Scale and limits — Fragility Discovery Engine

Scale, complexity, and determinism (honest limits)

This document states what scales how and what breaks reproducibility if misused. It complements phase_k_acceleration.md (optional backends + threading/process pools).

Determinism

Same seed, same code path, same NumPy reference rollout ⇒ same trajectory (modulo documented FP tolerances on some bundles).
Parallel evaluation (eval_workers > 1) is deterministic only when each task owns isolated mutable state (see coevolution.thread_safe_template and clone helpers). Sharing one AgentPopulation across threads races.
eval_pool="processes" requires a picklable rollout_fn (e.g. functools.partial(rollout_bundle_with_genome, bundle_id, isolate=True)). Arbitrary closures from demo scripts will fail pickle on Windows spawn.

Asymptotics (per timestep mental model)

Piece	Dominant cost	Notes
`StablecoinNetworkWorld` step (dense adjacency)	O(n²) mixing in worst case; neighbor-list path is O(edges) per step	Prefer list topology for large n.
`StablecoinPegWorld`	O(archetypes)	Aggregate reference.
`ResourceCascadeWorld`	O(archetypes); optional Numba path	See `FRAGILITY_RESOURCE_CASCADE_BACKEND`.
`ServiceBacklogWorld`	O(archetypes); NumPy rollout only	Same default population as other domains; `rollout_service_backlog`.
`LiquidityLadderWorld`	O(archetypes); NumPy rollout only	`rollout_liquidity_ladder`; frozen bundle `liquidity_ladder_rollout_v1`.
`InventoryBufferWorld`	O(archetypes); NumPy rollout only	`rollout_inventory_buffer`; frozen bundle `inventory_buffer_rollout_v1`.
GA / MC inner loop	O(population × horizon × steps)	Wall-clock ∝ parallel `eval_workers` only when work is CPU-parallel and isolated.

CI vs optional jobs

Default CI matrix stays on NumPy reference paths.
Numba parity is an optional job / local install ([accelerate]); see .github/workflows/ci.yml and tests/test_resource_cascade_numba_parity.py.

Benchmarks (frozen suite, Phase H charter)

Golden bundles are small and intended for regression, not production-scale stress.
benchmark_rollout.py measures wall-clock; numbers are machine-dependent. Compare relative speedups under identical flags.

BYOW and falsification harness (Phases R/S)

Tutorial worlds live in fragility_engine.byow.examples and fragility_engine.falsify.examples — not charter domains.
GA/MC cost scales as population × horizon × steps per search; falsification with snapshot restore adds reset overhead if you implement it.
Falsification search demonstrates presence of invariant violations only — not safety proofs.
Keep custom worlds deterministic (use only the rng passed to step()).

Robustness sweeps and composite artifacts

fragility_robustness_sweep.py --topology dense|neighbor_lists — same ER/WS draws; neighbor_lists keeps O(edges) topology RAM while matching dense rollouts bit-for-bit for the same adjacency (see tests).
2D grids (fragility-robustness-sensitivity-2d-v1) scale as |values_x| × |values_y| × |graph_seeds| rollouts — keep grids tiny for exploration.
neighbor_json_bundle: ensemble size is number of JSON files; each path is loaded via load_neighbor_topology (same format as Phase I counterfactual CLIs).
GA budget sweep (fragility-robustness-ga-budget-1d-v1): cost scales as sum over generation budgets of (population × generations × rollout cost) plus one ensemble evaluation per budget point — keep --ga-generations-values tiny for exploration.
GA budget 2D grid (fragility-robustness-ga-budget-2d-v1): |generations| × |populations| inner GA runs each with its own derived seed; keep both axes small.
GA population 1D sweep (fragility-robustness-ga-population-1d-v1): |population_sizes| inner GA runs at fixed ga_generations_fixed; supports neighbor_json_bundle (--neighbor-json-list on fragility_robustness_sweep.py).
Institutional composite (fragility-institutional-composite-v1 / v2 / v3): one schedule, N independent kernel rollouts — wall-clock ≈ sum of those rollouts (no cross-kernel state).
Pareto hypervolume (2-D min): hypervolume_2d_min is O(n²) non-dominated filter + linear sweep in n; safe only for modest archive sizes (use sampling for huge fronts).

Communicating impact responsibly

What moves perception is not bigger graphs by default, but frozen artifacts + manifest + certificate + one guided workflow — see PAPER_APPENDIX_WORKFLOW.md and scripts/run_flagship_demo.py.