Postnet × Cloudflare — federated NTK-Mirror gate-controller (Phase 40-4b)

Federated training of an NTK-Mirror gate controller — K = 5 000 signed log-gates on the residual streams of a frozen Qwen2.5-0.5B-Instruct. The base model runs FORWARD in your browser tab via onnxruntime-web; only the K = 5 000 gate values are trained, federated via the SPSA tournament from Phase 36-39 (DeComFL fusion + symmetric AIMD adaptive η + post-apply byzantine defense via Phase 40-3's audit loop).

Heavy demo. First Join downloads the ~906 MB int8 Qwen ONNX (hosted on HuggingFace Hub) into your browser's OPFS storage (cached for future visits). Forward passes in WebAssembly take ~1 second each; 2 SPSA trials per round = 6 forwards = ~6 sec/round. Plan for ~10 min per 100 rounds. WebGPU would be 3-5× faster but is gated on browser/driver compatibility.

attack mode (this tab fabricates seeds + lies about Δ → server's Phase 39 byzantine defense should quarantine this worker within ~10 wins; no forward passes, so attack ticks are basically free)

(idle)

worker—

round—

pool—

peers—

test loss—

η—

||θ||—

accept rate—

bandwidth↑0 B ↓0 B

(idle)

Protocol: SPSA tournament (DeComFL fusion, Phase 36+). Each round, this tab runs 2 SPSA trials on a fixed 4-example math corpus; each trial = 3 Qwen-0.5B forward passes with different gate values; submits the trial with the most-negative claimed Δloss as a 20-byte proposal (seed, scalar_g, delta) plus a 4-byte audit_loss_before (Phase 40-3). Server picks the best across all workers, applies θ ← θ − η · g · u(seed) to the K = 5 000 gate values, broadcasts the applied flip.

What this is, in one line: federated training of a real LLM (a tiny part of one — the gate controller, P = 5 000 = ~0.001% of Qwen-0.5B's parameters) across browser tabs of strangers, with byzantine tolerance, ~20-byte wire format, no install, free.

Postnet × Cloudflare — federated Qwen-0.5B gate training