Method and limits
How the game works, how a result is produced, what was left out on purpose, and what is known to be wrong.
[0.0]A grognard exercise, not science
Force on Force is a grognard exercise. A grognard is a wargamer of the old school: someone who will argue for an evening about a supply rule. This is that kind of project, done with care. The inputs are sourced, the guesses are labelled as guesses, every addition is tested and committed, and the paper trail is public.
It is not a scientific model. Nothing here has been peer reviewed. Many of its coefficients are judgement. It has two historical backtests, both against Iraq, and one of them is in-sample. It has never been checked against a war like the one it describes, because there has not been one. Some parts are known to be wrong and are listed as such.
So read every percentage as what this rule set produces, not as a forecast. The model also assumes the war has already started. It says nothing about whether one will.
[1.0]What it is
A campaign-level Monte Carlo wargame. It does not move units on a map. It tracks who runs out of troops, money, public support or will, and when. The first pairing is Canada and the United States; the engine does not know that, and other pairings can be added.
Every run is conditional on a war having started. Scenarios are a full invasion, an occupation of Alberta only, and economic pressure with no invasion.
[2.0]Sequence of play
- Opening fight, by day. Combat power is people times effectiveness times air superiority, with a penalty for the defender's dependence on the attacker's systems. Loss rates follow the power ratio. The phase ends when the defender breaks and the attacker has reached its objectives.
- Occupation, by month. A resistance index sets how many troops the occupier needs, following Quinlivan's force ratios. Control is troops deployed over troops needed. Too few troops feeds the resistance. Occupier deaths scale with troops, resistance and the shortfall.
- Politics, by month. US public support falls with deaths and economic loss; below a threshold there is a monthly chance of withdrawal, higher in election months. Canada's government has a monthly chance of capitulating, which rises with losses.
- Economy, by month. Lost trade, energy shocks, sanctions and war damage on both sides.
- Modules. Each switchable module adds to one or more of these steps. See Modules.
A war ends when the United States withdraws, Canada capitulates, the occupation is consolidated, or the clock runs out.
[3.0]How a result is produced
- Thousands of wars. Uncertain parameters are ranges. Each run draws its own values, so no two wars are the same.
- The same dice twice. To test a change, the same 3,000 wars are played again with one thing altered. Weather, luck and every other draw are identical, so the difference is the change and nothing else.
- Intervals. Differences are reported with 95% intervals. A difference whose interval spans zero is called noise.
- One switch per module. A module that is off must leave results exactly as they were. That is checked run by run.
[4.0]Checks
- 32 automated tests, including one that fails if any parameter lacks a source line.
- Harness checks on every version: same seed gives the same answer; outputs stay in bounds; a module is neutral when off, cannot affect scenarios it has no path into, and pushes in the direction it claims; the provincial model with identical provinces reproduces the single-area model.
- Stored reference results that must not change unless a version says they should.
- Two backtests against history, below. Two is not many, and the second is in-sample: the death rate was tuned to it.
| Case | Model median (p10–p90) | Observed |
|---|---|---|
| Iraq 2003 invasion, US KIA | 157 (93–278) | ~139 (DoD, all causes to 1 May 2003) |
| Iraq 2003–11 occupation, US hostile KIA | 3,545 (2,879–4,325) | ~3,500 |
[5.0]Left out on purpose
The game works with national and provincial aggregates. It names no sites, routes, facilities, methods or tactics. Infrastructure appears as a category that can be degraded, never as a list of components or weak points, and nothing here ranks ways to damage civilian systems. That is a design rule, not an omission to be filled later.
[6.0]Known faults
Flagged by recent versions, and next to be fixed:
- Training is invisible after the opening fight. A volunteer with one week counts the same as one with six months once the war is an occupation (0.1.3.7).
- A mobilized country counts its dead like an unmobilized one. That makes extra fighting look self-defeating for Canada in a full invasion (0.1.3.7).
- The effect of mobilization on Canada's will is one hand-set number. It decides whether the 400,000 plan is worth 4 points or 12 (0.1.3.7).
- Provincial governments are not actors. Whether one accommodates an occupier is a switch set by hand (0.1.3.5).
Standing since the first version:
- Defender capitulation hazard is uncalibrated. It drives the headline outcome in long wars. It needs a case base of occupied or coerced states (Denmark 1940, Norway 1940, France 1940, Czechoslovakia 1938–39, Baltic states 1940, Kuwait 1990, Crimea 2014, Georgia 2008) and should model a collaborationist-government path separately from capitulation by the legitimate government.
- The occupation backtest is in-sample. The KIA rate was tuned to Iraq 2003–11. It needs out-of-sample validation: Afghanistan 2001–21, Vietnam 1965–73, Soviet Afghanistan 1979–89 and Chechnya.
- Initial attacker public support is the dominant driver of the Canada–US results. It should become a scenario-level model fed by polling series (Angus Reid, YouGov, Gallup), not a single sampled number.
- Armed propensity has no Canadian precedent. Bracket it with comparative evidence on resistance in high-income democracies, and model civil resistance (Chenoweth) as its own channel with its own outputs.
- Economic coefficients are order-of-magnitude placeholders. Replace them with OECD TiVA Canada–US value-added, Bank of Canada / CBO trade-shock elasticities, and EIA refinery-dependence data (PADD 2 and PADD 4 runs on Canadian heavy crude).
- Some actor values are estimates: US population and GDP, Canadian land-combat share, and Alberta and territories population. Replace them with StatCan, BEA, Census and DND Departmental Results data.
The full list of what comes next is on the roadmap.
[7.0]Who made it
Force on Force is a project of Prime Rogue Inc.. Scope, direction and editorial judgement are Kevin Duska's. The model code, the test harness and the results documents were written with Claude, an AI system made by Anthropic, working to that direction; the commit history records it.