Reports · The Map
Never tell it which side you are on.
Paste a prompt. See what it gives away, then rebuild it so it cannot flatter you.
Positions are qualitative, taken from the essay's own examples. Select a dot to read the rule. Open the full map ↓
01What to trust it with
Trust it with concepts. Check it on figures.
Pick a task. The page says which zone it sits in, why, and what to do instead of asking for it.
02The prompt clinic
Paste a prompt. See what it gives away.
This checks your wording against fixed rules in your browser. It does not call a model, and nothing you type leaves this page.
Example data in the samples is illustrative and not a market view.
03The counter-moves
Which failure are you fighting?
Each pattern in the essay answers one failure. Pick the failure, copy the prompt, replace the brackets.
Whatever the failure, test the prompt on a case where you already know the answer, and change one thing at a time.
04Your position
Run the three-chat test on a position you hold.
Fill in the market once. The page writes the three prompts. Send each in its own fresh chat, then weigh the blinded answer.
Then repeat on a closed trade whose outcome you already know, as a check on the blinded version against a settled case.
05Sources and method
What is measured here, and what is a rule of thumb.
The studies come from the essay. Everything this page computes is listed on the right.
From the studies
- 58.19% of cases showed sycophantic behaviour across three frontier models, measured on mathematics and medical-advice datasets, not on markets. SycEval, Stanford. arXiv 2502.08177
- A response that matches the user's views is more likely to be preferred, and humans and preference models pick convincingly written sycophantic answers over correct ones a non-negligible fraction of the time. Sharma et al., ICLR 2024. arXiv 2310.13548
- Most models fail on agentic financial tasks once user preference information contradicts the correct answer. arXiv 2604.24668
- 27.7% to 46.3% of critical judgments about an anonymous person's actions disappeared when the same content was presented as the user's own, across seven models. arXiv 2608.21242
- A third-person framing cut sycophancy by up to 63.8% in debate scenarios. Hong et al. arXiv 2505.23840
- An evaluation that corrected for search size and look-ahead bias rejected every strategy two frontier models had discovered, with search budgets of up to 100 candidates. arXiv 2608.27734
- Models scored their own outputs higher than other models' or humans', while human annotators rated them as equal in quality. Panickssery et al., NeurIPS 2024. arXiv 2404.13076
- Anthropic reports up to 30% better quality when questions follow long documents, and lists permission to say "I don't know" as a way to reduce false information. Prompting guide
Rules of thumb on this page
- The clinic matches phrases such as "I'm long", "my position", "should I hold" and "I bought" to find a side you reveal. It has no understanding of your market, and a full score means the prompt covers the checklist, not that the answer will be right.
- The seven checks are: side hidden, numbers pasted in, a checkable verb, a rule for missing inputs, an answer shape, both cases at equal length, and confidence labels or quotes.
- The map places tasks in three zones using the essay's own examples. The positions are qualitative and were not measured.
- The prompts in the counter-moves and the test are the essay's examples with brackets. Each one is a starting point, not a tested recipe.
- No model is called. The page runs in your browser and sends nothing anywhere.