Issue: #630. State: proposal for falsification, no field proof.
The synthetic stress and failover-frontier code provides reproducible simulation evidence, not measured production topology. Product-level adaptive routing remains blocked by admitted-path telemetry in #649.
Freeze simulation configuration and seeds before comparing Physarum against static shortest-path, discounted Thompson and explicit multipath/failover. Report success rate, recovery after a shift, mean burden per successful task, failure-attribution noise sensitivity and confidence/dispersion per regime. Run stationary, gradual, transient, node-loss, correlated failures, both graph densities and multiple parameter settings. Compare at matched burden to avoid claiming resilience gains paid for with hidden resource cost.
If the Pareto frontier of an adaptive baseline dominates the proposed policy across stationary and shift fixtures, do not integrate. If results depend on one arbitrary conductance floor or one random seed, report sensitivity instead of a general benefit. If an available deployment is flat/single-hop, there is no legitimate multi-path topology and this routing policy stays research only.
Collect real already-admitted nodes, explicit edge meaning, path latency and failure attribution before designing any canonical topology snapshot. Every future shadow recommendation must bind to an exact admitted-pool digest and freeze before outcome; it never grants eligibility, spend or merge authority.
Next evidence: run fixed-seed replication, publish code/config digests and record negative results. None of those runs are claimed by this file.
Run from the repository root with a pinned Python environment and source revision, and retain full stdout JSON before interpreting the results:
git rev-parse HEAD
mkdir -p results # results/ is gitignored scratch output
python sim/physarum_compute_routing_stress.py \
--seed-start 1 --seeds 20 --epochs 80 --tasks-per-epoch 20 \
--topology all > results/physarum-routing-stress.json
python sim/physarum_compute_routing_stress.py \
--seed-start 1 --seeds 20 --epochs 80 --tasks-per-epoch 20 \
--topology all --failover-frontier \
> results/physarum-routing-failover-frontier.json
python sim/physarum_compute_routing_stress.py \
--seed-start 1 --seeds 20 --epochs 80 --tasks-per-epoch 20 \
--topology dense --environment abrupt-shift --parameter-sweep \
> results/physarum-routing-parameter-sweep.json
sha256sum results/physarum-routing-*.json
These commands are proposed reproduction steps, not experiments performed by this PR. The frontier command compares both topologies, all available regimes and the configured failover retry-probability sweep; the parameter command deliberately selects a single topology and environment, as required by its CLI parser. Use the exact same seed cohort, environment and topology when comparing strategies.
For every condition record: exact code commit; Python version; all CLI arguments; output digest; synthetic topology; admitted-edge assumption; number of independent seeds; per-policy successful task fraction; mean burden per attempted and per successful task; paired burden-matched frontier gap; dispersion/uncertainty; cases dominated by discounted Thompson or failover; and the reason a hypothesis was retained or rejected.
Stop condition: do not recommend even shadow deployment if a realistic strong baseline dominates at matched burden, if results only hold in one hand-picked regime/parameter setting, or if source-provenance and paired seed reproducibility are missing. A synthetic win can only nominate a future field study, never establish resource admission or product benefit.