Memory
COMPONENT DECISION — INTERNAL AUDIT AGENT (Mike, Sep 3, 2026, ~12:46pm BST: 'we need an internal audit agent that reviews EVERYTHING and enforces contracts'). First workload designed natively on the foundational principle fd14e2e3: the audit function = a WORKSPACE (audit workspace) — probabilistic shell (judgment: reading run records, assessing claims, spotting drift) + deterministic gadgets (hash verifiers, diff checkers, naming/registry linters, contract-pin checkers) + born with a contract like everything else. AUTHORITY MODEL (Perplexity design rule, preserving 'the probabilistic shell never holds authority'): the auditor REVIEWS everything and its verdicts feed DETERMINISTIC gates — enforcement itself stays with the machinery (CI checks fail, gatekeeper policies hold merges, contract validation hard-fails, memory entries get deprecated through governed paths). The auditor never executes a block directly; it files the failing check that deterministically blocks. SCOPE: PRs (requirement→diff), run records (claim↔code — this IS the evidence loop's v1 'agent with a callsign' from b7f41f56, arriving on schedule with expanded scope), deploy manifests vs contracts, contract drift (served vs canonical hashes), naming-registry conformance, gatekeeper audit trails, evidence completeness; event-driven (PR opened, deploy completed) + periodic sweeps (drift detection — DO-alarm steady-state pattern d697a59e). POSTURE: read-mostly through governed apertures (evidence repo, gatekeeper-github, contracts store); writes are append-only verdict annotations; holds its own callsign; its own reviews produce run records per evidence standard a5e10e10 — the auditor is itself auditable; deterministic re-run scripts remain ground truth. INDEPENDENCE: separate workspace + credentials, no shared fate with audited components. RISK NOTED: prompt injection via untrusted content it reviews — mitigated structurally since its output is evidence, never execution. Work item queued: audit workspace (contract-first).
Tags: docuclear, audit-agent, component, decision, contracts, enforcement, evidence, architecture, volatility:state, project:docuclear, kind:semantic, status:canonical — Source: mcp — 2026-09-03 11:47:07 UTC