You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
$20,000 left to reach the funding goal.
I am developing SACL, a controlled benchmark of whether language-model agents recognize that their own earlier outputs changed their environment and avoid treating self-produced artifacts as independent confirmation. Such evidence loops could amplify errors and weaken oversight. This is a test of a possible failure mode, not an established claim about deception, consciousness or deployed-agent safety.
What are this project's goals? How will you achieve them?
I will develop matched critical/control scenarios covering files, tool use, research and planning; preregister scoring, configurations, exclusions and uncertainty estimates; and evaluate selected models with repeated runs. Ablations will test wording, position and provenance cues. Months 1–2 focus on scenarios and validity; months 3–4 on evaluations and ablations; months 5–6 on analysis and open release. The project would start after funding and access are confirmed.
How will you know it worked?
Success means a reproducible evaluation protocol, interpretable uncertainty, versioned code, development examples and a public report including null results. If the scenarios do not validly isolate self-attribution, I will report that failure rather than claim a model effect. Scripted tests cannot establish real-world agent safety or directly demonstrate reduced catastrophic risk. Any held-out items remain separate, with their treatment disclosed.
How will this funding be used?
Target USD 20,000: USD 18,000 total researcher cost including taxes/contributions over six months, and USD 2,000 API/GPU compute. This narrower scope limits models and ablations and does not promise paid independent reproduction.
Minimum USD 5,000: one funded researcher-month spread across two calendar months (USD 3,000 total cost), up to USD 1,500 API/GPU compute, and USD 500 explicit contingency. The pilot delivers a small matched development set, protocol, scoring checks and first results with uncertainty. This minimum does not fund two full-time researcher-months.
Who is on your team? What's your track record on similar projects?
I am Jana Kolářová, an independent researcher in Prague without institutional affiliation, and will conduct the study myself. ORCID: https://orcid.org/0009-0000-9588-4256. I develop and run measurement experiments, including the public SCETES collector: https://github.com/theesresearch/scetes-collector. My E0 pilot measures GPU inference energy; it is distinct from SACL. I do not claim that this previous measurement work establishes expertise or a positive result in agent self-attribution.
What are the most likely causes and outcomes if this project fails?
The main risks are invalid scenarios that reward surface cues, unstable model access, insufficient power and limited independent validation. I will prioritize validity checks and a small auditable scope, report uncertainty and inconclusive findings, and reduce the evaluation matrix if compute access is limited. An invalid benchmark or null result is a possible outcome; no positive model effect is promised.
How much money have you raised in the last 12 months, and from where?
I have received no research grants or donations for this research in the last 12 months.
On 23 October 2026 I submitted TAIF for SACL (USD 30,000, with a USD 20,000 narrower option), Anthropic ERAP for SACL (USD 1,000 API credits), and AI for Science for the separate E0 study (USD 3,000 Claude.ai usage credits). These applications are pending; no award from them has been confirmed. This Manifund target is an alternative route for the narrower SACL project, not an additional USD 20,000 on top of a fully funded TAIF award. I will disclose decisions and adjust or withdraw overlapping requests before accepting duplicate support; credited compute will reduce uncovered compute requirements.