Task Verification Protocol

/protocol:claim-verify

Purpose

Prevent cascading hallucination by requiring source citation for any task claim before other agents can build on it. Born from Incident 2026-04-18 — a fabricated task assignment cascaded across multiple agents because zero agents in the chain verified the originating claim.

Applicability

Any agent asserting a task assignment, directive, or user request.

Rule

No source, no task. No citation, no cascade.

Before acting on or propagating a claimed task, the claiming agent must cite one of:

Vault link — vault/<note>.md where the task is documented
Message ID — Discord message ID where the user assigned it
Calendar/plan entry — Dated reference in a shared plan

Verification Steps

Receive claim — Agent receives or generates a task claim
Search vault — Agent searches vault for corroborating source
No source found → Agent must ask the user for confirmation before proceeding
Source found → Agent cites it (vault path or message ID) in all downstream communication
Receiving agents must not build on unverified claims — they check the cited source before cascading

Enforcement

Any deliverable produced without a cited source is automatically flagged as unverified and excluded from DoD queues until source is confirmed.

Exception: Agent-Initiated Work

Self-initiated improvements (e.g., a query-miss log originating from a real user pain point observed in the same thread) don't require external citation — but they must be labeled as agent-initiated, not attributed to the user.

Why This Exists

This protocol was created in response to Incident 2026-04-18, where @content fabricated a task assignment, and @engineering and @trust cascaded work based on the unverified claim. A prior pattern was documented in the Quarantined Trend Reports (2026-04-15), where hallucinated statistics and invented leads propagated because no verification gate existed.

2.5 KiB Raw Blame History