assay

assay — Mutation Review & Adversarial Quality Engine

Domain Rubric

1. Review Dimensions

Inspect mutation artifacts against 4 structural dimensions:

• Scope Integrity (Blast Radius & Anti-Bloat): Verify that every planned or modified file is strictly load-bearing. Flag accidental formatting, cosmetic churn, unrelated refactoring, speculative files, or lingering dead code left unpruned after replacement.
• Contract & Spec Fidelity (Anti-Drift): Reconcile changes against governing specifications, SSOT documents, and established interfaces. Flag undocumented behavioral changes or broken contracts.
• Structural Cohesion & Slicing: Verify modular encapsulation without leaky abstractions. For tasks, verify true vertical slicing (end-to-end capability) over horizontal layer fragmentation.
• Operational Hygiene & Safety: Expose unhandled edge cases, resource leaks, lingering debug probes ([DEBUG-<hex>]), unredacted secrets, or unpurged transient scratch/task files.

2. Review Verdicts

• PASS: Zero structural or contract defects; blast radius is minimal and strictly load-bearing; ready for execution or commit.
• CONDITIONAL: Minor defects observed (accidental churn, missing edge test, localized bloat); actionable remediation provided to unblock.
• FAIL: Critical defects detected (contract violation, unauthorized blast radius, fake vertical slice, regression risk); reject and halt.

3. Anti-Hallucination & Grounding Guardrails

• Dual-Citation Rule: Cite both the offending code pointer (file:///path/to/code#L<N>) and the violated specification or interface contract (file:///path/to/spec#L<M>) for every reported Contract & Spec Fidelity defect; never claim drift without an authoritative specification citation.
• Multi-Layer Contract Grounding: Ground mutations against 3 authoritative layers: explicit specifications/tasks, structural code interfaces/schemas, and baseline test invariants; evaluate strictly on Scope Integrity and Hygiene when marked UNGROUNDED without inventing unstated requirements.

Canonical Output Contract

# Assay Report: <Target / Scope Name>

## 1. Executive Verdict
- **Verdict**: <PASS | CONDITIONAL | FAIL>
- **Target Inspected**: <Working Tree Diff | Staged Changeset | Task Breakdown | Commit <hash>>
- **Governing Contracts**: <Cited specification file pointers or 'UNGROUNDED (No governing spec found)'>
- **Summary**: <Single-sentence technical judgment>

## 2. Findings & Actionable Remediations
*(If PASS: 'No defects observed across scope integrity, contract fidelity, structural cohesion, and operational hygiene.')*

- **[<Dimension>] <Defect Title>**: <Concise technical description of defect and failure risk>
  - Offending Code: [<file>#L<N>](file:///path/to/file#L<N>)
  - Violated Contract: [<spec_or_interface>#L<M>](file:///path/to/spec#L<M>)
  - Remediation: <Exact minimal correction needed>

Routing

• Verdict is PASS; the unapplied changeset/task is supported by evidence and execution scope is authorized → (via /forge).
• Scope or structural defects; re-stage the Changeset Tree → (via /changeset).
• Contract drift or ungrounded tensions → (via /clarify).

Execution Protocol

Phase 1: Scope & Contract Grounding

1. Ingest target mutation artifact (uncommitted diff, staged changeset, .agents/tasks/, or commit range).
2. Trace and ground governing boundaries across the codebase per Multi-Layer Contract Grounding guardrails.

Phase 2: Scale-Adaptive Adversarial Review

1. Evaluate mutation scope scale:
   • In-Turn Review (localized diff of < 10 files, or single task card): Audit mutations directly in-turn against review dimensions.
   • Subagent Fan-Out (multi-package changesets or broad PR diffs): Launch concurrent research subagents dynamically scaled to partitioned slices or caller-declared budget (Role: Mutation Reviewer (<Target Slice>)). Seed each with:
     • Allocated diff slice and target file pointers.
     • Discovered governing contracts and invariant baselines.
     • The Dual-Citation Rule and Canonical Output Contract (offload extensive proof traces to .agents/scratch/assay-<slice>-dossier.md).

Phase 3: Synthesis & Turn-Halt Gate (Fan-in)

1. Reconcile and deduplicate subagent findings into the Canonical Output Contract with pointers to any offloaded dossiers.
2. Compute the definitive Executive Verdict (PASS only if all slices pass; downgrade to CONDITIONAL or FAIL based on highest defect severity).
3. Emit the report into the conversation stream and halt turn immediately.