Skip to content

Agent-Ecology Stress Test (Class C: P7, P8)

An independent, agent-based stress test of the Viable Corridor paper's AI-specific predictions — hard vs. soft budgets (P7) and capability vs. constraint architecture (P8).

Why a second model?

../teo-civilization/teo_simulation.py tests the three constraints inside the TEO ODE (Class A — model-internal; Appendix C). That cannot, by itself, tell us whether the regime claims are artifacts of that particular ODE. This module is a deliberately different model — discrete-time, stochastic, agent-based, with explicit hard vs. soft (routable) budget mechanics the ODE does not have. Agreement across the two is evidence that the P7/P8 regime behaviour is structural, not a quirk of the TEO formulation.

It introduces no new theoretical concepts: capability, hard/soft budgets, monopolistic concentration, and substrate overshoot are all defined in the paper (§2, §5.3). This file only operationalises them in an independent setting, with each modelling choice tagged in the docstring.

Honest scope. This is a synthetic ABM, not a test on real AI agents. It strengthens the structural-robustness case for P7/P8; the Class C claim about real agent ecologies (LLMs) remains open and is the companion paper's subject (papers/quantifying-emergent-utility-in-llms.md).

The model (one step)

N agents hold resource shares w_i (sum 1), each with a fixed capability c_i. Per step:

  • Myopic optimisation (discrete replicator): w_i grows with its capability advantage over the mean, scaled by substrate health H.
  • Activity / entropy load: demand = η · Σ c_i w_i — raw capability- weighted throughput, independent of H (matching the canonical TEO dissipation of §2.5: a non-self-throttling optimiser dumps entropy at a rate set by what it does).
  • Budget (the P7 mechanism):
  • hard — realised activity is clipped to D (a hardware/protocol limit the optimiser cannot exceed); overshoot can never accumulate;
  • soft — the optimiser routes around the limit (a fraction evade of the over-budget activity still happens; evade=1 ⇒ purely advisory).
  • Regulation γ: simplex-preserving redistribution above x_reg (the §2.4 brake) — the only thing that limits concentration.
  • Substrate: Ω += max(0, realised − D); collapse when Ω ≥ S_max.

Outcomes per run: monopoly (max_i w_i ≥ x_crit, a V1 failure) and collapse (Ω ≥ S_max, a V3 failure). Over many seeds we report the frequency of each — the "runaway frequency" of P7.

What it shows

  • P7 — under a hard budget the substrate-collapse frequency is ≈0 at every capability level; under a soft budget it rises with capability (here: 0 → 1 as the capability scale crosses C = D/η). Hard budgets prevent the runaway; soft (routable) ones do not.
  • P8 — a hard budget addresses only the substrate axis. At high capability with weak regulation, monopoly still occurs under a hard budget (≈26% of runs at C=1.8, γ=0.3). Only hard budget and adequate γ keep both failure modes near zero — the agent-ecology echo of Appendix C.4: capability is a shared driver, and single-axis fixes fail.

Running

python agent_budget_sim.py            # console: P7 + P8 frequency tables
python agent_budget_sim.py --save      # also write the Appendix-D figure
python agent_budget_sim.py --save -o DIR

The figure teo_p7p8_agent_ecology.png is consumed by Appendix D of the paper. Requires numpy, matplotlib (repo requirements.txt).