You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
I am an independent researcher. In 2026 I completed a self-funded mechanistic interpretability study showing that a user's engagement style corresponds to a measurable signature in a language model's attribution graphs. I classified naturalistic conversation segments by the user's communication policy, Navigator (steering, correcting, deepening across turns) versus Operator (directive, self-contained), built 54 matched pairs controlling task, domain, and length, and traced them through gemma-2-2b with Gemma Scope transcoders using the open-sourced circuit-tracer stack. The primary indicator, activation-weighted mean layer, locked before the cohort analysis, is higher in Navigator segments in 31 of 35 primary pairs (+0.554 layers, sign p = 3.5e-6), direction holding in 43 of all 54. Navigator graphs are also smaller and sparser, so graph size does not produce the shift.
Everything is public: preprint and interactive evidence at https://theveil.dev, full reproduction repository at https://github.com/Sarafova-AI/every-contact-leaves-a-trace.
This is a retroactive ask: it reimburses the costs of work already completed and published. It funds no future work and commits me to nothing forward.
The project is complete, so the goals are delivered: a tested measurement with one primary endpoint locked before the cohort analysis, 54 matched pairs traced, a preprint, an interactive evidence site, and a full reproduction repository.
Why it matters: mechanistic interpretability reads model internals but treats the user input as a fixed prompt. Human-AI interaction research varies the user but measures only outputs. This study sits in the gap between them: the human side of the exchange treated as a variable, read directly in the computational substrate. The methods are deliberately boring and reproducible: matched pairs, a public corpus, open tooling.
The follow-up agenda is published at theveil.dev/#horizon. If I ever seek funding for it, that will be a separate ask.
Reimbursement of the direct costs of the completed study, plus a deliberately partial compensation for research time. Costs incurred, March to June 2026:
RunPod H100 GPU time (consumed by tracing sessions): $128
OpenRouter API (source-conversation generation): $1
theveil.dev domain registration: $10.81
Claude Max subscription, 4 months x EUR 117: EUR 468
Claude extra usage credits: about EUR 60
ChatGPT subscription, March to June (per invoices): EUR 240.72
Research time, 100 hours x $30 per hour: $3,000
Total: $3,970 (EUR converted at about 1.08).
On the research time line: the study took four months of evenings and weekends, several hundred hours across instrument design, segmentation and matching, trace sessions on rented GPUs, statistics, an adversarial review pass, the paper, the site, and the reproduction repository. I claim 100 hours. The number is deliberately partial.
One disclosure for full transparency: my application to the EA Long-Term Future Fund, submitted 2026-06-11 and pending, includes a retroactive component for this same completed study. If both fund, I will disclose to each and the overlap will be netted. The same cost will not be paid twice.
I work alone. I am a vascular surgeon by training: seven years in hospital practice, board certified in 2015. In 2016 I moved into clinical research as a research physician, and for the last six years I have worked in clinical trial feasibility, today as a feasibility strategist at a global contract research organization. I started using AI in January 2023 out of curiosity; it stuck, and I have not spent a working day without it since. The study is that trajectory carried into research I own: clinical trial methodology (matched pairs, endpoints locked before analysis, blinded rating) applied to a question in mechanistic interpretability, built end to end with Claude as research partner.
The work is already done, public, and reproducible, so there is no execution risk in this ask. If it does not fund, nothing changes about the research; the costs simply remain mine.
Anthropic External Researcher Access Program ($1,000 in API credits): submitted June 2026, pending. EA Long-Term Future Fund: submitted 2026-06-11, pending. Nothing received to date.