The interviewer listens for: Starts from constraints (bandwidth, latency, number of agents, who builds the IP, ecosystem), not from a favourite protocol; compares at least two options with their costs; mentions verification IP and reuse.

The interviewer listens for: Only when no standard fits a hard constraint; counts the full cost (specification, VIP, compliance, tools, training, every future derivative); prefers profiles or extensions of standards.

The interviewer listens for: Discusses congestion spreading and pause storms against retransmission cost and tail latency; workload-dependent; mentions end-to-end congestion control and isolation of lossless traffic.

The interviewer listens for: Reserved fields ignored by receivers, capability negotiation, defaults off, version fields; tests old against new both ways; notes that version 1's rules decide what is possible.

The interviewer listens for: End-to-end argument; per-link replay as a performance optimisation where error rates are high; end-to-end checks for paths through untrusted or unprotected logic; numbers from a reliability budget.

The interviewer listens for: Little's law with worst-case latency, the cost of tags, buffers and reorder logic, and what happens beyond the limit; notes interconnect ID width.

The interviewer listens for: Agent count and bandwidth; snoop filters as a middle ground; latency of indirection; storage and back-invalidation; workload sharing patterns.

The interviewer listens for: Reuse protocol layers (PCIe, CXL, streaming) over a short-reach physical layer; standards for multi-vendor chiplets; retry even on short links; latency and energy per bit; test and known-good-die.

The interviewer listens for: A dependency graph across channels, message classes and routing, an acyclicity argument, the consumption assumption stated, formal on the critical loops, and simulation with deadlock detection.

The interviewer listens for: Races between ownership changes (writebacks against forwards), message-class separation, snoop filter sizing, ordering at the home, and verification of the protocol tables formally.

The interviewer listens for: Specification as numbered rules, assertions per rule, monitors, reference model and ordering-aware scoreboard, coverage plan, formal targets, independent authorship, silicon debug features.

The interviewer listens for: Gathers evidence first (trace, counters, sticky errors), forms hypotheses tied to liveness rules, reproduces with targeted stress or emulation, and adds observability for next time.

The interviewer listens for: Every request answered, every credit returned, every grant eventually given, link eventually trains or fails; bounds justified from worst cases; formal where bounds are hard.

The interviewer listens for: Access control plus performance isolation: per-requester queues, limits, regulators, QoS with guarantees, timeouts that isolate the requester, and monitoring.

The interviewer listens for: Defines what each side drives and what happens to every piece of shared state; timeouts or wrappers for lost requests; re-initialisation of credits and sequence numbers; tests resets mid-traffic.

The interviewer listens for: Contain them: poison carried to the consumer, error responses, defined escalation when poison cannot be carried, logging with addresses and IDs, and no silent corruption.

The interviewer listens for: State and sticky errors, counters, timestamped trace with triggers, readable when hung; designed with the architecture, budgeted by risk.

The interviewer listens for: Separates protocol efficiency (headers, framing, encoding), flow control (credits versus round trip), concurrency (outstanding limit versus latency), arbitration and bubbles; measures before guessing.

The interviewer listens for: Credit round trip per hop, burst absorption, outstanding limits by Little's law, and the cost of overprovisioning; models before RTL.

The interviewer listens for: Queueing grows sharply near saturation; caps on queue depth, priority for latency-sensitive classes, and admission control; measures percentiles, not means.

The interviewer listens for: Break-even time per state, measured idle distributions, latency tolerance per device, handshake with deny, wake paths for every input, timeouts adapted to traffic.

The interviewer listens for: Hop count (three-hop forwarding), directory lookup, distance; owner and forward states, direct transfers, home placement near users.

The interviewer listens for: When payloads are small and bandwidth is the limit; compares header compression, larger payloads and flit packing against complexity and latency.

The interviewer listens for: Treats the specification as the bug: writes the disputed rule precisely, adds an assertion, updates the document, checks other instances of the same ambiguity.

The interviewer listens for: Numbered checkable rules, stated freedoms and limits, corner cases, examples labelled as examples, assertions delivered with it, versioning rules from version 1.

The interviewer listens for: Question by question, with arithmetic for every size and timer and a dependency argument for deadlock; looks for interactions between sections; ranks comments by severity with fixes.

The interviewer listens for: Any well-argued example showing how a constraint changed (rates, scale, security) and the mechanism that became a burden; shows judgement about longevity.

The interviewer listens for: The fifteen questions and the constraint-first checklist; dissect familiar protocols first; draw handshakes, credits and replay; then read a real specification with the checklist.

The interviewer listens for: Bandwidth efficiency, tail latency, error and replay counters, hang reports, verification bugs by area, and how hard derivatives were to make.

The interviewer listens for: A defensible single question, such as "what happens if the other side stops responding?", with the reasoning for why it exposes the most risk.