Experiments

Each experiment answers one question and feeds one number (or verdict) into the Link Contract. Method: emulated links first (netem/tc — controlled loss, delay, reordering, capacity), then confirmation on real ad hoc WiFi.

#	Question	Method	Output
E1	At what residual loss rate / reordering depth does mesh-QUIC throughput collapse?	Sweep loss/reorder on an emulated link under a C1 stack; measure goodput knee	R2 thresholds ("residual loss < X%, reordering < Y")
E2	What is the minimum link MTU at which the C1 stack works?	Shrink emulated MTU under mesh-QUIC (handshake + data), with and without link-level fragmentation	R4 MTU floor
E3	How much ACK-path delay does QUIC tolerate on a half-duplex/TDMA medium? The return path is gated until the medium turns around (up to a frame epoch): RTT inflates and jitters, ACKs arrive compressed in bursts; past some gap, PTO fires spuriously and the congestion window collapses. Where is that gap?	Alternating-gate emulation; sweep turnaround/epoch duration; watch RTT estimator, spurious retransmits, goodput vs full-duplex baseline	Max turnaround/epoch bound — the delay-variance term of R2, and an input to scheduled-MAC frame design (R9)
E4	What is the operational floor — the minimum link bandwidth and maximum delay at which Florete stays practically usable (link mTLS establishment, agent sync, useful goodput)?	Sweep a rate × RTT grid on an emulated link; measure handshake time, sync time, steady goodput	Operational floor for the contract; R3 handshake sizing falls out
E5	Which telemetry fields are both necessary and obtainable (R5)?	Top-down: enumerate what TCE/PCE/RRA computations actually consume. Bottom-up: instrument commodity WiFi (nl80211: RSSI, MCS/rate, retries, queue) and check coverage	R5 mandatory minimal set
E6	Can neighbor discovery run purely in software over R1's datagram service?	Beacon mTLS connection initiations over ad hoc WiFi broadcast; alternate initiator/responder roles on a schedule; sweep beacon rate	Discovery latency vs airtime overhead; verdict on the R6 software baseline

Results are recorded here per-experiment as they land (setup, raw numbers, conclusion), and the contract is updated in the same change.