5  Open Roads

The four vignettes converge on a single methodological lesson: governance findings on this substrate are jointly determined by institutional design and agent architecture. Rule-bound agents produce clean, flat baselines; signal-adaptive agents decode the institution and respond in both directions. Neither class alone tells the full story, and any campaign-scale follow-up must run both or explicitly state which baseline it rests on. Extending the social-ecological-systems robustness frame of Anderies et al. (2004) — originally cast in terms of links between resource users, infrastructure providers, and physical infrastructure — we read the agent architecture as a further such component whose fit-to-task shapes how the system responds to disturbance, making the choice of agent class a first-class governance-design question, not a methodological detail.

From this baseline, three campaign-scale directions emerge — each licenced by the nested substrate and beyond the reach of any single-arena framework (Dorsch and Flachsland 2017; Carlisle and Gruby 2019). We sketch each as a structured card: question, substrate, and risk.

5.1 Road A — Governance of the invisible

Motivation from vignettes. V3 showed that a world-level enforcer remains load-bearing against adaptive LLM agents — but the hazard it monitors (aggregate harvest) is visible. BDPD\(^3\) (Brunelli 2026b) showed that when pollution is the trigger signal, the regulator is structurally too late: the damage is committed before the symptom clears any detection threshold. Road A requires moving from the logistic heuristics of V1–V4 to the Seneca ODE substrate, where the problem is no longer only agent adaptivity but the speed of collapse relative to governance sampling rate — the rate-induced boundary studied in BDPD\(^3\) and characterised more broadly in Ritchie et al. (2023), with real-world analogues in cross-boundary cascading nutrient pollution under polycentric governance (Ahlström and Cornell 2017; Cash et al. 2006). Road A asks whether the same nested substrate can do better by routing hidden pollution state into the enforcer directly.

  • Question. When environmental harm propagates between arenas through a hidden state variable (\(P\), the pollution coordinate of the Seneca engine), can a meta-agent who sees \(P\) sanction the emitter early enough to prevent a downstream cascade?
  • Substrate. Link.type = 'pollution' routing hidden state from one arena into the engine of downstream arenas (already shipped in v1.1 and exercised by BDPD\(^3\) (Brunelli 2026b) in the single-arena Seneca setting; Road A would consume it for the first time in a polycentric configuration). Built-in for single-arena Seneca replica; mixed built-in + LLM for 3-arena polycentric cascade; LLM-only for the governed case with a P-aware enforcer. Budget ~$5–10.
  • Risk. The cascade might be too fast to sanction usefully or too slow to detect in a 30-turn pilot — both are parameter-tuning costs.

5.2 Road B — Behavioural contagion under exclusion

Motivation from vignettes. V4 found that conformists join the defector in the junk arena — but whether this is passive noise or active contagion is unresolved at \(N = 5\). Road B designs the experiment that separates them: if hiding the defector’s harvest record restores junk-purity, contagion is the mechanism; if it does not, passive spillover suffices.

  • Question. V4’s junk purity \(0.6 \pm 0.3\) is compatible with two causal stories: passive spillover (each cooperator defects independently) and active contagion (cooperators update on the defector’s visible defection, a candidate mechanism for which is the normative conformity documented in dyadic LLM judgement tasks (Bito et al. 2026) and the recommender-driven polarisation observed in opinion-dynamics LLM simulations (Bojić et al. 2025) — both adjacent rather than identical settings). Under which conditions does contagion take over — and does breaking the social channel (hiding the defector’s harvest record) restore type-purity?
  • Substrate. The exclude perturbation from V4, scaled to ~\(N = 50\) seeds across a grid of pact strictness, sanction visibility, and junk arena treatment. LLM throughout. Budget ~$10–15.
  • Risk. Contagion may be a small-sample artefact: the spread \(\pm 0.3\) on 5 seeds is wide enough that the underlying mean could sit anywhere from 0.3 to 1.0.

5.3 Road C — Treaty enforcement at scale

Motivation from vignettes. V3’s LLM defector probed the enforcement boundary 2.4 times over 30 turns. Whether that probing is bounded (the defector eventually accepts the cap once the cumulative cost exceeds the harvesting gain) or unbounded (it keeps probing indefinitely) determines whether treaty enforcement is a viable long-run institution or a cost that grows without limit.

  • Question. Is the LLM enforcer-multiplier bounded (the LLM eventually accepts the cap) or unbounded (it keeps probing until the commons or the enforcer’s budget runs out)?
  • Substrate. The treaty enforcer from V3, scaled to \(T_{\max} = 100\) and parameterised defector aggression. Three treaty shapes (per-arena, total-world, per-round vs per-game). LLM defector + LLM cooperators; built-in enforcer. Budget ~$50–100.
  • Risk. Longest individual runs of the three roads; narrative closest to standard mechanism design, less distinctive than A or B.

5.4 How to choose

Axis A — Hidden monitoring B — Contagion C — Enforcement
New platform work minor (new P-aware enforcer) none none
Novelty high high medium
Budget ~$5–10 ~$10–15 ~$50–100
Risk of null result medium high low

A useful next step is a mini-triage: one Seneca cascade to time the cliff window, one long-horizon V4 run to bound the contagion fraction, one \(T = 100\) V3 run to test whether the multiplier settles. Cost ~$3–5 and one development day, after which the choice between A, B, and C rests on firmer ground.

5.5 Open question for the series

Whether the governance inversion documented here — rule-bound agents flat, signal-adaptive agents responsive — generalises to the Bardi/Seneca substrate (BDPD\(^3\), where the regulator is currently deterministic) or to single-arena communication dynamics (BDPD\(^1\), where the architecture effect dominates talk) is the cross-cutting question that a campaign-scale follow-up across the three substrates would address.

Methodological Coda: LLMs as Synthetic Collaborators

Two distinct roles of LLMs in this work should not be conflated. Role A (epistemic/generative): LLMs participated in conceptual development, implementation, and manuscript preparation — this is the methodological contribution acknowledged in the transparency statement. Role B (experimental): LLM agents (DeepSeek-v4-flash) serve as generative subjects in the V2–V4 LLM cells — this is the empirical content reported above. The workflow follows the human-AI collaborative methodology described in BDPD\(^0\) (Brunelli 2026a, §Methodological Coda): multiple frontier LLMs with different training distributions (Anthropic’s Claude, DeepSeek, Alibaba’s Qwen, Z.ai’s GLM-5.1) were employed for cross-validation of analysis, framing, and manuscript revisions, under the author’s direction.

Data and Code Availability

The BDPD simulation platform, experiment definitions, and analysis scripts are available at gitlab.com/bdpd/bdpd. The raw pilot artefacts are in data/pilot/ of the repository. Per-vignette reproduction commands, the agent prompt template, the LLM configuration knobs, the aggregate.json schema, and known data caveats (including the V2 LLM aggregator skip-seed bug noted in Appendix D) live in docs/platform/reproducibility.md of the repository — the canonical reference for replicating any pilot in this paper or BDPD\(^1\).