Private access · by request
Compare the options of a decision on your figures
A decision engine for companies that have to choose: a price, a hire, a retention offer, a channel. Each option is simulated on your own figures, under the same random draws. The result names the winning option with its precision, or asks for the figure worth narrowing.
Engine
6 questions answered by the engine, replayed as recorded
QuestionWe have $180,000 in the bank and burn about $22,000 a month, with revenue of $9,000 a month that moves around a lot. What are the chances our cash lasts 12 months?
Engine answer
Conditional probability of lasting 12 months without going below zero: 93%, if your spending stays at 22,000 per month and your revenue evolves like that of companies with a similar profile. At the current pace (revenue 9,000 − spending 22,000), cash lasts 13.8 months.
- Answer
- 93%
- Read from the question
- “burn about $22,000 a month” · “We have $180,000 in the bank” · “our cash lasts 12 months” · “revenue of $9,000 a month”
- Answered in
- 6.65 s
Data table
| Months from now | Series | Cash in the bank | Low | High |
|---|---|---|---|---|
| 0 | Most likely | 180k | 180k | 180k |
| 1 | Most likely | 167k | 165k | 170k |
| 2 | Most likely | 154k | 150k | 160k |
| 3 | Most likely | 141k | 136k | 150k |
| 4 | Most likely | 128k | 121k | 140k |
| 5 | Most likely | 115k | 105k | 130k |
| 6 | Most likely | 102k | 89.8k | 121k |
| 7 | Most likely | 89.4k | 74.4k | 112k |
| 8 | Most likely | 76.5k | 58.8k | 104k |
| 9 | Most likely | 63.6k | 42.8k | 95.6k |
| 10 | Most likely | 50.8k | 27.2k | 86.6k |
| 11 | Most likely | 37.9k | 12.1k | 79.2k |
| 12 | Most likely | 25.1k | −3,901 | 70.1k |
Recorded from the engine on 2026-09-30, commit 0510ba55f4e7
Difference
Figures computed by the simulation, from the question you ask
A language model reads your question and picks out the figures you gave. Each number of the answer comes from the simulation, run on those figures. Below its precision thresholds, the engine asks for a narrower figure or names no winner.
Read the proof protocol →Method
From a question to a winner, in four steps
- 1 · Question
- a decision asked in plain words, with the figures you have
- 2 · StoneIR program
- the question compiles into a typed program, and each number carries its justification
- 3 · Same worlds
- each option is simulated on the same random worlds, so the draw cancels out
- 4 · Winner or refusal
- the winner with its precision, or no winner below the thresholds
Measured
Stated probabilities checked against observed frequencies
On 75,998 probabilities served, the calibration error is 0.0107, for a pass threshold of 0.03. A perfect forecaster scores 0.0018 on the same cases.
- Level forecasts
- 80.5% inside the 80% interval · 90.0% inside the 90% interval · 300 cases
- Cold-start forecasts
- 79.1% inside the 80% interval · 89.5% inside the 90% interval · 600 cases
- Monthly failure counts
- 82.3% inside the 80% interval · 92.9% inside the 90% interval · 105 cases
- Cases with no winner
- 0.2%
- Registered runs
- 26 runs over 13 families of decisions
- Measured
- 2026-09-29 on the development lot, engine commit 7a86f874
Data
The data under the engine.
- Weekly store × product observations
- 22,717,635
Compute core
Compute core in Rust
Under The Cairn, the simulation runs on a deterministic compute core. The same program and the same inputs give the same result, world by world.
Founder
A decision engine built by a quant researcher
Mahé Chadeffaud built Stochastly, a quantitative research terminal where the work is to tell a real edge from a false positive.
Compare your next decision on your figures
Private access, opened organisation by organisation, with packages by request.
The packages →