research-workflow

Go/no-go checkpoint before an expensive or irreversible run (high-impact-checkpoint).

Inspect or enable the research-workflow hook decision log (RWF_HOOK_DEBUG).

Reference-parity audit against an external reference, loading the matching domain lens.

Capture a reproducibility contract for the current work — env lock, seeds, precision, inputs, commit.

Multi-lens scientific code & figure review of a changeset, running the Review-cluster skills (correctness, numerics, JAX, robustness, craft, figures).

Use when you have a result you are about to trust — to red-team it against confirmation bias and the stable-but-wrong failure mode (numerical artifact, latent bug, boundary effect, or a mundane alternative explanation that fits the same data). Produces the strongest attacks plus the cheapest discriminating test for each. Don't use for reviewing CODE (→ scientific-code-reviewer and the Review cluster), the neutral close-out format (→ verification-gate), reporting the result's uncertainty budget (→ uncertainty-reporting-gate), or explaining a convergence/refinement-floor behavior (→ numerical-method-validation).

Use when relying on code, math, or facts the AI assistant itself produced — apply EXTRA scrutiny precisely because the model's signature failure is confident fabrication: hallucinated library APIs, plausible-but-wrong algebra, invented constants/citations, and tests written to pass rather than to catch. The stance that points generic skepticism at the assistant's own output. Don't use as the concrete check itself — route to the specific gate: API existence (→ verify against docs), formulas (→ derivation-before-implementation), constants/citations (→ provenance-of-constants), result size (→ plausibility-envelope), test/claim integrity (→ evidence-first-execution).

Use when a research session produces meaningful results and you need durable manifests, payloads, plot scripts, and a completion note so later sessions can reason from artifacts instead of memory. Don't use for the in-the-moment command discipline (→ evidence-first-execution), the go/no-go close-out (→ verification-gate), or pinning the runtime environment itself (→ reproducible-environment-contract).

Use when a result or model rests on simplifying assumptions, approximations, fixed parameters, or regime-of-validity choices — keep an explicit running ledger of what each result depends on, so when an assumption later breaks you know exactly which conclusions die with it. Don't use for citing the source of a value (→ provenance-of-constants), recording a decision and its rationale (→ decision-log-and-commits), or quantifying the numeric error a kept assumption induces (→ uncertainty-reporting-gate).

Use when verifying performance claims, characterizing scaling, comparing implementations, or validating against analytic solutions for performance-critical kernels (integrators, force solvers, renderers). Don't use for small analysis scripts (overkill) or for designing a test plan before code exists (→ testing-strategist).

12 hooks across 4 events