Two reviewers declare a reset defect and a third declares the design correct. A majority vote feels decisive, but all three may have inferred the same missing requirement. A coordinator's job is to connect claims to evidence, not to count confident voices. Separate the requirement, the implementation observation and the inference linking them. This structure makes disagreements small enough for an engineer to resolve by reading the actual supplied material.
Synthetic fixture; line labels are not executable code.
S1: At a rising edge, rst_n=0 clears q, regardless of en.
S2: Otherwise en=1 captures d; en=0 holds q.
R1: always_ff @(posedge clk) begin
R2: if (en) begin
R3: if (!rst_n) q <= 1'b0;
R4: else q <= d;
R5: end
R6: end
Synthetic candidate A: "Asynchronous reset missing; add negedge rst_n."
Synthetic candidate B: "Reset may be blocked by enable; inspect R2-R3."First normalize these into separate claims. Candidate A asserts an asynchronous requirement; candidate B asserts that enable can block a required synchronous action. Ask each reviewer for its requirement citation and a minimal distinguishing input. Do not merge them into 'reset issue' because that would lose the different assumptions and possible remedies. The coordinator records supported candidates, rejected claims and unresolved questions with reasons. Rejected claims matter when you later count false alarms.
A counterexample is a short hypothetical input history that would violate the requirement if the proposed interpretation is correct. It is not a simulation transcript. Record the prior state, reset and enable at the sampling edge, and the expected and inferred next states. Another engineer should be able to trace the conditionals manually. If an approved simulator is later used, keep its actual version, stimulus, output and scope separate from this paper trace.
For the fixture, decide which candidate has support and produce one edge with a specific prior q value. Explain whether the coordinator should recommend a source change, declare signoff or ask a human to verify a candidate. Also decide what to do if a third reviewer repeats candidate A word for word. Keep the answer focused on the given requirements; do not add reset-release, CDC or timing assumptions that the packet does not provide.
B survives as a candidate: with prior q=1, rst_n=0 and en=0 at a rising edge, S1 requires q=0, while R2 skips all assignments and q retains 1. A lacks support because S1 requires a sampled synchronous reset, not an asynchronous one. Repetition of A adds no evidence. Send B and its paper trace for human verification; do not edit RTL or declare signoff. This is reasoning from the excerpt, not a simulated result.
This method transfers to timing and physical reports. If one reviewer says negative slack proves a setup problem while another says the report is a hold check, inspect the report header and convention before discussing a fix. Different assumptions can produce superficially conflicting yet internally consistent interpretations. When the input cannot distinguish them, the coordinator should publish the missing information as an explicit question instead of choosing the more familiar explanation.
Independent verification uses the source and an appropriate check outside the draft's own assertions. Another model's agreement is still model output, especially when it saw the first answer. A human paper trace is useful for this small example but does not replace broader RTL verification. NIST AI RMF 1.0, January 2023 provides risk-management background; the adjudication procedure here is a teaching workflow.