Rtl P5s Func Verify Policy

Tier 3 Functional Verification Policy

Testing Tier Context

Tier 1: Smoke Test     — connectivity, R/W, basic ops (rtl-p4-implement Wave 4)
Tier 2: Unit Test      — reference comparison, uarch features (rtl-p4s-unit-test)
Tier 3: Module Regr.   — cocotb multi-seed (THIS POLICY)
Tier 4: Integration    — cross-module, end-to-end (rtl-p5s-integration-test)

Tier transition rules:

• Tier 2 PASS (rtl-p4s-unit-test) → Tier 3 eligible
• Tier 3 PASS or PARTIAL_PASS (this skill) → Tier 4 eligible (rtl-p5s-integration-test) (PARTIAL_PASS = some Critical/High ac_ids PARTIAL; WARNING at Stage 1, escalated to FAIL at Stage 3)
• Tier 3 FAIL → fix via rtl-p4s-bugfix, re-run Tier 2 then Tier 3

cocotb Signal Naming Convention

cocotb test files MUST use correct signal names matching RTL port conventions:

• Signal access: dut.i_data (NOT dut.data_i), dut.o_valid (NOT dut.valid_o)
• Clock: dut.clk (single domain) or dut.sys_clk (multiple domains) — NOT dut.clk_i
• Reset: dut.rst_n (single domain) or dut.sys_rst_n (multiple domains) — NOT dut.rst_ni
• cocotb clock utility: cocotb.clock.Clock(dut.sys_clk, 10, units="ns")
• Reset sequence: drive dut.sys_rst_n.value = 0, wait, then dut.sys_rst_n.value = 1

Multi-Seed Strategy

• Default 5 seeds: 1, 42, 123, 1337, 65536 (configurable via sim/regression/seed_list.txt)
• Each seed tests different random stimulus ordering
• A module passes multi-seed regression only when ALL seeds pass
• Use COCOTB_RESOLVE_X=RANDOM for X-state handling
• Use RANDOM_SEED={seed} for reproducibility

Execution Strategy

• Local-first runtime: default to local execution on current host (--mode local)
• Default parallel budget: use max(1, nproc-2) unless user explicitly overrides --parallel
• AWS usage policy: aws-batch is allowed only when the user explicitly asks to use AWS and explicit gate/runner wiring exists (RTL_ALLOW_AWS=1, RTL_AWS_BATCH_RUNNER)
• Pipelined: as each module's TB completes → immediately launch sim (don't wait for all TBs)
• Module-level parallelism: each module's TB + sim runs as an independent parallel task
• Multi-seed parallelism: queue 5 seeds × N modules; local active workers stay within max(1, nproc-2)
• Incremental coverage: coverage-analyst starts partial analysis on completed modules
• Early termination: >5% failure rate across seeds → halt and report immediately

Coverage Targets

Below target: testbench-dev generates additional tests → re-run regression.

Coverage Collection

# Verilator compilation with coverage
make -C sim/{module} sim SIM=verilator EXTRA_ARGS="--coverage --trace-fst" TOPLEVEL=dut MODULE=test_dut

# Merge multi-seed coverage data
verilator_coverage --write-info merged.info seed_*/coverage.dat

# Coverage HTML report
genhtml sim/coverage/merged.info -o sim/coverage/html/ --title "Regression Coverage"

Regression Scripts

# Automated multi-seed regression ({plugin_root} = plugin root resolved from .rat/state/spawn-context.json)
bash {plugin_root}/skills/rtl-p5s-func-verify/scripts/run_regression.sh \
  --mode local --seeds "1 42 123 1337 65536" --sim verilator

# Optional override when user explicitly asks
bash {plugin_root}/skills/rtl-p5s-func-verify/scripts/run_regression.sh \
  --mode local --parallel "$(($(nproc)-2))" --seeds "1 42 123 1337 65536" --sim verilator

# Coverage merge
bash {plugin_root}/skills/rtl-p5s-func-verify/scripts/merge_coverage.sh \
  --format verilator --output sim/coverage/merged.info

Requirement Traceability Matrix Format

Save to reviews/phase-5-verify/requirement-traceability.md:

AC-Level Format (when structured acceptance_criteria with ac_id exist in iron-requirements)

When iron-requirements.json contains structured acceptance_criteria entries (object arrays with ac_id fields), the RTM uses AC-level granularity:

# Phase 5 Review: Requirement Traceability
- Date: YYYY-MM-DD
- Reviewer: func-verifier
- Upper Spec: iron-requirements.json
- Verdict: PASS | PARTIAL_PASS | FAIL

## Feature Coverage Checklist
| REQ ID | AC ID | Description | Test Case | Status |
|--------|-------|-------------|-----------|--------|

## Findings
### [severity] Finding-N: ...

## Verdict
PASS | PARTIAL_PASS | FAIL: [reason]
- PASS: all Critical/High ac_ids VERIFIED or FORMAL
- PARTIAL_PASS: some Critical/High ac_ids PARTIAL (WARNING at Stage 1, escalated to FAIL at Stage 3)
- FAIL: Critical/High ac_ids UNTESTED or tests failing

AC-level status values per criterion:

• VERIFIED — AC covered by a passing test
• FORMAL — AC proved by formal verification
• PARTIAL — AC partially covered (some test cases pass, scope limited)
• UNTESTED — AC exists but no test covers it
• NOT_VERIFIABLE — AC has verifiable: false (inspection-only); document in RTM, excluded from automated coverage tracking

VERIFIED judgment is made at the individual criterion level when ac_id fields are present.

REQ-Level Format (backward compatible — when no structured AC)

When acceptance_criteria is absent or contains a plain string array (P1/P2 format), the RTM uses the existing REQ-level format:

## Feature Coverage Checklist
| REQ ID | Test Name | Result | Status |
|--------|-----------|--------|--------|

Verdict rules (Tier 3 module-level, aligns with Stage 1 2-tier model):

• PASS — all requirements (or acceptance criteria) verified with passing tests
• PARTIAL_PASS — some Critical/High ac_ids are PARTIAL (not yet VERIFIED/FORMAL). At Stage 1 module graduation: WARNING (proceed). At Stage 3 final audit: escalated to FAIL.
• FAIL — M requirements/criteria UNTESTED, K with failing tests
• Any REQ (or AC) UNTESTED → testbench-dev must generate additional tests
• PARTIAL Critical/High ac_ids → WARNING at this stage, must be resolved before Stage 3

cocotb Ecosystem Quick Reference

• cocotb-bus: Base classes for Driver, Monitor, and Scoreboard
• cocotbext-axi: Ready-to-use AXI4/AXI4-Lite/AXI4-Stream masters and slaves
  • Example: AxiLiteMaster(AxiLiteBus.from_prefix(dut, "s_axi"), dut.sys_clk, dut.sys_rst_n, reset_active_level=False)
  • Stream: AxiStreamSource(AxiStreamBus.from_prefix(dut, "s_axis"), dut.sys_clk, dut.sys_rst_n)
• cocotb-coverage: Functional coverage with @CoverPoint and @CoverCross decorators
  • Example: @CoverPoint("top.data", bins=[range(0,64), range(64,256)])
• TestFactory: Parameterized test generation
  • Example: TestFactory(run_test).add_option("width", [8,16,32]).generate_tests()

See {plugin_root}/skills/rtl-p5s-func-verify/references/cocotb-ecosystem.md for complete API reference.

Tier 2 Baseline Utilization

When Tier 2 unit test results (sim/{module}/{module}_unit_results.json) are available from Phase 4, the CDTG pipeline MUST operate incrementally:

1. Load baseline: Read Tier 2 coverage metrics (line_pct, fsm_pct, toggle_pct) and already-covered features from unit_results.json
2. Prioritize gaps: CDTG Round 1 focuses on uncovered FSM states, untested code paths, and features not exercised in Tier 2
3. Avoid duplication: Do not regenerate test vectors that duplicate Tier 2 coverage. Extend coverage, not repeat it
4. Graceful degradation: If Tier 2 results are absent (e.g., module skipped P4 unit testing), proceed from zero baseline. Log warning but do not block

Coverage targets remain unchanged: Line ≥ 90%, Toggle ≥ 80%, FSM ≥ 70%. The baseline only affects CDTG prioritization, not target thresholds.

Backward Traceability Policy

Test failure reports MUST include requirement impact analysis:

• Every failed test must list its affected req_ids and ac_ids (from # Covers: comments)
• Failed tests without coverage comments are flagged as UNTRACEABLE
• Failure Impact Summary table is mandatory in regression reports
• Priority classification: Critical/High → BLOCKING, Medium/Low → WARNING, unmapped → UNTRACEABLE

This policy ensures that test failures are immediately connected to requirements, enabling prioritized debugging and requirement-level risk assessment.

Escalation & Stop Conditions

• cocotb not installed → halt and route setup through /rtl-agent-team:rat-setup
• Failure persists after 2 RTL fix rounds → escalate to rtl-architect with waveform analysis
• Coverage below 80% after full regression → invoke rtl-p5s-coverage-analyze skill
• cocotb signal name mismatch error → testbench-dev must fix to use i_/o_ convention
• Requirements with NO TEST COVERAGE after additional test generation → escalate to user
• Requirement traceability verdict FAIL with persistent test failures → escalate to rtl-p4-implement

Final Checklist

• All cocotb tests use correct signal names (dut.i_*, dut.o_*, dut.sys_clk, dut.sys_rst_n)
• All test vectors run to completion
• RTL vs ref model comparison done per vector
• Waveform analysis done for all failures
• sim/coverage/coverage.xml generated
• sim/regression/*_result.json written per test
• Multi-seed regression passed (5 seeds per module: 1, 42, 123, 1337, 65536)
• Per-module pipelined execution used (TB → sim without waiting for all TBs)
• Coverage merged across seeds (sim/coverage/merged.info or sim/coverage/coverage.xml)
• Coverage targets met: line ≥ 90%, toggle ≥ 80%, FSM ≥ 70%
• sim/regression/seed_{seed}_results.json written per seed
• Early termination applied if failure rate >5%
• Requirement Traceability Matrix produced (AC-level when structured acceptance_criteria exist, REQ-level otherwise)
• Every REQ-NNN in docs/phase-1-research/iron-requirements.json (canonical) or legacy docs/phase-1-research/requirements.json (fallback) covered by at least one test
• When structured AC exists: every Critical/High ac_id is VERIFIED, FORMAL, or PARTIAL (PARTIAL = PARTIAL_PASS verdict at Stage 1, escalated to FAIL at Stage 3 final audit). UNTESTED = FAIL.
• All covered requirements pass their tests (or failures escalated)
• Traceability verdict is PASS or PARTIAL_PASS (PARTIAL_PASS = WARNING at Stage 1, escalated to FAIL at Stage 3)
• reviews/phase-5-verify/requirement-traceability.md saved