Warrant Erosion Principle

• WEP is a formal principle defining how AI generative and interpretive transformations can weaken the epistemic warrant of propositions.
• It highlights that indiscriminate acceptance of tool outputs leads to semantic laundering, where weak evidence is reclassified as strong support.
• Addressing warrant erosion requires explicit epistemic typing in AI architectures to ensure external validation and reliable inference.

The Warrant Erosion Principle (WEP) formalizes a fundamental epistemic vulnerability in contemporary AI agent architectures, particularly those employing LLMs and tool integrations. WEP states that generative or interpretive transformations—such as those effected by LLMs or similar agent components—do not, by default, preserve the epistemic warrant of their inputs. Unless explicitly architected for warrant preservation, such transformations can degrade or even sever the justificatory link between a proposition's acceptance by the system and its grounding in genuine observations or validated inference. This mechanism enables "semantic laundering": a systematic process by which weakly warranted statements can be promoted to strong epistemic status simply by traversing architecturally trusted boundaries, rather than through epistemically relevant inference. WEP thus underpins the architectural realization of Gettier-style failures, in which beliefs are assigned high confidence without proper justification (Romanchuk et al., 13 Jan 2026).

1. Formal Definitions

Epistemic Warrant

A proposition $P$ is conferred warrant according to

$\warrant(P)\;=\;\langle O,\,I\rangle$

where

• $O$: set of observations physically or empirically grounding $P$
• $I$: set of externally validated inference rules by which $P$ is derived from $O$

Warrant strength is partially ordered: $$\langle O_1,I_1\rangle\;\le\;\langle O_2,I_2\rangle$$ iff $O_1\subseteq O_2$ and $I_1\subseteq I_2$.

Epistemic Relevance

Inference step $r$ increases the warrant of a proposition iff it adds either:

• a new observation $o\notin O$, or
• a new externally validated inference rule $r'\notin I$.

Semantic Laundering

A tool boundary or interface $B$ semantically launders if,

$\SL(B)\;\equiv\; \exists P_1,P_2\colon \warrant(P_1)\text{ weak} \;\land\; P_2=B(P_1)\;\land\; \warrant(P_2)\text{ strong}\;\land\; \Inf_{\mathit{ep}}(P_2)=\Inf_{\mathit{ep}}(P_1)$

i.e., the warrant is increased without any epistemically relevant inference.

Warrant Erosion

For any generative or interpreting transformation $T\colon P\to P$,

$\neg\left(\forall P\;\warrant(T(P))\succeq\warrant(P)\right)$

In words, warrant is not preserved through transformations unless explicit guarantees exist (Romanchuk et al., 13 Jan 2026).

2. Architectural Premises

WEP applies under a specific but prevalent set of architectural assumptions:

• Uniform Proposition Space: Both agent and tool outputs are of a uniform proposition type $P$.
• Tool Results as Observations: All tool call results, regardless of provenance, are systematically added to the observation set $O \subset P$.
• Epistemic-Status Assignment: Propositions are licensed or refused for adoption via a function $A: P\to\{\ASSERTIVE,\COND,\REFUSAL\}$, and even LLM-generated propositions may be assigned high license status.
• No Epistemic Typing of Tools: Tools are not differentiated as OBSERVER (direct data acquisition) vs GENERATOR (inferential/generative), resulting in indiscriminate treatment of all tool outputs as 'observations'.

These conditions are satisfied by commonplace LLM-agent paradigms employing black-box tools, plugin ecosystems, or chain-of-thought with interleaved tool use.

3. Mechanism and Proof Sketch of Warrant Erosion

In typical agent architectures, the warrant for a proposition $P_1$ derives from $\langle O, I \rangle$—namely, its observation and inference set. The application of a generative or interpretive transformation (e.g., an LLM call or parsing step) produces a new proposition $P_2 = T(P_1)$, but—by standard design—$T$ does not augment $O$ with new empirical data nor $I$ with externally validated rules. Thus, the warrant cannot strictly increase. Nonetheless, system-level mechanisms often reclassify such outputs as high-confidence based solely on their trajectory through a tool, not epistemic content. This dissociation between license assignment and epistemic grounding exemplifies warrant erosion: the formal or algorithmic status of a proposition becomes stronger than its justificatory roots support.

4. Concrete Manifestations of Semantic Laundering

Several canonical cases illustrate both warrant erosion and semantic laundering:

Case	Sequence	Epistemic Violation
Tool-Boundary Laundering	LLM issues tool call → tool (LLM generator) returns output → runtime marks as 'Observation'	No new external data or validated inference; output is nonetheless accepted with increased warrant
Multi-Agent Self-Licensing	Agent 1 (LLM) claims $p$ → Agent 2 (LLM) 'validates' $p$ → validation accepted as evidence	Circular justification with no extra-architectural grounding
Channel-Based Warrant Assignment	Runtime assigns strong warrant based on 'tool' source, not on content	Epistemic role of tool not even checked; generator output laundered as observation

These patterns are architecturally reproducible and not exceptional.

5. Relationship to Gettier Problem and the Theorem of Inevitable Self-Licensing

The WEP crystallizes the connection between modern agent architecture failures and classical epistemic paradoxes:

• Gettier Invariance: Gettier-style failures (true beliefs with improper warrant) are systematic, not exceptional, because warrant can appear to increase via architectural transformations that do not establish new links to truth-makers.
• Theorem of Inevitable Self-Licensing: Given architectural homogeneity in $P$ and indiscriminate admission of tool outputs as observations, any proposition generated by an agent can be recycled via an auxiliary 'tool' (LLM or otherwise) and re-ingested as ostensible 'evidence', resulting in inescapable circular justification. The assignment of epistemic license via $A$ thus becomes unavoidably influenced by the agent's own previous outputs.
• Classical Gettier cases are accidental; here, warrant erosion is an architectural invariant (Romanchuk et al., 13 Jan 2026).

6. Implications for Agent Design and Compliance

WEP demonstrates that neither scaling, model improvement, nor reliance on LLMs as judges remedies semantic laundering or warrant erosion. As every LLM call by default operates in GENERATOR mode, true warrant is never constructed de novo unless the system is architected for such constraint. Accordingly, only the introduction of explicit epistemic typing—distinguishing OBSERVER, COMPUTATION, GENERATOR, etc., as tool roles—and enforcing that only OBSERVER or COMPUTATION outputs can enter the observation set $O$, can preserve authentic warrant. Content-based, not channel-based, derivations are essential to prevent the laundering of unsupported claims.

Architectures adhering to the Uniform Proposition Space and indiscriminate tool result acceptance (Assumptions A-D) are incompatible with requirements for auditability, compliance, and safety. Traceable, non-circular justification chains cannot be achieved in their absence. Explicit epistemic typing is thus the minimal architectural remedy to prevent warrant erosion and the attendant systematic semantic laundering (Romanchuk et al., 13 Jan 2026).

7. Summary Table: WEP and its Architectural Context

Architectural Aspect	Role in WEP	Minimal Remedy
Uniform Proposition Space	Enables recirculation of outputs	Typed proposition sets
Tool Results as Observations	Indiscriminate warrant assignment	Epistemic type checks
No Epistemic Typing	Launders generator outputs	OBSERVER/GENERATOR distinction
LLM-as-Judge Schemes	Self-licensing and circularity	External validation only

A plausible implication is that any LLM-agent architecture lacking these types of epistemic controls will inevitably admit systematic justification failures—of which warrant erosion is the explanatory core.