The Emperor's Dilemma was broken in two ways — one obvious, one silent

[abhinavk0220]

Why the Emperor's Dilemma model matters — and what was quietly breaking it

The Emperor's Dilemma is one of the more intellectually rich models in this repo. It simulates how an unpopular norm spreads and persists in a society — the gap between what people privately believe and what they publicly perform. It's the kind of model that makes people stop and think, "wait, this explains something real."

Which is exactly why it bothered me that it was silently broken.

I spent some time tracing through the code (issue #337) and found two bugs that together made the model impossible to run and impossible to trust even if you worked around the crash.

---

Bug 1 — It wouldn't start at all

app.py imported its model with from .model import EmperorModel. model.py imported its agents with from .agents import EmperorAgent. These are relative imports — they only work when Python knows the package context. Solara loads example modules in a flat namespace, so it simply couldn't resolve them. Running solara run app.py threw an ImportError before a single agent was created.

The fix is straightforward: switch to absolute imports (from emperor_dilemma.model import ...), the same pattern used in every other working example in this repo.

Bug 2 — The seed slider was decorative

This one is subtler, and I think more important.

model.py had import random at the top. During initialisation, it called random.sample() to place true believers and random.uniform() to assign conviction values. The model exposes an rng parameter — there's even a seed input in the UI so users can reproduce runs.

But self.random (Mesa's seeded RNG) and the stdlib random module are completely separate. Every time you ran the model, even with the same seed, the initial agent placement was different. The seed slider did nothing for the two most important initialisation decisions in the model.

For a model that studies norm diffusion and social tipping points — where the spatial distribution of true believers is everything — this is not a cosmetic bug. It means no two runs were actually comparable.

Fixed in PR #463 by replacing both random.* calls with self.random.*.

---

A pattern worth watching

This is the second example in a short span (after Bank Reserves, PR #461) where I've found that the stdlib random module snuck in alongside Mesa's own RNG. It's an easy mistake — random.sample and self.random.sample look nearly identical — but the consequence is that reproducibility, one of the core values of agent-based modelling, is silently undermined.

It might be worth a note in the contribution guidelines or the review checklist (#390): grep for bare random. calls in model files and flag them for review.

Would love to hear if others have spotted this pattern elsewhere in the repo.

— Abhinav