BDPD — Be Different Play Differential

The full arc: from a one-aggressor collapse to a four-paper programme on the Seneca cliff, governance, and the politics of timing

Roberto Brunelli

2026-06-29

“He cut down the cedar, he conquered the forest of cedars.” — Epic of Gilgamesh, Tablet V

Four papers, one laboratory. From a fragile commons that a single defector dooms, to a polycentric Seneca world where governance has to read the right signal at the right time.

Roberto Brunelli · 2026 · BDPD⁰ · BDPD¹ · BDPD² · BDPD³

The Cliff

“Increases are of sluggish growth, but the way to ruin is rapid.” — Seneca, Lucilium 91

Socio-ecological systems grow slowly but collapse rapidly — the Seneca Effect. Resource extraction accelerates through capital accumulation while pollution feeds back with a fatal delay.

Productivity peaks one instant before the cliff. By the time the decline is visible, recovery is already impossible.

Bardi’s canonical three-variable Seneca model: Resources decline monotonically; Capital rises, peaks, then crashes; Pollution ignites late and outlasts everything. Collapse is roughly twice as fast as growth.

One Is Enough

“He who controls the spice controls the universe.” — Frank Herbert, Dune

BDPD⁰ asked the first question: what does a single aggressive agent do to a logistic commons under empirically realistic regeneration (\(r \le 0.32\))?

The answer is binary. Across pool sizes, schedulers, endowments — one aggressor dooms the system. Rescue requires regeneration \(r \approx 0.95\), roughly \(8\times\) the canonical default of \(r = 0.12\).

Higher regeneration in the canonical band does not save the commons. It only buys a slower collapse.

Gate pass rate heatmap: regen rate × aggressive fraction. The only green region is zero aggressive agents. *One is enough.*

gate pass rate with one aggressive agent in the slow-regeneration band (\(r \le 0.32\), fisheries / slow-growth forests), all pool sizes, all schedulers, all endowments.

Smart \(\neq\) Good · Safe Betrayal

“It pays to be obvious, especially if you have a reputation for subtlety.” — Isaac Asimov, Foundation

Reactive < conservative. Individually rational adaptation — pull back when the stock falls — is collectively worse than unconditional moderation. The reactive agent creates an exploitable vacuum the aggressor fills. The smarter the response, the larger the subsidy.

Safe betrayal. In the card game Forest of Humbaba, late defection against a conservator profits the defector by up to +27% with no increase in collapse rate. The conservator’s healing is the slack the defector captures.

The tragedy of the compensator Unilateral conservation without exclusion is structurally counter-productive — it subsidises the very agents it cannot constrain.

It transforms the commons from a shared resource into a subsidised extraction ground. Compliance is punished; defection is rewarded.

A Dual-Mode Laboratory

“The map is not the territory — but the right map reveals the territory’s traps.”

BDPD is an open-source experimental and didactic testbed — infrastructure for the research community, not a predictive model with universal claims.

A configurable Node.js platform with two engines (logistic and Seneca R/C/P), three schedulers, configurable observation noise, perturbations, and agent slots (HTTP, sandboxed JS, LLM).

A physical card game — Forest of Humbaba — that partially reproduces the platform’s resource and governance dynamics on a tangible board-game substrate: Forest Deck, Box Reserve, Forest Die, Seneca Clip.

Convergence across substrates increases confidence. Divergence reveals hidden assumptions worth resolving analytically.

Platform → statistics Sweeps across thousands of runs reveal thresholds, slopes, and boundary conditions.

Card game → intuition

Discrete, stochastic play makes the strategic structure felt — and contrasts whether platform patterns survive a fundamentally different substrate.

Together → dual-mode science Findings on one substrate are cross-checked on the other for qualitative consistency.

The Governance Test

“What you cannot enforce, do not command.” — Sophocles

BDPD¹ takes the collapse from BDPD⁰ as given and asks the second question: on a fragile commons where one aggressor suffices, can communication and sanctioning restore sustainability?

The platform now ships a pact registry, a six-tool cheap-talk surface (broadcast / send_private / announce_intended_harvest / propose_pact / accept_pact / pledge), five governance metrics (cooperation_index, lie_score, sanction_rate, announce_frequency, silent_defection), and flat + graduated sanction perturbations — all engine-agnostic.

Four mini-pilots (C1 / D1 / D2 / D3), each at \(N = 5\) seeds with DeepSeek-flash LLM agents at the canonical collapse configuration.

Talk Is Cheap, Talk Is Not Enough

D1 — incomplete 2×2 factorial. Agent architecture (built-in vs LLM) × cheap-talk channel (off vs on), with the built-in + talk cell left out of scope (heuristic bots would only mechanically replay a fixed template, not signal). Three cells; the dominant effect is architecture, not talk.

Built-in cooperators preserve the commons in 0/5 seeds. LLM agents without any communication channel preserve it in 4/5 seeds. Adding the cheap-talk surface drops the binary preservation rate to 2/5, a sample-variance fluctuation at \(N = 5\) (Cohen’s \(d = 0.17\) on the continuous stock difference) — i.e. no detectable preservation benefit from talk.

Talk does, however, redistribute wealth — the aggressor’s earnings drop 33% when communication is enabled, while cooperators gain +17%. Cheap talk is an instrument of equity, not preservation.

BDPD¹ D1 headline: commons stock trajectories across the three implemented cells of the 2×2 (built-in + talk is out of scope). The vertical separation is architecture; the cheap-talk axis adds wealth redistribution but not preservation.

The Ladder That Pays

“It is not the strongest of the species that survives, but the one that is most adaptable to change.” — attributed to Darwin (Megginson)

D2 / D3 — graduated vs flat sanctioning. Same expected cost, same pact terms; only the shape of the escalation differs.

A graduated \([1, 3, 10]\) ladder preserves the commons in \(5/5\) seeds. A flat sanction of equal expected cost preserves it in only \(3/5\).

A ladder-geometry sweep then isolates which feature of \([1, 3, 10]\) carries the effect. A flat-5 control preserves only \(1/5\), while all three graduated schedules preserve \(\ge 4/5\). Graduation itself — not amount — is the load-bearing parameter.

The graduated cells trace a Pareto frontier between commons preservation and violator-wealth retention.

BDPD¹ D3 ladder sweep: soft \([1, 2, 4]\), hard \([1, 5, 25]\), and flat-5 across \(N = 5\) seeds. The flat schedule is in a different basin.

A World Above the Arena

“The thing that hath been, it is that which shall be — but never in the same arena.” — Ecclesiastes, paraphrased

BDPD² lifts the substrate one level. A World composes \(N\) existing arenas under a global clock, with resource links (one-way per-round flow), treaties (cross-arena constraints read by meta-agents), world-level meta-agents (read full world state and emit perturbations), and an exclude perturbation that migrates pact-violating players into a junk arena.

Four vignettes survey the surface:

V1 — Collapse isolation The platform single-arena collapse cleanly reproduces across 10 parallel arenas in one World. The world abstraction does not leak.

V2 — Reflexive vs voluntary sanctioner A visible sanctioner — whether reflexive or voluntary — deters the LLM defector. A delayed sanctioner that waits for repeat violations triggers collapse.

V3 — World-level treaty enforcer A meta-agent that reads world state and applies a cross-arena treaty fires rarely — and that is precisely why the treaty works.

V4 — Junk-arena exclusion Pact violators migrated into a designated junk arena. Built-in purity \(1.00\) — junk arena does what its name promises. Under LLM agents purity drops to \(0.6 \pm 0.3\): conformists co-breach the pact and follow the defector into the penalty arena.

The Inversion

A finding recurs across V2 / V3 / V4: the qualitative direction of governance effects (collapse yes/no, sanction-ladder ordering, enforcer firings) is robust to seed and substrate.

But the agent class is not just noise. Built-in heuristic agents produce flat baselines across all three sanctioner policies — there is nothing for the institution to discriminate against. LLM agents invert both the survival ordering and the defector-wealth ladder.

The institution does not act on the substrate. It acts on the signal-adaptive behaviour of the agents the substrate happens to host. A governance design that works under one population can systematically fail under another. The substrate is necessary; it is not sufficient.

BDPD² cross-cutting: the governance inversion across vignettes. V2 reflexive-cell violations 16 → 3 (LLM at N=5); V3 enforcer firings 1.0 → 2.4; V4 junk purity 1.00 → 0.6 — same institutions, different agent populations.

Seneca’s Real Engine

“All things are full of weariness; a man cannot utter it; the eye is not satisfied with seeing, nor the ear filled with hearing.” — Ecclesiastes 1:8

BDPD³ swaps the logistic engine for the Bardi/Seneca three-variable ODE:

\[ \dot{R} = -k_1 R C - l_3 R \qquad \dot{C} = +k_1 R C - k_2 C P - l_1 C \qquad \dot{P} = +k_2 C P - l_2 P \]

A natural resource \(R\) feeds industrial capital \(C\), which generates pollution \(P\), which degrades both. The Seneca cliff is the asymmetry: \(C\) grows slowly while \(R\) sustains it, then collapses rapidly once \(R\) is exhausted.

Three jurisdictions A, B, C are coupled by directional pollution links: a fraction of A’s \(P\) adds to B’s and C’s \(P\) as an externality without depleting A. The leading-following cascade is the smallest topology that admits two independent victims downwind of one emitter.

The Wrong Signal, Too Late

“By the time the pearl is rotten, the oyster has already drowned.”

S2 ships the cascade with no governance. Arena A emits; B and C inherit the pollution downwind.

Capital downwind drops by ~76%. Collapse arrives ~16 turns earlier. Peak \(P\) is ~7.1× the baseline.

And — counter-intuitively — the resource signal points the wrong way. The poisoned neighbours read as richer on \(R\) in the early phase: their own industrial activity has not yet drawn down the local resource, because the cliff hits them before \(C\) can take off.

A regulator armed by pollution \(P\) is structurally too late. The signal they are watching is a lagging tail of the boom that has already foreclosed recovery.

BDPD³ S2 cascade. Arena A’s pollution chokes B and C; the resource trajectory of B/C is misleading until the cliff arrives.

Win-Win Hides in the Capital

“The right question, asked at the right time, is itself a kind of answer.”

S3 asks: is it the lever (cap, levy, fine) or the trigger signal (pollution \(P\) vs capital \(C\)) that decides whether governance works?

The binding axis is the trigger, not the lever. The same cap-on-\(k_1\) applied when \(P\) crosses a threshold fails — \(P\) is lagging, the signal arrives too late.

The same cap-on-\(k_1\) applied when capital \(C\) crosses a threshold — the leading, visible signal — prevents the cascade and keeps the emitter alive. A genuine win-win.

The fine lever, by contrast, fails as a negative control: there is no demand → production channel in the model, so punishing households cannot slow the producer. Levers are not substitutable without a causal theory of the channel they act on.

BDPD³ S3 win-win. The cap-leading cell preserves A and rescues B/C; the cap-reactive (P-triggered) cell loses all three.

What’s Next

“You can prove anything you want by coldly logical reason — if you pick the proper postulates.” — Isaac Asimov, Foundation

The four-paper arc closes the v1.1 cycle. The directions that follow are scale-up, not new mechanisms:

Campaign-scale LLM sweeps Multi-model (DeepSeek, Claude, Qwen, GLM), multi-temperature, \(N \gg 5\). The mini-pilots flag effects; campaigns will measure them.

Bardi/Seneca polycentric campaigns S0–S3 establish the substrate; the next step is sweeping pollution-link topologies, multi-leader cascades, and treaty designs across the new trigger-signal axis.

Human-subject experiments The card game makes the strategic structure playable. Whether human players reproduce safe betrayal, the compensator trap, and the lagging-signal fallacy is an open empirical question.

Open infrastructure gitlab.com/bdpd/bdpd · AGPL-3.0 (code) + CC BY 4.0 (docs/game) · print-and-play · four preprints

“Only the gods live forever. As for us men, our days are numbered. Our occupations are a breath of wind.” — Epic of Gilgamesh, Tablet III

“The platform produces statistics. The card game produces intuition. The governance question is not which lever — but which signal, and when. The questions are open. We invite others to use it.”

Open-source platform · Print-and-play card game · Four preprints

gitlab.com/bdpd/bdpd · Roberto Brunelli · v1.1 · 2026

Authored by Roberto Brunelli with the support of frontier large language models; model outputs were source-grounded by prompt, validated by extensive triaging with different models and finally integrated under author supervision.

Epigraphs throughout are mood-setting — paraphrased and attribution-light by design (the “Darwin” is Megginson; Ecclesiastes is a paraphrase).