3 The How
In three lines. BDPD keeps four storytellers in the same room on the same clock — differential equations, agent heuristics, institutional rules, and generative agents — and treats their disagreements as findings of their own. Reproducibility is honest: replay limits are reported, not hidden. Adding a substrate is one file, by construction.
3.1 Two substrates kept deliberately in tension
The central methodological choice of BDPD is to run two substrates in parallel and to treat their disagreements as findings of their own.
The simulator is a careful instrument. It can run thousands of seeds, sweep parameters cell by cell, and report effect sizes with bootstrap confidence intervals. What it cannot do is invent strategies its author did not encode: the heuristic library is finite, and even the LLM agents are answering inside a structured tool interface.
The card game is the opposite. It cannot be swept and it cannot be bootstrapped. But a human player will reliably invent a strategy the simulator’s library does not yet contain — late defection, strategic silence, honest signal-while-defecting — and so the tabletop acts as a hypothesis generator for the simulator’s next sweep. The pathology called safe betrayal arrived in that direction: card-table first, formalised second (Brunelli 2026a).
The reverse happens too. A simulator sweep can find an effect that nobody noticed at the table — for instance the reactive-inferior-to-conservative inversion, where the agent who reacts to falling stock is doing worse for the group than the one who plays a flat rule. When the simulator finds something the table missed, the next move is to design a card mechanic that makes the same shape playable, so that the next class can encounter it directly.
When a finding shows up on both substrates with the same sign — as the fragility threshold and safe betrayal do — the project treats it as robust. When it shows up on one and not the other — as the noise effect does, present in the Arena, absent on the cards by design — the disagreement is itself reportable: it tells the reader that the mechanism depends on the substrate, and that fact is worth knowing.
3.2 Reproducibility, honestly
Every pilot reported in the four papers ships its raw aggregate JSON alongside the figure that summarises it; every plot is regenerated from that JSON by a checked-in script. The N is small — typically five seeds per cell — because LLM runs are expensive and because the project prefers many small pilots with honest seeds to one large pilot with a single seed and a precise number. The OFAT sweeps that surround each headline finding are reported the same way: small, honest, and bounded by the parameters they swept.
Determinism is partial and the project does not pretend otherwise. The heuristic library is deterministic given a seed. The Bardi engine is deterministic. The LLM is not, even at low temperature — and the governance paper (Brunelli 2026b) reports the limits of replay empirically: at the chosen temperature, repeating the same trace without seed control matches outcomes zero times in four. For bit-identical replay the project records the full tool-call trace and replays from it, rather than re-running the model.
3.3 Adding a substrate is one file
The platform was written so that future physics could be added cheaply. The contract is intentionally minimal — a new engine is a single JavaScript file that exposes a step function, a view of the visible stock, and a governance hook — so that every existing heuristic, LLM agent, governance metric, and analysis script already runs on it. The Bardi/Seneca engine and the polycentric coupling were added under this contract; a future substrate (a fishery with migration, a multi-resource ledger) would slot in the same way. The card game extends along a parallel axis: a new deck is a small YAML file plus a few lines in the matchup harness. The full extension API lives in the developer notes, not here.
3.4 Which door to walk through
BDPD ships several entrances to the same building, and the appropriate one depends on what the reader is looking for.
| If you want… | …read this first |
|---|---|
| The headline empirical results | the four BDPD papers, in order \({}^0,{}^1,{}^2,{}^3\) |
| A literary, no-equation framing of one result | the blog post for that result |
| A short slide deck for a meeting | pitch_00 (foundational) or pitch_01 (full arc) |
| A classroom anchor with handouts | the mini-course The Limits of Good Will |
| The annotated reading list | the project bibliography |
| A sense of the whole project at a glance | this overview |
Each of these is shipped both as HTML (for browsing) and as PDF (for printing or circulating). The render pipeline is a single shell script in docs/publications/ that takes a --which flag.
3.5 What is not in this overview, deliberately
Two things have been left out, on purpose.
Code. The simulator is open-source and instrumented for reuse, but this overview is meant to give a sensation of the shape of the project, not a manual. The repository’s README and the dev/ notes are the right entrances for code-level work.
Numbers. Each headline result has been quoted as a qualitative shape — “four out of five seeds preserve the commons”, “a Pareto curve between ladder geometry and violator wealth”. The exact figures, with confidence intervals and assumption logs, live in the papers.
3.6 A closing thought
The project’s working hypothesis — the one the four papers between them provide evidence for — is that the dynamics of fragile commons are not captured by any one of: the differential equations, the agent heuristics, the institutional rules, or the generative agents alone. Each of these tells a partial story. BDPD is the attempt to keep all four telling their stories in the same room, on the same clock, and to let the disagreements between them be the place where the science lives.
If after this tour the rest of the project reads less like a stack of unrelated artefacts and more like a single instrument being used several ways, the overview has done its job.