Same Quality, Cheaper and Fairer: Multi-Agent Reinforcement Learning for Joint Bitrate and Origin Selection in Hybrid CDN-P2P Swarms

We report a complete, adversarially audited study of per-node multi-agent reinforcement learning (MARL) for the joint decision of video quality and chunk origin in hybrid CDN-P2P swarms, carried from an event-driven simulator to a physical platform of six node processes serving real video. The headline is a parity result with a price tag, and because it is not the same arm in both regimes we state it one arm at a time. Under origin scarcity - the regime the masked policy was trained and deployed in - one arm carries the whole conjunction on hardware, over ten seed-paired scenarios, against a threshold heuristic re-tuned inside that regime: mean QoE is a statistical tie (-1.37 +-1.32 and -0.16 +-1.08 for the two declared calibrations), origin traffic falls by 26-29 %, the worst-served user rises from -11.71 to -3.12 5th-percentile QoE and Jain fairness from 0.82 to 0.93. With an abundant origin - where the deployed arm is the unmasked predecessor - the same weights never deliver quality and saving together: read by argmax they buy QoE parity (-0.76 +-1.01, n.s.) and no saving (-1.2 % +-1.3, n.s.); read by sampling they buy 19.0 % +-1.5 less origin traffic (10/10 seeds) at a real cost of -3.68 +-0.66 QoE (0/10). Only fairness holds at both operating points there - Jain +0.116, 10/10, under either decision rule, which differ by 0.000 (n.s.); the worst-user gain is argmax's alone (5th-percentile QoE +13.37 +-2.04 against +7.25 +-3.43 for the sampled rule). Mean QoE never won outside the simulator: on the platform it is a statistical tie in every campaign, and we measure a mechanism consistent with that (real delivery micro-dynamics inflate the player buffer and let a buffer-based ABR climb to rungs the simulator does not concede; its signature appears in three independent campaigns and is amplified, not created, by physical scarcity). We locate and measure that mechanism; we do not claim to have proven it causally. Our main technical finding concerns the action space. The wall this method hit at four times the largest trained swarm was neither missing information nor insufficient optimisation but expressibility: with two independent action heads, "ask the peer who actually has the chunk" is a relational constraint between heads and is literally inexpressible, so for the operating point that actually uses P2P, 51-69 % of its P2P decisions had no viable candidate at all. Conditioning the origin head on the already-sampled quality level moves the N=200 gap from -24.30 to -1.58 QoE without adding a single parameter - while the policy keeps choosing peers at chance. A zero-training ablation completes that picture and costs us the word in our own title: in the simulator, amputating the learned origin head and letting a threshold rule choose among the same viable peers improves mean QoE by 0.68 to 3.39 points at every swarm size inside the trained range - a composite we never took to hardware. We also report the negative scaling result plainly: victory inside the trained range (N=20 and N=50, the first project claim to survive a three-replica between-seed interval), parity at 2x and at 4x that range (the latter with a negative point estimate), and a pre-registered project gate that does not pass. We release the simulator, the platform, 209 tests, and a 95-row truth table that records every claim the project made, its verdict and the number that decides it - including the seven claims we retracted, one of them in this paper's own review.