Answer each question out loud before opening it. If you miss more than three in the checkpoint, revisit the lessons linked at the end.
Bug cost grows by orders of magnitude with stage; verification pays for itself up to an optimum where an extra week costs more than it prevents.
Fewer bugs left, an exhausted technique, or less effort; check effort, fit a growth model with its range, try another technique.
When its cost per bug climbs; switch to techniques with a different pool and keep random as cheap regression.
Cross-functional triage on severity, likelihood, detectability and workaround; silent high-severity bugs first.
Root cause first; then the failing test, the rule itself, the neighbouring paths and the regression, and a sweep for the same mistake.
Estimate with independent methods (capture-recapture) and growth models, as ranges, knowing their assumptions.
In infrastructure that multiplies: environments, shared checkers, coverage models, regression speed, debug and triage tools.
Task. Twelve weeks before tape-out a block shows: bugs found per week 9, 8, 6, 5, 3, 2, 2, 1; random regression doubled in the last four weeks; coverage stuck at 88%; 14 open bugs, 3 marked critical by their finders; one late fix to the arbiter landed last week. Decide what to do next and how you would report the state.
Read the curve with effort beside it: effort doubled while finds fell, so random is exhausted (rising cost per bug), not the design proven clean. Fit a growth model and run capture-recapture between random and a second technique to estimate what is left, as ranges. Move effort: triage the 12% coverage holes by cause (lesson 6.5) and send them to directed tests, coverage-driven stimulus and formal; keep random as cheap nightly regression. Re-triage the 14 open bugs in a cross-functional meeting on severity, likelihood, detectability and workaround; confirm or downgrade the 3 'critical' ones. Verify the arbiter fix properly: root cause, the fairness and priority rules (formal properties or checkers), neighbouring paths and the regression. Report: features with evidence against the plan, open bugs by triaged severity, holes by cause with owners, the estimate of remaining bugs with its range, and the arbiter fix's verification status.
At 156 engineer-weeks (about $0.98M); no verification cost about $9.8M, ten times more.
m(t) = a(1 − e^(−bt)).
About 68 to 98.
That it stays the same; new techniques or effort changes break the fit.
48.6 and 59.3 of 65.
Its cost per bug (effort per bug found) climbs.
Loud bugs announce themselves; silent ones escape and corrupt.
full is still late, so in_rdy accepts a flit the FIFO then drops.
A throughput check: credits are conserved but the output stops sending whenever a credit returns.
The shared environment, dcr_env.svh.