0005. Remove the ai-compliance-auditor agent; Álvaro owns §6 compliance directly
Date: 2026-07-13 Status: accepted
Context
ADR-0001 set up an ai-compliance-auditor agent to audit Álvaro's AI-usage
practices against the TFG regulation (§6) — declaration completeness, authorship
balance, data protection, thesis consistency, and attribution — and mandated
running it before each milestone release.
In practice, §6 compliance is a judgement the author must make and defend
personally: the regulation makes Álvaro responsible for reviewing all AI output
and for the truthfulness of the AI-usage declaration. Delegating that judgement
to an agent risks the exact failure mode §6 guards against — treating a
compliance verdict as something the AI produces rather than something the author
owns. The audit trail the agent would read (docs/ai-usage/, Co-Authored-By
trailers, ADRs) already exists and is maintained independently.
Decision
We will remove the ai-compliance-auditor agent entirely. §6 AI-usage
compliance is Álvaro's direct responsibility: he reviews declaration
completeness, authorship evidence, data protection, thesis consistency, and
attribution himself, using the existing audit trail and the ai-declaration
skill. No agent is run for compliance.
- Delete
.claude/agents/ai-compliance-auditor.mdand every instruction to run it (CLAUDE.md,docs/roadmap.md,docs/development-plan.md,doc-curator). - The pre-release gate keeps only the
codebase-sanityagent (code quality). - Unchanged: the §6 obligations themselves — the audit trail, the
Co-Authored-By: Claudetrailer, theai-declarationskill, the quarantineddata/private/, and the data-protection rules all stand.
This supersedes the ai-compliance-auditor element of ADR-0001's AI-governance
decision; the rest of ADR-0001 stands.
Alternatives considered
- Keep the agent, run it before releases (ADR-0001). Rejected by Álvaro: a compliance verdict is the author's to make; an agent producing it blurs the authorship line §6 exists to protect.
- Keep the agent but make it advisory/on-demand only. Rejected: leaves a standing tool whose output could be mistaken for the compliance sign-off, for no benefit over the author reading the same audit trail directly.
Consequences
- Clearer authorship: the §6 compliance judgement is unambiguously Álvaro's, which is itself the strongest evidence of human authorship.
- One less agent to maintain; the pre-release ritual is just
codebase-sanity. - Accepted: no automated pre-flight for declaration/attribution gaps — Álvaro performs that review manually (the audit trail makes it a quick pass).