rtl-model-consistency

<Use_When>

• All three models exist (refc/, bfm/, rtl/) and need consistency validation.
• Checking for drift between models after independent updates.
• A pre-regression gate is needed to confirm baseline consistency.
• A discrepancy between two models requires the third to arbitrate. </Use_When>

<Do_Not_Use_When>

• Only two models exist → use rtl-p5s-func-verify (RTL vs ref) or rtl-p5s-perf-verify (RTL vs BFM).
• Models are known to be out of sync — fix the diverging model first, then re-run.
• Only functional RTL verification is needed → use rtl-p5s-func-verify. </Do_Not_Use_When>

<Why_This_Exists> When ref model, BFM, and RTL are developed independently, silent divergence accumulates. 3-way comparison isolates which model is wrong: ref == BFM != RTL means RTL has a bug; ref != BFM == RTL means the ref model diverged; ref == RTL != BFM means BFM has an issue. Two-model comparisons cannot make this determination. </Why_This_Exists>

Prerequisites

• refc/build/ref_model built and executable.
• bfm/build/bfm_smoke built and executable.
• RTL files present under rtl/ with simulation infrastructure (iverilog or cocotb).
• Test vectors at sim/consistency/test_vectors.bin (or will be generated — minimum 10 vectors).

If any model binary is missing: WARNING — run that model's build skill first; proceed with 2-way comparison and note the absent model in the report.

<Responsibility_Boundary>

• Scripts (compare_3way.py) handle deterministic pairwise binary comparison and first-divergence extraction.
• LLM handles diagnosis (which model is the likely source of error), report narrative, and escalation decisions.
• Contract surface: all three models must run on identical input vectors; bitexact is the default criterion unless tolerance is documented. </Responsibility_Boundary>