Executive Brief
A research agent queries a corpus and returns a claim. A synthesis agent receives that claim, treats it as verified, and builds a summary on top of it. A decision-support agent receives the summary and presents a recommendation with high confidence to a human approver. At no point did any agent verify the original claim. At no point was the original uncertainty preserved. By the time the claim reached the human, it had been laundered through three layers of processing until it appeared indistinguishable from a verified fact.
This is not an edge case. It is the default behavior of every multi-agent pipeline that does not explicitly design for the opposite.
The problem is structural, not model-level. Individual agent hallucination rates have been studied extensively, see Rawte et al. [1] and Huang et al. [2]. But the compounding dynamics that emerge when agents hand outputs to each other without provenance metadata have received almost no practitioner-facing treatment. The NIST AI Risk Management Framework [3] addresses trustworthy AI at a system level but does not specify inter-agent provenance requirements. AutoGen [4] and similar orchestration frameworks give agents communication channels without epistemic guardrails on what travels through them.
The cost of inaction is asymmetric. In a single-agent system, a hallucination is visible at the output layer and can be caught by a human reviewer. In a six-agent pipeline, the same hallucination arrives at the output layer dressed in three layers of confident synthesis. The human reviewer sees a polished, well-sourced-sounding recommendation with no signal that the foundation is sand.
Who owns this problem: Chief AI Officers and CTO organizations building agentic pilots. Legal and compliance teams approving AI-assisted decisions. Any team deploying a pipeline where one agent's output feeds another agent's context window.
Why now: Multi-agent architectures have moved from research to enterprise pilot deployments at scale. CAMEL [5], AutoGen [4], and ReAct [6] patterns are appearing in production pilots across financial services, healthcare, and legal. The window to design epistemic integrity into these pipelines is before they are embedded in critical workflows, not after the first high-visibility failure.
The Three Coined Constructs
Definition: The process by which an unverified or uncertain claim produced by one AI agent is consumed and re-stated by a downstream agent with increased confidence, stripping the original uncertainty from the lineage. The claim has been "laundered" through an intermediate processing step that removes the provenance signal.
Formal analog: Like financial money laundering, the mechanism does not alter the content of the claim; it alters the traceability of its origin. A claim that originated as a 60% confident inference is indistinguishable, at the output layer, from a claim derived directly from a verified source, once it has passed through one or more intermediate agents.
Why the term matters: "Hallucination propagation" describes the content problem. Epistemic Laundering names the architectural mechanism: the stripping of provenance. These are distinct failure modes requiring distinct mitigations. A practitioner who conflates them will address the wrong layer of the stack. This term originates with this work and is subject to the license stated in the footer.
Definition: The accumulated epistemic risk in a multi-agent system that arises from agent boundaries where outputs are passed downstream without provenance metadata. Each boundary that strips uncertainty adds to the debt. The debt is invisible in any single interaction and compounding across pipeline depth and breadth.
Formal analog: Technical debt accrues when engineering shortcuts defer structural problems. Trust Propagation Debt accrues when epistemic shortcuts defer verification. Like technical debt, it is cheap to address early (at design time) and expensive to address late (after incidents have occurred). Unlike technical debt, it does not show up in any code review, test suite, or linting tool unless specifically instrumented.
Measurement proxy: The debt of a pipeline segment can be estimated as: depth x (1 - provenance coverage rate) where provenance coverage is the fraction of inter-agent handoffs that carry uncertainty metadata. A five-agent chain with zero provenance coverage has five times the debt of a single-agent system, plus a compounding factor for correlated downstream decisions. This term originates with this work.
Definition: The measurable phenomenon where a claim stated with lower confidence by an upstream agent is re-stated by a downstream agent with higher confidence, because the downstream agent has no visibility into the original uncertainty. Each agent in the chain treats the prior agent's output as a fact rather than an estimate, and synthesizes it with the fluency and assertiveness of a confident conclusion.
Why it happens mechanically: Large language models are trained to produce coherent, assertive output. When a model receives a claim as part of its context, it has no native mechanism for propagating the original uncertainty unless that uncertainty is explicitly encoded in the context. A claim received as: "The quarterly revenue declined by 4%" is processed identically to "I am 60% confident the quarterly revenue declined by something in the range of 2% to 6%." The model does not know which it received unless the text says so. This term originates with this work.
Architecture: How Epistemic Laundering Happens
The diagram below shows a representative four-agent pipeline and the two boundaries where Epistemic Laundering typically occurs. The trust boundary (dashed clay border) marks the zone where unverified claims enter the pipeline. The blue arrows show data flow. The annotations show where provenance is present versus stripped.
Fig. 1: Confidence Amplification Across Agent Hops
The chart below is a directional illustration of Confidence Amplification: a claim's stated confidence inflates across agent hops when no provenance is carried. The verified-claim line is flat because uncertainty is preserved at each hop. The unverified-claim line rises toward saturation by hop 4. Both curves are directional; precise rates depend on model, prompt design, and task type.
Fig. 2: Trust Propagation Debt by Architecture Pattern
Different multi-agent architectures accumulate Trust Propagation Debt at different rates. A single-agent system with no handoffs has near-zero debt. A deep orchestration chain with no provenance metadata has the highest debt. Patterns with provenance-carrying protocols (such as structured tool call results with uncertainty fields) accumulate debt more slowly.
Fig. 3: Epistemic Laundering Risk Matrix
The matrix below maps pipeline depth against claim verifiability. High-depth pipelines processing inherently low-verifiability claims (opinion synthesis, forecasting, legal interpretation) carry the highest laundering risk. Single-agent systems processing high-verifiability claims (database lookups, arithmetic, document retrieval) carry the lowest.
Decision Framework: Four Variables That Determine Severity
Not every multi-agent system carries the same epistemic laundering risk. The severity depends on four structural variables:
- Pipeline depth. How many agent-to-agent handoffs occur before the output reaches a human or a consequential action? Each hop without provenance metadata is a laundering event. Severity scales with depth.
- Claim verifiability at origin. Is the upstream claim a lookup (high verifiability) or an inference (low verifiability)? Inferences compound more dangerously through downstream agents than factual retrievals because the original uncertainty is higher and less legible.
- Decision reversibility at the output layer. Is the downstream action reversible (a draft recommendation) or irreversible (a trade execution, a clinical decision, a legal filing)? Laundered confidence reaching an irreversible action is categorically more dangerous than the same confidence reaching a draft for human review.
- Human review quality at the output layer. Does the human reviewer have domain expertise sufficient to catch a confident-sounding error? In high-volume workflows, reviewers often lack the expertise to challenge a polished AI recommendation. The downstream human review is not a reliable backstop against Epistemic Laundering.
If your pipeline has depth greater than two, and any agent in the chain makes inferences (not just lookups), and the output feeds a consequential or irreversible action: you have a material Trust Propagation Debt that requires structural remediation, not a prompt engineering fix.
Minimum Viable Team
A pilot remediation program for Trust Propagation Debt in an existing multi-agent system requires:
- 1 Senior ML Engineer (owns uncertainty propagation protocol): Designs and implements the provenance metadata schema, instruments inter-agent handoffs, and builds the confidence scoring layer. Should have experience with structured output from LLMs and familiarity with at least one orchestration framework (AutoGen, LangGraph, or equivalent).
- 1 Platform/Data Engineer (owns observability infrastructure): Instruments the pipeline for provenance logging, builds the audit trail store, and ensures downstream systems can consume and display uncertainty metadata.
- 1 Domain Expert (owns claim verifiability taxonomy) part-time: Defines which claim types in the organization's specific pipeline are high versus low verifiability. This person must understand the domain (legal, financial, clinical) rather than the AI stack.
- 1 Product Owner with AI literacy (owns requirements and evaluation criteria): Defines what "acceptable epistemic integrity" means for each use case, sets the go/no-go criteria for the pilot, and owns communication to the business stakeholder.
Scale-up adds a second ML engineer to own evals and red-teaming, and a dedicated governance liaison if the pipeline operates in a regulated environment.
Implementation Roadmap
Map every agent-to-agent handoff in the current pipeline. Classify each by claim type (lookup vs. inference), depth position, and current provenance coverage. Produce a Trust Propagation Debt register. Go/no-go gate: stakeholder sign-off on debt register and prioritized remediation scope.
Implement structured provenance metadata at the two highest-risk handoff boundaries identified in Phase 1. Define the uncertainty schema: confidence score, source reference, claim type tag, inference chain depth. Run the modified pipeline against test cases and red-team outputs. Go/no-go gate: measurable reduction in Confidence Amplification rate on test set vs. baseline.
Extend provenance protocol to all handoff boundaries. Integrate uncertainty display into the human-facing output layer. Build review UX that surfaces original confidence and source chain alongside the synthesized recommendation. Define monitoring metrics: provenance coverage rate, Confidence Amplification delta, and human override rate as a proxy for reviewer trust calibration.
Risk Register
| Failure Mode | Early Signal | Mitigation | Severity |
|---|---|---|---|
| Schema drift: Provenance metadata format changes between agents and the schema is no longer consistent at the output layer. | Downstream agent begins ignoring or dropping metadata fields. Confidence scores become null at output. | Version the provenance schema. Validate schema conformance at every handoff boundary with a lightweight guard layer. | High |
| Confidence score gaming: An upstream agent is prompted (intentionally or by training) to report artificially high confidence, defeating the provenance system. | Confidence scores cluster near 0.9+ across all claim types, including low-verifiability inferences. | Calibrate confidence scores against a ground-truth eval set. Red-team upstream agents with prompts designed to elicit overconfident outputs. | High |
| Reviewer anchoring: Human reviewers at the output layer defer to AI confidence scores without exercising independent judgment, even when scores are visibly low. | Human override rate drops to near-zero across a review cycle, including for low-confidence recommendations. | Introduce reviewer calibration training. Design UX to require explicit acknowledgment of uncertainty before approving low-confidence outputs. | High |
| Provenance coverage debt in new pipeline segments: New agents are added to the pipeline without being instrumented for provenance, silently re-introducing laundering at new boundaries. | Trust Propagation Debt register diverges from the live pipeline architecture during routine sprints. | Make provenance instrumentation a mandatory checklist item in the agent deployment process. Gate agent deploys on coverage validation. | Medium |
| Orchestration framework incompatibility: The chosen orchestration layer does not natively support structured metadata on inter-agent messages, requiring a brittle workaround that breaks under version upgrades. | Metadata loss occurs after a framework version upgrade. Provenance coverage drops without a clear root cause. | Abstract the provenance layer above the orchestration framework. Do not rely on framework-native message metadata that is not version-stable. | Medium |
Three Enterprise Scenarios
ROI and Cost of Inaction
A single high-visibility decision made on laundered AI inference, and subsequently reversed or litigated, carries direct costs in settlement, remediation, regulatory response, and reputational damage. In financial services, clinical, and legal contexts, these costs are material. The pattern of laundering is the liability, not just the individual decision.
Multi-agent AI programs that generate high-profile errors are frequently suspended rather than fixed, because the architectural root cause is not understood. The cost of suspension includes sunk development investment, lost productivity from reverting to manual workflows, and organizational credibility loss for the AI program overall.
A provenance protocol implemented at the pilot stage, before the pipeline reaches critical workflows, costs a fraction of retrofitting. The primary cost is engineering time (two to four weeks for a four-agent pipeline) and schema design. The cost of adding provenance to an already-deployed, enterprise-embedded pipeline is an order of magnitude higher.
The payback on epistemic integrity investment is event-driven, not recurring. A single averted high-severity decision error in a regulated context typically recovers the full investment in the provenance protocol. The more useful frame is not ROI but risk-adjusted cost: what is the expected cost of a laundering-driven decision error, discounted by the probability of occurrence given pipeline depth and use case sensitivity?
Executive Checklist
Before deploying or expanding any multi-agent pipeline in a consequential workflow, a CTO, CISO, or Chief AI Officer should be able to answer all of the following.
-
01
Can you identify every agent-to-agent handoff in the pipeline and what information travels through each?Good answer: a documented handoff map with message schema for each boundary.Red flag: "the orchestration framework handles that" with no further detail.
-
02
Does any agent in the pipeline make inferences rather than pure lookups? If so, what uncertainty metadata travels with those inferences?Good answer: inference-class claims carry a confidence score and source reference. Downstream agents consume and preserve that metadata.Red flag: "the model is pretty accurate" with no mention of inter-agent provenance.
-
03
What is the pipeline depth (number of sequential agent handoffs) before a claim reaches a human or triggers an action?Good answer: three or fewer, with provenance tracked at each hop. Or: deeper, with explicit mitigation for each additional hop.Red flag: "it depends on the query" with no maximum depth control.
-
04
Can the human reviewer at the output layer see the original confidence scores and source chain for the claims in the recommendation?Good answer: yes, the review interface surfaces the provenance chain on demand.Red flag: the reviewer sees only the synthesized recommendation with no access to the underlying evidence quality.
-
05
Is there a minimum confidence threshold below which a claim requires human verification before it can be passed to a downstream agent?Good answer: yes, with a defined threshold per claim type and a routing mechanism for below-threshold claims.Red flag: no threshold; all claims pass through regardless of confidence.
-
06
Has the pipeline been red-teamed specifically for Confidence Amplification? That is: has a team deliberately injected low-confidence upstream claims and measured the confidence of downstream outputs?Good answer: yes, with documented results and mitigations applied to failure cases.Red flag: red-teaming focused on individual agent accuracy but not on inter-agent propagation.
-
07
Is there a process to update the Trust Propagation Debt register when new agents are added to the pipeline?Good answer: provenance instrumentation is a required step in the agent deployment checklist, gated before production.Red flag: new agents are added on an ad-hoc basis with no structured review of their provenance handling.
-
08
Is the output of this pipeline feeding any irreversible actions: trade executions, clinical decisions, legal filings, or contract executions?Good answer: yes, and there is a mandatory human review gate with explicit uncertainty display before any irreversible action is taken.Red flag: yes, and the human review is a rubber-stamp step with no visibility into claim provenance.
Build vs. Buy vs. Configure
- Provenance metadata schema: Build in-house. This is domain-specific: the claim types, verifiability taxonomy, and uncertainty representation must match your specific pipeline and use case. No vendor provides a turnkey schema appropriate for every enterprise context.
- Orchestration framework: Configure from existing stack (AutoGen, LangGraph, CrewAI, or equivalent). Do not rebuild orchestration. Abstract the provenance layer above it to avoid framework lock-in.
- Observability and audit trail: Buy (a structured logging and tracing solution) and configure for provenance-specific fields. Building a custom audit trail store from scratch is not a good use of the pilot budget.
- Review interface for human reviewers: Build, with design input from domain experts. The UX for surfacing provenance to a non-technical reviewer is a design problem specific to your workflow. Generic dashboard tooling will not surface uncertainty in a way that is legible to a clinical, legal, or financial reviewer.
- Red-teaming and evaluation: Build internally using your domain team to craft adversarial cases, with optional external augmentation for regulated contexts. Outsourcing red-teaming without domain expertise misses the most dangerous failure modes.
References
- Rawte, V., Sheth, A., and Das, A. (2023). "A Survey of Hallucination in Large Foundation Models." arXiv:2309.01219. arxiv.org/abs/2309.01219
- Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., and Liu, T. (2023). "A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions." arXiv:2311.05232. arxiv.org/abs/2311.05232
- NIST. (2023). "Artificial Intelligence Risk Management Framework (AI RMF 1.0)." DOI:10.6028/NIST.AI.100-1. doi.org/10.6028/NIST.AI.100-1
- Wu, Q., Bansal, G., Zhang, J., Wu, Y., Li, B., Zhu, E., Jiang, L., Zhang, X., Zhang, S., Liu, J., Awadallah, A.H., White, R.W., Burger, D., and Wang, C. (2023). "AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation." arXiv:2308.08155. arxiv.org/abs/2308.08155
- Li, G., Hammoud, H., Itani, H., Khizbullin, D., and Ghanem, B. (2023). "CAMEL: Communicative Agents for Mind Exploration of Large Language Model Society." arXiv:2303.17760. arxiv.org/abs/2303.17760
- Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., and Cao, Y. (2022). "ReAct: Synergizing Reasoning and Acting in Language Models." arXiv:2210.03629. arxiv.org/abs/2210.03629
- Park, J.S., O'Brien, J.C., Cai, C.J., Morris, M.R., Liang, P., and Bernstein, M.S. (2023). "Generative Agents: Interactive Simulacra of Human Behavior." arXiv:2304.03442. arxiv.org/abs/2304.03442