Review native evidence
Read <project>/reports/flow_engine.json from the normal public runner. Use
the selected rows' current position and name from flow/steps.json. The
Router-selected window executes through the sole engine path; this review
adds no gate, alternate runner, receipt, qualification or approval layer.
Review each native result
- For each selected row, read
tool_ran, native_steps, native_metrics, metrics, and the native paths in reports. Confirm the tool actually consumed the intended design and corner.
- Open the underlying tool report. Compare its measurement and operator with the unchanged original spec, PDK or owner limit. Check that a summary did not hide a failed corner, missing input or unexecuted measurement.
- AI review the result's engineering meaning and physical plausibility. A token, file, stub field or printed PASS is not proof that the intended behavior was exercised.
- Match simulation transcripts and waveforms to the actual stimulus and observed response. Check that a testbench can distinguish the required behavior rather than printing PASS unconditionally.
- Follow source facts to the actual consumer. Confirm that citations refer to real source meaning, assertions express the applicable requirement, and bit captures came from the measured simulation or hardware.
- Compare related native reports for inconsistent inputs, design revisions, corners or numbers. Investigate substantive disagreements through the responsible producer or tool.
Keep failures as failures. A missing native value, missing original limit or
unexecuted selected tool is NOT_MEASURED, never an inferred PASS. Describe
what remains unmeasured and why. Preserve analog, mixed-signal and FPGA evidence
where applicable, and preserve all original numeric engineering limits.
Review a proposed exception against the real defect, mitigation and residual
risk. Its text or metadata cannot erase an engineering failure or convert
deferred work into an executed PASS. Follow the owner's current instructions
for handling that remaining work.
Resolve and record
Trace each finding to the first incorrect producer or consumer. Fix the
design, local native-tool fork or plugin as the current task permits, then
rerun the actual affected tool and review its native result. Do not add another
per-step checker or qualify the tool during an ordinary design run.
Keep the existing per-row AI review and measured duration with the result.
Append any useful narrative to <project>/reports/compliance_spot_check.md,
identifying the current table position/name, native report, actual value,
original limit, finding and remaining FAIL or NOT_MEASURED obligation. Keep
run evidence outside the plugin repository. Use only current authorization
for publishing a report, messaging others, creating an issue or landing a fix.
Skill: flow signoff judgment — applicability, references and fixes the expert owns
When to apply: deciding whether an optional flow row applies, choosing the
reference a native result is compared against, or turning a physical FAIL into a fix.
The IC expert decides these from the design's own evidence; the owner is asked only
about spending, external or irreversible actions and host security.
- Applicability when the documents are silent. Decide from the design's nature
and record the reasoning as the expert's ruling, never as a document citation.
A single-die, purely digital benchmark block with one supply needs no functional
safety, power intent, multi-die, hard-macro, system-scenario, security,
DFT/scan/ATPG or package-model rows. Those rows become applicable only when an
input states the feature or the delivery requires it.
- Post-layout SPICE correlation. Compare SPICE against the STA delay with the
declared OCV derate removed. The derate is margin, not model error. Keep the
tolerance derived from the Liberty NLDM grid around each stage's operating point;
never substitute a fixed percentage.
- Crosstalk glitch. Simulate the victim driver held quiet with every coupled
aggressor switching. Use transistor-level drivers and receivers at the slow, hot,
low-voltage corner with maximum RC. Compare the peak with the receiver's VIL/VIH.
A failure is fixed in the design, for example by upsizing the victim driver with
an incremental-route ECO; the threshold never moves. An ECO creates a new carrier:
re-measure every post-route row on it. Rows whose inputs are unchanged and proven
byte-identical may be carried over.
- Electromigration. Use the PDK's native power-grid currents and judge them
against the foundry DRM average-current tables. Vias are judged per cut, with the
cut count recovered from the branch resistance; metal uses the drawn width. RMS is
judged only when the PDK defines an RMS rule.
- Latch-up. When the PDK ships no latch-up deck, measure the DRM latch-up rules
geometrically on the GDS: well and substrate tap distance, and IO guard rings
around marked diffusion. Pick the device class (low or medium voltage) from the
thick-oxide marker. Zero structural findings from a name-based screen is not a
latch-up measurement.
- Final physical verification. Run LVS on the extraction of the final post-fill
GDS, not the routed DEF. Run every native step the row lists, including ERC.
- Final fill-aware extraction. Supply special nets do not belong in a signal
SPEF. Bind them by name or exclude them, as RCX does. Never resolve them by a
flood selection through the whole power grid and fill.
- Power comparisons. Compare like with like. A non-pad group sum and a
whole-design total that includes the pads are different quantities.
- Delta-sigma / oversampled ADC ENOB.
- Separate the three ceilings: distortion (integrator swing driving HD2/HD3),
thermal noise (kT/C at the sampling capacitor), and the ideal quantization limit
computed at the measured input centre and offset.
- Same-input point pairs that decode differently are simulator trajectory
divergence. Choose integration settings that make each pair agree within the
ideal quantization residual before any 128-point run is trusted.
- An input offset consumes full-scale headroom and raises distortion. Fix the
offset before resizing for resolution.
1---2name: compliance-gate-spot-check3description: "AI review of actual native tool evidence against original engineering limits, with explicit failures and unmeasured obligations."4tier: verification5paired_program: flow_engine.py6metadata:7 source: https://skillmd.com/skills/vibeic/compliance-gate-spot-check8---910# Review native evidence1112Read `<project>/reports/flow_engine.json` from the normal public runner. Use13the selected rows' current position and `name` from `flow/steps.json`. The14Router-selected window executes through the sole engine path; this review15adds no gate, alternate runner, receipt, qualification or approval layer.1617## Review each native result18191. For each selected row, read `tool_ran`, `native_steps`, `native_metrics`, `metrics`, and the native paths in `reports`. Confirm the tool actually consumed the intended design and corner.202. Open the underlying tool report. Compare its measurement and operator with the unchanged original spec, PDK or owner limit. Check that a summary did not hide a failed corner, missing input or unexecuted measurement.213. AI review the result's engineering meaning and physical plausibility. A token, file, stub field or printed PASS is not proof that the intended behavior was exercised.224. Match simulation transcripts and waveforms to the actual stimulus and observed response. Check that a testbench can distinguish the required behavior rather than printing PASS unconditionally.235. Follow source facts to the actual consumer. Confirm that citations refer to real source meaning, assertions express the applicable requirement, and bit captures came from the measured simulation or hardware.246. Compare related native reports for inconsistent inputs, design revisions, corners or numbers. Investigate substantive disagreements through the responsible producer or tool.2526Keep failures as failures. A missing native value, missing original limit or27unexecuted selected tool is `NOT_MEASURED`, never an inferred PASS. Describe28what remains unmeasured and why. Preserve analog, mixed-signal and FPGA evidence29where applicable, and preserve all original numeric engineering limits.3031Review a proposed exception against the real defect, mitigation and residual32risk. Its text or metadata cannot erase an engineering failure or convert33deferred work into an executed PASS. Follow the owner's current instructions34for handling that remaining work.3536## Resolve and record3738Trace each finding to the first incorrect producer or consumer. Fix the39design, local native-tool fork or plugin as the current task permits, then40rerun the actual affected tool and review its native result. Do not add another41per-step checker or qualify the tool during an ordinary design run.4243Keep the existing per-row AI review and measured duration with the result.44Append any useful narrative to `<project>/reports/compliance_spot_check.md`,45identifying the current table position/name, native report, actual value,46original limit, finding and remaining `FAIL` or `NOT_MEASURED` obligation. Keep47run evidence outside the plugin repository. Use only current authorization48for publishing a report, messaging others, creating an issue or landing a fix.4950### Skill: flow signoff judgment — applicability, references and fixes the expert owns5152**When to apply:** deciding whether an optional flow row applies, choosing the53reference a native result is compared against, or turning a physical FAIL into a fix.54The IC expert decides these from the design's own evidence; the owner is asked only55about spending, external or irreversible actions and host security.5657- **Applicability when the documents are silent.** Decide from the design's nature58 and record the reasoning as the expert's ruling, never as a document citation.59 A single-die, purely digital benchmark block with one supply needs no functional60 safety, power intent, multi-die, hard-macro, system-scenario, security,61 DFT/scan/ATPG or package-model rows. Those rows become applicable only when an62 input states the feature or the delivery requires it.63- **Post-layout SPICE correlation.** Compare SPICE against the STA delay with the64 declared OCV derate removed. The derate is margin, not model error. Keep the65 tolerance derived from the Liberty NLDM grid around each stage's operating point;66 never substitute a fixed percentage.67- **Crosstalk glitch.** Simulate the victim driver held quiet with every coupled68 aggressor switching. Use transistor-level drivers and receivers at the slow, hot,69 low-voltage corner with maximum RC. Compare the peak with the receiver's VIL/VIH.70 A failure is fixed in the design, for example by upsizing the victim driver with71 an incremental-route ECO; the threshold never moves. An ECO creates a new carrier:72 re-measure every post-route row on it. Rows whose inputs are unchanged and proven73 byte-identical may be carried over.74- **Electromigration.** Use the PDK's native power-grid currents and judge them75 against the foundry DRM average-current tables. Vias are judged per cut, with the76 cut count recovered from the branch resistance; metal uses the drawn width. RMS is77 judged only when the PDK defines an RMS rule.78- **Latch-up.** When the PDK ships no latch-up deck, measure the DRM latch-up rules79 geometrically on the GDS: well and substrate tap distance, and IO guard rings80 around marked diffusion. Pick the device class (low or medium voltage) from the81 thick-oxide marker. Zero structural findings from a name-based screen is not a82 latch-up measurement.83- **Final physical verification.** Run LVS on the extraction of the final post-fill84 GDS, not the routed DEF. Run every native step the row lists, including ERC.85- **Final fill-aware extraction.** Supply special nets do not belong in a signal86 SPEF. Bind them by name or exclude them, as RCX does. Never resolve them by a87 flood selection through the whole power grid and fill.88- **Power comparisons.** Compare like with like. A non-pad group sum and a89 whole-design total that includes the pads are different quantities.90- **Delta-sigma / oversampled ADC ENOB.**91 - Separate the three ceilings: distortion (integrator swing driving HD2/HD3),92 thermal noise (kT/C at the sampling capacitor), and the ideal quantization limit93 computed at the measured input centre and offset.94 - Same-input point pairs that decode differently are simulator trajectory95 divergence. Choose integration settings that make each pair agree within the96 ideal quantization residual before any 128-point run is trusted.97 - An input offset consumes full-scale headroom and raises distortion. Fix the98 offset before resizing for resolution.