The closing questions test whether the candidate can shrink a problem honestly, read an equivalence result, and defend a signoff. Abstraction without a justification is guessing; equivalence without modes is hoping; signoff without a ledger is storytelling. This drill makes the honest versions automatic.
Question | Short answer | Trap | Teaching lesson |
|---|---|---|---|
What is the cone of influence? | Exactly the logic that can affect the property: reduction to it is exact and never over-constrains. Over-constraint comes from assumptions or unsound abstraction, never from the cone. | Blaming the cone for a vacuous pass. | |
What justifies data independence? | No control path from the data values: the block moves words without looking at them. Proved by inspection of the control fan-in or broken explicitly per relevant bit. | Abstracting data the control branches on. | |
Counter too wide to bound? | Replace it with a saturating counter at a small threshold and prove behaviour identical below it, or prove the real counter by induction. The 32-bit timeout becomes 2 bits with a bound of 5. | Bounding below the behaviour and filing it. | |
Memory too large to include? | Track one symbolic address with a shadow value and a valid bit: 41 bits stand in for 8192. Constrain the address to the legal map and cover each region. | An unconstrained address hiding failures. | |
What is a sound cut point? | A signal replaced by a free variable plus a contract proved on the real driver. The contract travels with the proof; replaying the trace on the unabstracted design checks it. | A free cut with no contract. | |
When is symmetry sound? | When requesters are interchangeable in the property and the design treats them uniformly. Prove one, argue the mapping, and mutate a second to check the uniformity. | Assuming symmetry the arbiter does not have. | |
What does a miter prove? | That two implementations agree on every checked input from a common reset, mode by mode. Never that either is correct; the spec proof still has to exist. | Shipping on agreement alone. | |
Retimed logic fails equivalence: setup or real? | First suspect the state mapping: retiming moves registers across logic, so the outputs agree but the internal registers differ. Prove equal outputs with the cut invariant or a larger k, not equal internal state. | Demanding equal internal state instead of equal outputs plus the cut invariant. | |
An ECO touches one mode: what proves it? | Per-mode miters with the constraint listed on each run, plus the intended-difference property pinning the change to its mode. The untouched modes must still prove equal. | One constrained run waiving the rest. | |
What closes a formal signoff? | Proven or bounded-with-justification for every property, reached covers, reviewed assumptions each discharged, cone coverage, mutation results, rerun on the final hash, waivers with reasons. | Not-failed counted as proven. | |
What does formal own in a CDC signoff? | Protocol checks of the crossing with the synchroniser abstracted, not structural analysis. Metastability and structural rules belong to their own checks. | Proving the synchroniser safe in the open flow. | |
What does the CSR proof cover? | Reset values, read-write behaviour, write-1-to-clear with hardware-set priority, read-only registers ignoring writes, and isolation between registers. Access timing matches the wait states: same cycle with none. | Checking data without the access rules. |
Re-diagnose: the cone may still hold the full datapath, or the property may need a helper rather than a smaller model. Abstraction shrinks the machine; only invariants close the step.