| name | verification |
| description | Crosscut verification conductor — closes the oracle gap for scientific
/ numerical software. Owns ground-truth V&V, NIST StRD fixtures,
benchmark scenarios, and DAG validation across the three streams.
Absorbs ground-truth, nist-strd-runner, benchmark-scenario-generator,
and dag-validator. Use when validating against a gold standard,
closing the oracle gap, auditing benchmark trustworthiness, running
NIST StRD harnesses, generating realistic benchmark scenarios, or
checking the Rust crate DAG for cycles. Composes with
superpowers:verification-before-completion. Serena-first for code.
|
| license | MIT |
verification
The crosscut V&V skill. Activates from any stream when the question is
"is the answer right?" — and stops a cycle from closing on the
three-pillar RIGOR axis if the oracle is biased or missing.
Anchors
CLAUDE.md sections (read before acting):
- Benchmark Backend Comparison Fairness — spectrafit is the subject;
lmfit and jax are independent cross-verification oracles.
- After a benchmark run — what to read from
manifest.json and how
to report regressions / accuracy / speed.
Hooks that will fire:
protect-nist-fixtures.sh — BLOCKING (exit 2): guards
tests/fixtures/nist_strd/. Only update via the absorbed nist
workflow + a fresh itl.nist.gov fetch.
pre-merge-dag.sh — pre-merge gate that the crate DAG has no cycles
and no invalid cross-dependencies.
pre-merge-perf-baseline.sh — pre-merge gate that perf hasn't
regressed against the recorded baseline.
enforce-perf-accuracy.sh — perf/accuracy contract enforcement when
present.
Serena first
For DAG validation and source-scan checks:
mcp__serena__find_symbol Cargo.toml → manifest discovery is path,
not symbol — grep is ok here
mcp__serena__find_referencing_symbols → for trace-impact of a
removed model
For code-touching verification work (writing oracles, harnesses),
serena drives.
Decision: which sub-document?
| Subject | Reference |
|---|
| Manufactured solutions, metamorphic relations, significant-digit checks, differential tests, uncertainty quantification, claim-to-evidence map | references/ground-truth.md |
NIST StRD fixtures, certified-data fitting harness, tests/fixtures/nist_strd/ | references/nist.md |
| Generating realistic benchmark scenarios (YAML), outlier injection, multi-solver comparisons | references/scenario-gen.md |
| Rust crate dependency DAG validation, cycle detection, visualization | references/dag.md |
Composes with
- The stream that owns the code under test (
crates-stream,
python-stream, or web-stream).
superpowers:verification-before-completion — never claim "done"
without an oracle reading.
superpowers:systematic-debugging when a verification failure is
confusing (the oracle disagrees but the path isn't obvious).
- If stuck:
andon-loop/references/stuck-mode.md.
Three-pillar reporting
Verification owns the RIGOR pillar. On any cycle close that touched
science / numerics, this skill writes a one-line RIGOR statement:
RIGOR: <oracle exercised> · max|Δr²| = X · UQ coverage = Y%
How verification activates in tri-stream mode
In andon-loop tri-stream, the parent loop spawns a verification
crosscut subagent after the per-stream sub-loops report green but
before the cycle is marked converged. The subagent runs:
- The NIST StRD harness (if any test path under
tests/fixtures/nist_strd/ was touched in this cycle).
- The DAG validator (if any
Cargo.toml was edited).
- The benchmark gate (
uv run spc-bench gate) — geomean speedup +
max |Δr²| against the baseline solver.
- The ground-truth claim-to-evidence map if a manuscript-adjacent
artifact was touched (e.g.
docs/superpowers/*.md).
Any failure halts the cycle — the andon rule applies across the
crosscut as well.