4reacting characters — a regulator, an incumbent, a rival, the press
replayablefork any turn and try a different move to compare outcomes
not a crystal balland it says so on every page it produces
How it works
- Cast of characters. Each one — a banking regulator, an incumbent bank, a
fintech rival, the press — gets a written profile: what it wants, what it will never do,
and exactly what information it can see. You read and edit these before the game starts.
- Turn by turn. You make a move; each character responds using only what it
can see; a referee keeps everyone honest and injects surprises — a currency shock, a leak.
- The payoff. A debrief of the reactions you provoked and the assumptions
that mattered — the "we never considered that" moves worth war-gaming further.
The honest frame is the product
Rehearsal is a practice room, not a fortune teller. Simulations don't
predict the future — they surface the questions you should have asked, and every debrief
says exactly that at the top. What it can guarantee is discipline: a referee
blocks any character from crossing its stated red lines or spending money it doesn't have,
so the rehearsal stays believable.
Proof it works
The built-in scenario is "launch SME lending in Mexico at aggressive
pricing" — the exact decision Bridge underwrites. Run it
and the story unfolds plausibly: the incumbent lobbies then matches your price, the regulator
escalates from an inquiry to a binding pricing rule, the press picks up the consumer angle,
and a peso shock tightens funding. Start with fair pricing instead, and the run ends at
roughly half the regulatory heat — a concrete "what if we'd chosen differently."
For technical readers
- Actors decide from bounded information (only the indicators their card exposes) and self-enforce their red lines and budgets; a facilitator independently re-checks every move and blocks any that would cross a red line or overspend — a unit test proves it catches and substitutes a bad move.
- Deterministic rule-based actors (reproducible, CI-gated) plus a Claude actor path under the same constraints.
- Branching forks any turn and replays; scripted shocks perturb the world mid-run.
- A Checkpoint eval gate verifies the core invariant — zero constraint violations across every tested strategy (8/8).
Honest limitations — this is the research bet
- Consistency is measurable and enforced; plausibility is not yet validated against expert raters.
- The real test — retrodiction, re-creating how a past regulation rollout unfolded from only what was public beforehand — is the next research step, not built yet.
- It's a rehearsal, deliberately framed as one. Its characters draw realism from
Compass briefings and Radar's regulatory history.