Scenarios

Test the decision before you make it

Not a model of an average biotech. Cambrian reads your milestones, sites, batches, budgets and dossier, and shows what a decision does to them as ranges you can defend, with the assumptions written down.

An experiment comparing three scenarios on runway, cash, readout probability, required N and power

Engines

Seven engines, one contract

Engines are pure functions of the twin, the assumptions and a seed. They never read the clock, the database or the network, so a run reproduces exactly and every forecast names the observable that resolves it.

Engine	Model	Headline outputs
Timeline	Dependency graph over playbook steps, tasks, milestones and submissions; slip distributions propagated in order	Completion date per milestone, on-time probability, critical path frequency
Runway	Monthly cash with burn drift, milestone-driven costs, grants with award probability and a planned raise	Months of runway, cash-out date, cash at 12 and 24 months, funded probability per milestone
Enrollment	Gamma-Poisson site model with startup times, added sites, competing trials, dropout and screen failure	Last patient in, readout, months to target, evaluable N
Statistical design	Closed-form sample size and power for continuous, binary and time-to-event endpoints; placebo erosion; assurance	Required N, power at planned N, probability of success, minimal detectable effect
Supply	Monthly stock by batch with expiry, demand from enrollment, reorder policy, yield, failure and lead times	Stockout probability, first stockout month, expired units, batches ordered
Regulatory pathway	505(b)(1), 505(b)(2), IDE, EU CTA and 510(k) templates with meeting lead times, dossier completion, statutory clocks and hold branches	Submitted date, study-may-proceed date, hold probability, on-time probability
Outcome priors	Elicited effect, adverse event and adherence priors propagated through the design to a readout	Positive readout probability, safety-stop probability, observed effect

Workflow

One durable workflow per experiment

Steps are journaled. A crashed step resumes where it stopped and nothing runs twice. Without a model key the workflow still completes with catalog and record-derived assumptions.

01
Snapshot
The twin is frozen into a content-hashed snapshot. Identical states reuse it, so runs reproduce bit for bit.
02
Evidence and priors
Connectors pull comparable trials, papers and adverse event data from ClinicalTrials.gov, PubMed and openFDA. Elicitation strategies produce priors with a confidence label and citations.
03
Scenarios
Decision variables expand into candidate scenarios: more sites, a later raise, a different pathway or endpoint, a shifted batch.
04
Engines
Each engine runs per scenario from the same seed and horizon. Every number is a distribution summary or a probability.
05
Ranking and panel
Scenarios are ranked against constraints. Role personas, and the agency reviewer in a war game, return concerns and questions.
06
Report
A credibility record and a report are written into the Data Room as a versioned, signable document.

Credibility

Results state what they rest on

Every experiment carries a credibility record shaped after ICH M15 and the FDA AI credibility framework: question of interest, context of use, decision consequence, model influence, model risk, engines and versions, prior provenance, verification and validation, limitations.

The platform never claims to predict clinical efficacy or regulatory approval, and the interface has no place to show a bare point estimate of either.

A credibility record with question, context of use, model risk, engines, priors, verification, validation and limitations

Learning loop

Reality scores the platform

Each run emits forecasts about milestone dates, site activations, enrolled counts, batch depletion and the 12-month cash balance. A nightly resolver reads the live records, scores them (CRPS, Brier, 80% interval coverage, bias in days) and feeds the measured bias back into the engines. A drift detector re-runs the stress test when records move and files risks for critical findings.