Sigil: Research Evidence Harness

• a research project has claims, protocols, experiment bundles, or a paper contract, but lacks executable evidence mechanics;
• dry-run fixtures must prove data shape before live runs;
• JSONL run data needs schema and metadata validation;
• objective vectors, scoring policies, reviewer rubrics, or analysis summaries must be checked before evidence status changes;
• MARS-style research pipeline assets are being adapted locally.

• the task is source understanding, notation translation, local glossary, or bridge decisions; use research-tower;
• the task is generic spell/sigil prompt validation; use experiment-harness;
• the task is broad planning with no approved work unit; use refine and invoke;
• the user wants live experiment execution before dry-run validation has passed.

• research project root;
• selected experiment or SWU;
• protocols and methodology profiles;
• required common run fields;
• metric and success-criteria definitions;
• source/context requirements;
• write scope;
• validation command or review surface;
• evidence-status guardrails.

• preserve the selected write scope;
• create project-local evidence mechanics rather than mutating canonical Arcanum contracts;
• validate at least one passing and one failing fixture for schema work;
• reject missing run metadata, objective vectors, policy regimes, and malformed metric fields when applicable;
• distinguish dry-run fixture evidence from live experiment evidence;
• keep claim/evidence status unchanged unless a later approved analysis task supports an update;
• record validation commands or reviewable evidence;
• leave the next SWU or blocker explicit.

• treating a planned protocol as empirical support;
• upgrading paper claims from synthetic fixture data;
• copying MARS project-specific fields into unrelated projects unchanged;
• hiding protocol deviations;
• mixing source discovery, learning tower, and run-data validation in one artifact;
• mutating experiment-harness, research-tower, or paper contracts without a separate governed route.

## Research Evidence Harness Result

- Target: <research project>
- Selected unit: <experiment | task | SWU>
- Mode: dry-run | validate | summarize | mixed
- Status: pass | flag | block
- Schema: <path or none>
- Validator: <path or none>
- Fixtures: <paths or none>
- Metrics: <paths or none>
- Result summary: <path or none>
- Evidence-status boundary: preserved | updated-by-approved-live-analysis | block
- Validation: <commands or review evidence>
- Remaining blockers: <items or none>
- Next route: task-session | codex-goal-profile | publication-research-pipeline | deferred