Napoleon at Brienne, by Job (1908)
Napoléon à Brienne — Job (J. Onfroy de Bréville), in G. Montorgueil, Bonaparte (1908). Public domain.
Berthier
Adjudication Tool for Flight of the Eagle
Version 1.0.0-beta.713 · BETA

Berthier is a digital umpire for Didier Rouy’s Napoleonic matrix kriegsspiel, Flight of the Eagle. It holds the true state of the campaign, adjudicates movement, combat, supply and the passage of dispatches by the published rules, and shows each commander only what his couriers, scouts and maps could have told him.

For the research program it is the instrument of record. Human, A.I. and A.I.-assisted staffs play the same historical campaigns under the same rules and the same fog of war, and every order, dispatch and outcome is logged for analysis.

It suits the task because the campaigns are bounded and documented: every result can be measured against what actually happened, with Napoleon’s own record as the baseline. Play departs from history on its own, so each game becomes a counterfactual case study, and advice is judged as it is in war — by whether it works when it is followed. One umpire engine applies one set of rules to every seat, so differences in outcome come from the decisions, not the referee.

Named for Marshal Berthier, Napoleon’s chief of staff, who kept the orders, the reports and the map.

About Berthier
For ResearchersFor EducatorsFor ProgrammersFor Hobbyists

Berthier

Adjudication Tool for Flight of the Eagle

Berthier is a browser-based digital umpire for Didier Rouy’s matrix kriegsspiel Flight of the Eagle, covering Napoleonic campaigns from 1792 to 1815. It holds the true state of the campaign; adjudicates movement, combat, supply, weather, sieges and the passage of courier dispatches by the published rules; and gives each player a fog-of-war view limited to what that commander’s couriers, scouts and maps could have told him.

It is the wargaming platform for an experiment that benchmarks A.I. as a strategic advisor. Human, machine and machine-assisted seats command formations in the same historical campaigns, under courier-lagged, partly false intelligence, against an opponent trying to win, and every order is stamped against what its seat knew when it wrote it. The pre-registered design states four hypotheses: that expert raters cannot identify a seat’s type above chance (indistinguishability); that machine and machine-assisted orders are rated no worse as written (quality); that those seats close their decision cycle faster relative to the adversary who must react (tempo, reported only together with effectiveness); and that they contribute at least as much to campaign outcome as human seats in the same role (effectiveness, measured against what the historical commanders achieved from the same position).

The design inverts John Searle’s Chinese Room. The umpire and the expert reviewers are the people in the room, and the A.I. slips its orders under the door. The umpire, blind to which seat is which, has to understand each order well enough to carry it out, under rules that do not care who wrote it, and effectiveness is scored on the campaign’s outcome, never on winning the umpire’s adjudications. If an A.I.’s orders, understood and executed that way, win against a live opponent, the A.I. has shown strategic understanding, or at least acted as if it had it: a bold claim, but an empirical one that the experiment can confirm or refute. Indistinguishability and quality are measured alongside outcome for a pragmatist reason: experts who do not trust advice will not follow it, however good it is. The authors’ prior is that the machine-assisted (centaur) seats will outperform both the unaided human seats and the machine seats, and the design states in advance how each other ordering of the three would be read. The most interesting result would be machine advice that is consistently better than that of human seats and of centaur seats, that blinded reviewers cannot identify as machine-written, and that arrives at a significantly faster tempo; human seats outperforming both would be as telling, because it would mean the advice degrades the commanders who take it.

It is a single-file React application with no build step that runs in a browser or from a local file, with a Node.js WebSocket relay for online play: per-role fog-filtered packets, assistant umpires with assigned players, and hosting handover. It carries sixteen scenarios from 1792 to 1815 and a map pipeline that traces and georeferences period campaign maps, and about three hundred regression harnesses gate every change.

More: For Researchers · For Educators · For Programmers · For Hobbyists

Built by Daniel Tyler Brooks. Game design © Didier Rouy. Developed with A.I. assistance (Claude, Anthropic). Title image: Napoléon à Brienne, by Job (1908), public domain.

Loading the umpire…