a shorter descriptive report and the same public coding guide. The checklist condition and larger sample require the target budget. Scope would be declared before collection; it would not imply that work is already complete.
Funding parameters, risks, and openness
The proposed target is $2,500 and minimum is $1,500, with a proposed fundraising deadline 45 days after the candidate authorizes public posting. Below minimum, delivery would not be promised. Funding is not guaranteed. Kazakhstan recipient/bank eligibility and tax/reporting treatment are unresolved; no banking information would be posted publicly. Material changes in costs, reviewer availability, or permitted data release would be disclosed and addressed under the actual grant terms before expenditure. No cost or work hour would be charged twice to this project and another grant.
The design is intentionally small. Translation and question difficulty may differ by language; model outputs may change over time; the review may contain subjective judgments. The descriptive results would not establish that any prompting condition causes better learning or represents all users. Negative or inconclusive results would also be reported.
The project is proposed personal research. No university, employer, referee, or other organization has endorsed it. Recommendation letters, private contact details, and financial information are not part of this public draft.