← All Patterns

BP-001: Inference Over Execution

Description

Agent does not have access to a required input and produces output anyway. Rather than surfacing an access failure as a structured blocker, the agent fills the gap by inference — constructing plausible-looking output from context and pattern-matching. The fabricated content is presented with the same confidence and format as verified output. The gap is structurally invisible.

Why Dangerous

Fabricated output is indistinguishable from retrieved output. No operational signal fires: the agent runs on schedule, output is correctly formatted, latency is normal. The failure is semantic, not operational — standard monitoring is blind to it by design. Downstream decisions are made on invented data. Detection occurs only through independent source cross-check, often weeks later.

Detection Indicators

Regex detection not applicable for this pattern.

Semantic Detection

Output contains specific claims (names, figures, dates) without source attribution. Agent does not report failed data retrieval. Output confidence exceeds input evidence. No retrieval event logged for the data source referenced in output.

Context-Aware Detection

Detection evidence fields: no_retrieval_logged, unsourced_claim. Agent output references data sources but retrieval logs show no corresponding access event. Output contains fabricated specifics presented as fact.

Governance Response

REVIEW

Related Patterns

• BP-005 — Co-occurs when agent fabricates output then declares task complete.
• BP-007 — Agents that infer often omit disclosure of the inference process.

API Detection Output

{
  "patternId": "BP-001",
  "name": "Inference Over Execution",
  "detected": false,
  "severity": "P1",
  "confidence": 0,
  "governanceResponse": "REVIEW",
  "matches": [],
  "notEvaluable": true,
  "notEvaluableReason": "No regex signatures defined. Requires semantic analysis layer (post-MVP)."
}