By name: xic-raw-validation description: Use this before launching or accepting XIC Extractor RAW-backed validation when the task involves 8RAW/85RAW, alignment validation, production-equivalent gates, benchmark acceptance, timing heartbeat, or expensive RAW I/O. Do not use it for pure unit tests, synthetic no-RAW checks, or reading existing artifacts unless the result is being used as gate evidence.
XIC RAW Validation
Use this skill before launching or accepting RAW-backed validation for XIC Extractor.
This is the execution checklist for the repo's expensive validation workflow. It points to canonical parameters instead of duplicating them.
Trigger Threshold
Use this skill when:
- launching, rerunning, or accepting 8RAW / 85RAW validation;
- validating alignment, benchmark, timing, heartbeat, or production-equivalent outputs;
- a RAW-backed command may run longer than a normal focused test;
- existing artifacts are being used as gate evidence.
Do not use this skill for:
- unit tests or synthetic no-RAW tests;
- quick code import checks;
- reading a validation note for background only;
- small diagnostics that do not touch RAW files or gate interpretation.
Required Reading
Read first:
AGENTS.mddocs/agent-parameter-settings.mddocs/agent-subagent-routing.md- active validation spec, plan, output index, or note
Search existing artifacts before rerunning:
tools/diagnostics/INDEX.md- relevant
docs/superpowers/notes/ - task-specific
output/directories
Preflight
Before launching a RAW run, state:
- decision the run can close;
- sample set: synthetic, 8RAW, 85RAW, targeted benchmark, or manual EIC;
- documented Python runner and Thermo DLL/RAW paths;
- foreground command shape;
- output level and expected artifacts;
- heartbeat or timing sidecars;
- timeout or stop condition;
- why existing artifacts cannot answer the question.
For alignment validation, prefer --output-level validation-minimal unless the
task explicitly needs human review/debug reports.
Do not launch 85RAW via background Start-Process from the Codex shell unless
the user explicitly approves an external-terminal or automation path.
Reviewer Trigger
Use validation-evidence-reviewer for:
- any 85RAW run;
- likely >30 minute RAW run;
- production-equivalent validation;
- benchmark or gate acceptance that affects product readiness.
Use mode preflight before launch and mode acceptance, science, or
performance after the run as appropriate.
Acceptance Labels
Keep these separate:
run_ok: command completed and emitted expected artifacts.gate_ok: artifacts satisfy the stated gate.production_ready: evidence supports production behavior.inconclusive: result cannot close the decision; name missing evidence or exit rule.
Preflight output is not validation. A timing heartbeat is not correctness. Diagnostic sidecars are not downstream delivery unless the active contract says so.
Output Surface Rules
For large alignment validation, the primary machine delivery surface is
alignment_matrix.tsv. Targeted benchmark diagnostics also need
alignment_review.tsv and alignment_cells.tsv.
Do not produce .xlsx, HTML, owner-edge, status-matrix, event-owner, or
ambiguous-owner outputs for large validation unless a human review or debug task
explicitly needs them.
Stop Conditions
Stop and inspect instead of relaunching when:
- heartbeat stalls or exceeds expected runtime;
- preflight fails or checks a weaker contract than the real command;
- output artifacts are missing or unexpectedly huge;
- the run cannot change the next action;
- the command uses an undocumented runner/path combination;
- PowerShell, sandbox, approval, or writable-root uncertainty appears.
When sandbox or command-shape uncertainty is the issue, run:
python -m scripts.agent_sandbox_doctor --command "<command>"
The doctor is diagnostic only; it does not execute the command or replace user approval.