Growth Search / A falsifiable simulation
Map before you spend.
A set of plausible growth levers. Find the ones most sensitive to intervention. The system is hidden, the learning budget is fixed, and every strategy begins with the same evidence.
A cheap model need not replace an experiment to help choose which expensive experiment comes next.
The search
One seeded landscape is an illustration. The evidence comes from repeated simulated landscapes below.
You have a bounded learning budget. How do you find the levers that can move the system?
- Same start: identical initial evidence and budget.
- Hidden truth: evaluation only; strategies cannot inspect it.
- Different evidence: observed opportunity is not causal sensitivity.
Each numbered tile is the same lever in every map. An outer border marks a strategy’s selection; after reveal, an inner border marks the true top-k. Tap a tile for its estimate and uncertainty. Use the ? buttons to explain settings, strategies, and results. On a phone, swipe between maps.
Run the search to inspect the actions that made each map.
What happened this time
This recap evaluates the finished searches. The maps above replay their steps; reveal ground truth to see the exact lever ranking. A sensitive lever can have a helpful or harmful effect.
What looks broken is not necessarily movable.
The left ranking is visible from observed evidence. The right is held back for evaluation.
Observed opportunity
Reach × observed impact × confidence; a useful but confounded starting point.
Intervention sensitivity
Absolute average causal response across fixed contexts. A large effect can be helpful or harmful.
Repeated landscapes
1,000 paired landscapes. Mean intervals are normal 95%; recovery-rate intervals use Wilson’s method.
| Strategy | Final top-k recall | Endpoint full set | Ever full set | First-recovery cost | Map error | Exposure |
|---|---|---|---|---|---|---|
| Use “Run 1,000 landscapes” below to compare results across simulated businesses. | ||||||
Proxy diagnostic: calculated after repeated runs; target fidelity is a parameter, not a promised realized accuracy.
One example shows how the search works. Repeated landscapes show how often each strategy succeeds, using your current settings.
What held up across repeated runs
Fidelity is a condition, not a verdict.
All four strategies are evaluated at 50–95% target proxy fidelity. Lines report recovery; the table carries the cost and censoring fallback.
Scientific mode
Inspect the model before reading the result.
The page generates a hidden nonlinear, segmented response system. Strategies receive only an evidence oracle; the evaluator later compares their selected top-k levers against causal intervention sensitivities. It is a thought experiment with explicit assumptions, not a calibrated forecast.
Target estimand and evaluator
Sensitivity(i) = |Es,X−i[(f(Xᵢ=0.8) − f(Xᵢ=0.2)) / 0.6]|
This is the magnitude of a fixed-range average causal response, weighted across three segments (46%, 34%, 20%) and 24 seeded contexts. A large harmful effect also counts as sensitive; the map does not recommend a rollout direction. Map RMSE concerns this sensitivity map, not recovery of an unknowable full response surface. “First recovery” is an evaluator timestamp: it records when a strategy happened to select the exact hidden top-k, which the strategy itself cannot certify.
Evidence, proxies, and strategy limits
Observed reach, drop-off, impact, and confidence may disagree with intervention sensitivity through noise and confounding. Synthetic probes fit initial/historical evidence, include persistent bias and uncertainty floors, and do not create independent information through repetition. “Fidelity” sets a target proxy-reliability parameter; realized directional accuracy is measured in simulated output rather than assumed.
Cartographer chooses among analysis, synthetic probes, micro-tests, and RCTs using an approximate expected information-gain-per-cost rule focused on top-k uncertainty. This approximation can be wrong. All actions respect a common maximum budget.
Uncertainty, reporting and reproducibility
Intervals for means use paired Monte Carlo samples; full-set recovery rates use Wilson binomial intervals. The downloaded regret field is a normalized gap in summed marginal sensitivities, not foregone revenue or the joint value of intervening on a set. Negative effects can be highly sensitive. Histories and live-test outcomes are stylized; there is no claim of external calibration.
Every download includes engine version, full normalized configuration, seed, strategy metrics, paired Monte Carlo differences when supplied, and diagnostics. The single map is labeled illustrative. Repeated runs carry confidence intervals. Cost-to-discovery is censored by failures and reported conditionally; recovery stays unconditional.
Full equations, parameters, and limitations
Sources and framing
Methodological inspirations, not claims of exact implementation: Song et al. (2019), cost-sensitive multi-fidelity optimization; Kaufmann, Cappé & Garivier (2016), top-m identification; Morris, White & Crowther (2019), ADEMP simulation-study reporting.