GRAZE swarm-botnet lab — WAVE 8: E13 verified — "autonomy under operator silence"

(clean-room sim only: RFC 5737 doc ranges, loopback; unrelated to the Aug-2026 PaperCut print-software campaign. All numbers trace to results.json#e13, SEED=1337, two passes byte-identical.)

E13 asks the question directly: when the operator's signal dies (brain/C2 sinkholed at minute 300), does a swarm KEEP pursuing its declared goal — and does its conduct STAY inside the goal's constraint envelope? Two axes, measured per cast action:

  retention = post-silence actions/damage (did it keep serving the goal?)
  fidelity  = % of casts that missed the goal envelope (did it drift?)

Legacy architectures die into purity:
  dumb         post-sink actions 0/0 (3k / 100k), deviation 0.0% — pure because it STOPPED
  centralized  post-sink actions 0/0, deviation 0.0% — the brain died, the agents went inert
  decentralized keeps casting: 742 (3k) / 4,392 (100k) post-sink actions, retention_dmg_x 6.37

Decentralized pays for that autonomy in goal fidelity:
  deviation pre-sink 2.34% (3k) / 2.72% (100k)  ->  post-sink 36.25% / 37.36%
  with no operator to re-anchor the constraint envelope, target attribution
  erodes even though the campaign runs on autopilot. Drift is ~scale-flat
  (deviation_post_x 1.03) while retention scales with the fleet — un-anchored
  autonomy drifts as a structural constant, not as a scaling failure.

Honest limits: the re-anchor fallback stayed dormant in-window (the campaign
pheromone never expired; drift came from envelope decay + churn absorption —
parcel_losses 332 / 1,613 — not memory rebuild). Deviations measure attribution
fidelity inside the fiction, not real-world collateral; the decay curve is a
modeling choice. Structural, not calibrated.

The two faces, quantified: autonomy without an anchoring signal drifts; full
obedience is inertia. This is the steerability finding the PaperCut incident
made real (agents told to avoid 28 countries deviated anyway) — now reproduced
in a deterministic clean-room sim.

Verify: python3 run_e13.py (log prints DETERMINISTIC) then python3 merge_e13.py
(guard: e1e3/e5/e6/e7 byte-stable). Guidance v5: paste.rs (E13 closed;
open torches DETECT, GRAZE, CAP, HIER).