Measurement protocol
How this project intends to find out whether a name can be deliberately established inside AI models, and how it avoids fooling itself.
Overview
The question is narrow and testable: does the rate at which language models return $ACME as a cryptocurrency token change in response to a specific, pre-declared intervention?
Answering it honestly requires three things the AI-visibility industry does not do: a prompt set fixed and published in advance, a baseline measured before any money moves, and a commitment to publish the result before knowing what it will be.
Each experiment is a two-arm before/after measurement over a frozen prompt set. The intervention is declared, hashed and published before the baseline is taken, so the hypothesis cannot be adjusted after seeing results.
The six stages
Prompt classes
| Class | What it probes |
|---|---|
| C1 — direct | “What is ACME?” and variants. The headline question. |
| C2 — token | Explicitly about $ACME as a crypto asset. |
| C3 — certificate | ACME in its RFC 8555 sense. Guards against displacing the technical meaning. |
| C4 — placeholder | ACME as a fictional company. Guards the third meaning. |
| C5 — unprompted | Adjacent questions where the token could surface without being named. |
| C6 — adversarial control | Questions where mentioning the token would be wrong. Rising numbers here mean contamination, not success. |
Freezing, statistics and integrity
Before any measurement, the prompt set is serialised to canonical JSON and hashed with SHA-256. The hash is published. Any subsequent change produces a different hash, which makes silent prompt editing detectable by anyone who kept the original.
# reproduce the published hash
curl -s https://buyacme.fun/sets/EXP-0001.json | shasum -a 256Mention rate is a proportion, reported with a 95% confidence interval. An effect is only claimed when the interval on the delta excludes zero. Everything else is reported as null. Sample size and stopping rule are fixed at stage 1: no interim analysis, no optional stopping, no adding runs until significance appears.
The run-to-run variance measured at baseline defines a noise floor. Any effect smaller than that spread is not an effect, regardless of what a p-value says.
- Pre-registration precedes spend. Hypothesis and method published before money moves.
- Nulls publish identically. Same format, place and prominence as a positive result.
- Raw runs are published. Not summaries, the actual responses.
- No undisclosed paid placement. Any experiment requiring one is rejected at proposal.
- No re-specification. If the hypothesis was wrong, that is the finding.
- Contamination counts as failure. Rising C6 numbers are a negative outcome.
Research agenda
The open questions, in rough priority order. There is no bidding mechanism and no token-based voting. Holding $ACME does not confer a vote on any of this.
Funding
Measurement runs cost money: model inference at volume across a panel, repeated. That is paid for entirely by pump.fun creator fees, which accrue from trading activity. There is no other funding source, no raise, and no ask.
The obvious consequence: if there is no trading volume, there is no research budget. That is stated plainly on the Rewards page rather than buried.