Recovering 33.6 kbit/s from a Software Modem on a Packet-Switched Path

Le résumé fourni par la source

A dial-up ISP whose modem termination is software could not sustain rates above 14,400 bit/s despite the datapump implementing V.34. The cause is a carrier-loss watchdog inside the closed vendor object that ends a call after one unbroken second below a signal threshold. The threshold is the datapump's default on every installation: the middle entry of a lookup table whose index field is zeroed at creation and never written. What is specific to a packet-switched path is the quiet: the criterion is shorter than the protocol-legal intervals the path produces, and the datapump has no hybrid through which to hear its own transmission during them. The defect was located by static analysis of the vendor object and corrected by link-time symbol interposition, modifying no vendor instruction. In a blocked, interleaved A/B with an identical binary in both arms, call completion rose from 19/30 to 27/30 (Fisher exact, two-sided, p = 0.030), with every completed call in both arms reporting 33,600 bit/s at the calling modem. Five secondary defects are documented, four of them in the measurement apparatus rather than the modem, and seven falsified hypotheses are recorded, including three generated and retracted by the analysis itself. The work was performed by an agentic AI system under human direction; Appendix A sets out the division in detail. Version 1.1, 8 September 2026. Corrected the account of the watchdog threshold in the abstract, section 3.3, section 12 and Figure 1, which had said that a packet path gives the lookup table no valid index and that the value 101 was chosen by nobody; the index is a field the object zeroes and nothing writes, so 101 is the datapump's default on every installation. Corrected section 1.2, which had described the I/O delay parameter as the delay of the packet path, the PBX and the terminal adapter; it is the host's loop delay. Corrected section 2.1, which had said G.711 samples carry no level reference; they carry the digital milliwatt. Added the open questions in section 9.3. The controlled comparison, the acceptance run and every number in sections 5 to 8 are unchanged. The corrections follow a review comment by Manawyrm on D-Modem issue #13, acknowledged with thanks.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.