Series Coverage Matrix
Table 1 maps each platform feature to the paper that describes it in detail. The matrix is intended as a reading guide: a reader interested in a specific mechanism can locate the paper that presents it and the one that stress-tests it.
| Feature | BDPD\(^0\) | BDPD\(^1\) | BDPD\(^2\) | BDPD\(^3\) |
|---|---|---|---|---|
| Logistic engine (\(S\), \(r\), \(K\), \(\theta\)) | described | used | used | — |
| Seneca engine (\(R\), \(C\), \(P\)) | described | — | — | exercised |
| Forest of Humbaba (card game) | described | — | cross-checked (V5) | — |
| Observability noise | described | — | calibrated (V6) | — |
| Perturbation surface (capacity cap, regen nudge) | described | — | — | — |
| World / nested arenas | — | — | described (V1) | — |
| Resource links | — | — | described | — |
| Pollution links | — | — | — | described |
| Pact registry (harvest cap, commons floor) | — | described | used | — |
| Cheap-talk channel (broadcast, announce, propose) | — | described (D1) | — | — |
| Governance metrics (cooperation, lie, silent def.) | — | described | — | — |
| Sanction perturbation (flat) | — | described (C1) | — | — |
| Graduated sanction ladder | — | described (D2, D3) | — | — |
| Arena-level meta-agent (sanctioner) | — | — | exercised (V2) | — |
| World-level meta-agent (treaty enforcer) | — | — | exercised (V3) | — |
| Exclude perturbation (junk arena) | — | — | exercised (V4) | — |
| Treaty (harvest cap per arena per round) | — | — | exercised (V3) | — |
| Capital shock / wealth shock | — | — | — | exercised |
| Pollution-aware regulator (leading signal) | — | — | — | described |
| OFAT robustness sweep | — | — | — | exercised |
| Built-in heuristic agents (conservative, aggressive) | described | — | used | used |
| LLM agents (DeepSeek-flash, Agent SDK) | — | exercised (D1–D3) | exercised (V2–V4) | — |
| Seed-robustness protocol (\(N = 5\)) | — | used | described (V6) | used |
Gaps not covered by the current series:
- RCP agent (built-in Seneca heuristic): exercised in BDPD\(^3\) pilot scripts but not described in a dedicated methods section.
- Agent SDK internals (
@tooldecorator, tool-call loop,on_tool_callhook): implementation documented in the repository; no paper section describes the SDK architecture. - Multi-model LLM comparison: all LLM results use a single model family (DeepSeek-flash). Model-pool diversification is deferred to campaign-scale work.