Your lab deliverable is an evidence note and a comparison plan, not a claim that three agents improved RTL quality. Use paper role cards or separate local drafts. Work through the single-assistant baseline first without reading the answer. Then perform the spec and RTL passes separately, keeping their initial drafts hidden from each other until the coordinator stage. If one person plays all roles, record that limitation because memory prevents a truly blind second attempt.
Snapshot lab-enable-v1; illustrative RTL excerpt only.
S1: rst_n is active low and sampled on posedge clk.
S2: reset clears q regardless of en; otherwise en loads d.
S3: en=0 without reset 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
Scope: one-bit q and d; no timing, CDC or physical claims.
Required output: claim, S/R references, paper trace, unknowns.Give the baseline the full packet and a single response budget. Give the team the same packet and a total budget that includes all three roles, context repetition and repairs. Either hold total budget constant or explicitly state that budgets differ. The spec role lists obligations. The RTL role traces assignments. The coordinator compares each claim with both sources and writes a disposition. No role changes the implementation. Use the same reporting template for both routes so presentation does not become the hidden advantage.
Before opening the check, produce an acceptance sheet. It should distinguish finding the relevant behavior, citing the right evidence, supplying a reproducible counterexample and avoiding unsupported claims. A human independently traces the conditionals using the packet. If approved engineering tools are available in another environment, any tool check must have its own recorded inputs and actual results. This lab itself requires no simulation, API call, MCP connection or installed dependency.
Write your baseline report and team report, then inspect the answer. Count missed supported findings, unsupported findings, incomplete outputs and human correction minutes separately. Record wall time and all actual resource use if tools were used. For the offline role-play, write 'no model usage measured' and record only your own exercise time. Do not turn a fabricated transcript or a made-up token total into a performance measurement.
The supported candidate is reset blocked by en at R2. With prior q=1, en=0 and rst_n=0 at the edge, S2 requires q=0 but the code makes no assignment and retains 1. Cite S2 and R2-R3. Do not claim an asynchronous-reset requirement, simulation pass or full-design correctness. Score each route against this same answer and retain false alarms. A team wins only if actual quality and effort evidence supports that choice; a tie or a worse result is valid.
For a worked accounting illustration only, suppose a baseline takes 12 human minutes and a team takes 9 drafting minutes plus 7 reconciliation minutes. The team uses 16 human minutes, even if its parallel drafts finish sooner on the clock. These invented values explain bookkeeping and demonstrate no model's performance. In a real comparison, replace them with observed totals, record the tool and prompt versions, and repeat on fresh cases before making a deployment decision.
This tiny design exercises evidence handling rather than realistic verification coverage. Hand your reports to another engineer and ask them to reproduce the paper trace from the frozen inputs. For evaluation background, see NIST AI RMF 1.0, January 2023. Carry the same evidence discipline into the capstone, where missing reports, malformed responses and held-out cases test the whole workflow.