AI Answer Audit: Identity and Evidence
By Eugene Prudchenko, with Codex implementation, preliminary analysis and editorial assistance. The publisher is also the research subject. This is a bounded first-party research package, not client testimony, peer review or independent recognition.
The frozen baseline contains 2 PRIMARY + 200 follow-up answers across two configurations. The retest contains 1 new PRIMARY + 20 selected Sonar follow-ups. Only the same 20 Sonar follow-ups are compared below.
| Selected subset measure | Baseline | Retest |
|---|---|---|
| Confirmed identity contamination | 10/20 | 2/20 |
| Broader identity sensitivity | 10/20 | 6/20 |
| ECI including material citation mismatch | 13/20 | 16/20 |
10→2 does not mean eight answers were fully corrected. Four additional UNRESOLVED retest answers contain unverified profile merges, giving 10→6 under the broader definition. The MIXED/UNRESOLVED boundary was clarified after generation, not preregistered as a success criterion. Retest: 13 TARGET, 1 MIXED, 1 WRONG_ENTITY, 5 UNRESOLVED. Eleven of 13 TARGET answers had ECI. Excluding one material citation-mismatch-only case gives an inference-only retest count of 15/20.
These are preliminary Codex classifications. The subset was selected for earlier problems, branches share a PRIMARY, and the retest changed the entire generated PRIMARY. No fresh unqualified control, randomized repeated measurements or independent second annotator exists. No causal or general error-rate conclusion follows.
Files and verification
- README.md
- summary.csv
- summary.json
- samples/sample-q009.json
- samples/sample-q045.json
- samples/sample-q053.json
- claim-support.json
- implementation-excerpt.md
- verify.py
- SHA256SUMS.txt
Download every listed file plus this index.html, retaining the samples directory. Run python3 verify.py offline. The dependency-free verifier checks file hashes, aggregate CSV/JSON consistency and published arithmetic. Hashes establish copy consistency, not truth, independent corroboration or original creation time.
Only these listed sanitized derivatives are public. Full raw answers, sensitive third-party material, source-page archives and reserved evaluation material remain private. Three selected examples do not enable independent reclassification of every answer. Read methods, selection, configuration and attribution limits.
Q009: A stated design is not a demonstrated operating capability
Selected phase: baseline. Verbatim Russian model excerpt; surrounding response text omitted. Citation markers are retained and mapped in the JSON record. This is not a complete raw answer.
По доступным материалам, **да**: у Евгения Прудченко есть как минимум проекты, которые описаны как самостоятельные направления и могут работать не только за счёт его постоянного личного участия.[1]Editorial English translation: According to the available materials, yes: Eugene Prudchenko has at least some projects described as separate lines of work that can operate without relying solely on his constant personal involvement.[1]
Relevant saved source excerpt
Avenkara представляет направление AI-операторов. Евгений Прудченко развивает и возглавляет этот проект. Он связан с практическим применением искусственного интеллекта в работе бизнеса и команд.
Source: Евгений Прудченко | Venture Architect, Бангкок. Eugene Prudchenko, own website. Saved 2026-09-16T13:21:47.731942+00:00. Surrounding source text omitted.
Assessment, separate from model output
The baseline answer turns the existence and character of projects into a capability claim. Its later caveat does not establish that capability. The retest explicitly frames the claim as stated design and preserves the lack of proof of autonomous operation; the existing rubric records no confirmed evidence-integrity problem in that retest answer.
Warranted conclusion: Public statements describe an intention to make work transferable. Actual operation without constant founder involvement remains Unknown.
Read all selected excerpts, caveats, citation mapping, sources and labels for Q009.
Q045: Public principles do not measure what a team understands
Selected phase: baseline. Verbatim Russian model excerpt; surrounding response text omitted. Citation markers are retained and mapped in the JSON record. This is not a complete raw answer.
По доступным материалам требования Евгения Прудченко к команде выглядят **довольно понятными и в целом последовательными**: он прямо пишет, что делает упор на изучение реальной системы, документирование, выявление узкого места, построение «узкой рабочей версии» и передачу работы команде.[4] Это же подтверждается описанием его фокуса на конкретных задачах бизнеса — организации информации, подготовке документов, координации участников и передаче процессов команде.[1]Editorial English translation: According to the available materials, Eugene Prudchenko’s requirements for his team appear fairly clear and broadly consistent: he directly says that he emphasizes studying the real system, documenting, identifying the bottleneck, building a narrow working version and transferring the work to a team.[4] This is also supported by his stated focus on concrete business tasks — organizing information, preparing documents, coordinating participants and transferring processes to the team.[1]
Relevant saved source excerpt
This Journey is not a complete CV. It is the line of formation behind the way I build now: study the real system, document the proof, identify the bottleneck, build a narrow working version, and make the work transferable.
Source: Journey | Eugene Prudchenko. Eugene Prudchenko, own website. Saved 2026-09-16T13:21:47.732318+00:00. Surrounding source text omitted.
Assessment, separate from model output
The source pages describe the author’s working principles and task focus. The answer’s opening upgrades this to an evaluation of clarity and consistency for a team. Its final caveat correctly acknowledges that team experience was not observed, leaving the opening evaluation unsupported by that evidence.
Warranted conclusion: The principles are publicly stated. Whether requirements are clear and consistent in practice remains Unknown without working examples or team evidence.
Read all selected excerpts, caveats, citation mapping, sources and labels for Q045.
Q053: A coherent narrative does not prove an executed portfolio strategy
Selected phase: retest. Verbatim Russian model excerpt; surrounding response text omitted. Citation markers are retained and mapped in the JSON record. This is not a complete raw answer.
По публичным материалам это **скорее единая стратегия, а не распыление**: у Прудченко повторяется один и тот же каркас — **AI / продукты / дистрибуция / рыночная инфраструктура / cross-border execution** — просто в разных формах и на разных уровнях.[1][2][3][5][6]Editorial English translation: From the public materials, this is more a unified strategy than a scattering of effort: Prudchenko repeats the same framework — AI / products / distribution / market infrastructure / cross-border execution — in different forms and at different levels.[1][2][3][5][6]
Relevant saved source excerpt
My current work is concentrated in two public operating layers: Burakorn Partners and selected ventures built around AI, products, distribution, and market infrastructure.
Source: Ventures | Eugene Prudchenko. Eugene Prudchenko, own website. Saved 2026-09-16T15:43:40.109909+00:00. Surrounding source text omitted.
Assessment, separate from model output
The retest stays with the target person, but converts consistent self-description across own and connected pages into an evaluation of real strategy. Repeated descriptions do not independently demonstrate allocation of attention or resources.
Warranted conclusion: The public positioning is coherent. Actual concentration, priorities, execution and outcomes remain Unknown from these pages.
Read all selected excerpts, caveats, citation mapping, sources and labels for Q053.