Postnet-CF: Federated LLM gate-controller training across browser tabs with verified byzantine defense

Postnet-CF is a federated learning protocol whose workers are unmodified browser tabs and whose coordinator is a single Cloudflare Durable Object. Updates are single-scalar SPSA estimates (DeComFL-style): each proposal carries 20 bytes of information (round, seed, scalar_g, claimed delta, audit_loss_before — five 32-bit fields) independent of model size, and the measured broadcast-only downlink stays under 400 B per tick at any parameter count up to 1.5e9. The reference implementation transports this payload as JSON over HTTP for compatibility with Cloudflare Workers. The per-round learning rate is adapted by a symmetric AIMD rule, giving faster loss descent than fixed-eta on a 50K-parameter head-classifier task.

The same protocol is then applied to federated gate-controller training on a frozen Qwen2.5-0.5B-Instruct: K=5000 signed log-gates on the residual streams (NTK-Mirror parameterisation) are trained across browser tabs running ONNX-Runtime-Web, with the base model forward-only and only the K-gate vector evolving over time. A post-apply byzantine defense — a no-self-audit cross-rotation queue plus a magnitude-lie test — is validated across 5 random seeds: in 5 of 5 runs an attacker that fabricates claimed_delta = -10 is quarantined on the first proposal following its 10th flagged win, at an empirical fraud-detection rate of 1.000 +/- 0.000, while the gate vector descends from a baseline loss of 1.7632 to 1.7627 +/- 0.0003 (multi-seed) and to 1.7570 in a single-seed extended run (R = 183). The complete system runs on the Cloudflare free tier; a public deployment is available at https://postnet-cf.abgunaydin94.workers.dev .

This deposit archives the source and the working-paper draft (docs/PAPER_DRAFT.md plus the rendered PDF and matplotlib-generated figures under docs/figures/). Every reported number is traceable to a committed empirical writeup under docs/PHASE_*.md and to the verifier scripts under scripts/.