| name | architectural-spike-audit |
| description | Audit an architectural-spike@1 against architectural_spike_rubric@1.0: enforces structural completeness (SPK-STRUCT - frontmatter, five sections, recommendation names exactly one probed option), checkable evidence on every option (SPK-EVID - repo path, command+output, or URL; confidence cross-checked against evidence depth), timebox discipline (SPK-BOX - plan and actual recorded, >1.5x overrun carries a recorded operator HALT verdict), and a non-empty discard log when options were rejected (SPK-DISC). Emits a `score / 10` verdict; refuses to pass on <10/10. Use when user asks to "audit this architectural spike" or "check the architectural spike". Do NOT use for "draft a new architectural spike" (use architectural-spike-author instead). |
| license | Apache-2.0 |
| metadata | {"version":"1.0.0","module":"skill","stage":"d","cyberos-template":"architectural-spike@1","cyberos-rubric-target":"architectural_spike_rubric@1.0"} |
| allowed_memory_scopes | {"read":["project:*","module:*"],"write":["project:task/{task_id}/architectural-spike-audit"]} |
| audit | {"row_kind":"architectural_spike_audited","required_fields":["spike_id","task_id","verdict","score","findings"]} |
| inputs | [{"name":"architectural_spike","format":"architectural-spike@1","required":true}] |
| outputs | [{"name":"spike_audit","format":"verdict: pass|fail|needs_human + score /10 + findings[]"}] |
| triggers | ["an architectural-spike@1 artefact exists and has not passed audit"] |
| blockers | ["artefact is not architectural-spike@1 (unknown version) - needs_human, never guess"] |
| untrusted_inputs | {"wrap_in_marker":"untrusted_content","injection_scan":"required","on_marker_hit":"surface_to_human"} |
architectural-spike-audit
1. Purpose
Make spike verdicts reproducible: an auditor cites SPK rule ids instead of paraphrasing prose. Only 10/10 passes; evidence is checked by RESOLUTION (does the citation actually check out at audit time), not by presence.
2. Verdict semantics
pass = every rubric rule green (10/10). fail = any rule red, findings name each rule id + location + what resolves it. needs_human = ambiguity the rubric cannot decide (unknown artefact version, contradictory frontmatter, an operator-verdict question).
See RUBRIC.md for the rule families, AUDIT_LOOP.md for the iteration protocol, and REPORT_FORMAT.md for the report shape.
