The Missing Layer in AI: Source Integrity for Decision Systems

Introduction: The Misunderstood Risk

AI systems are increasingly used to process source material and produce outputs that inform decisions. These outputs are often fluent, structured, and internally consistent. As a result, they are treated as reliable representations of the underlying information.

Hallucination is one well-known failure class: cases where the model generates content with no basis in any source. Source transformation creates another.

The output can remain substantially grounded in its source while conditions, qualifications, scope, certainty or relationships change.

When AI transforms source material, it produces something cleaner than what the source actually said. Conditions may be removed, uncertainty may be resolved, and contextual relationships may be simplified or omitted. The output remains plausible and coherent, but no longer fully reflects the structure and constraints of the original material.

Because the output appears complete and authoritative, returning to the source is an extra step that may not be taken. Decisions are made based on the AI's representation of the information rather than the information itself.

This creates a new class of risk. The issue is not that the output is obviously incorrect. It is that the output is subtly misaligned with the source in ways that are difficult to detect and easy to accept.

As AI becomes more embedded in professional workflows, this risk scales with it. The volume of decisions informed by AI-generated outputs increases, while the ability to check whether those outputs still say what the source said does not.

Understanding and addressing this gap is not a matter of improving model accuracy alone. It requires a different approach: one that looks at how meaning changes as information moves, not just whether the final output looks correct.

The Problem: Invisible Meaning Loss

When AI processes source material and produces an output, it does not preserve the original information. It transforms it.

Prior work in summarization faithfulness and factual consistency has identified similar transformation effects, where outputs remain fluent while diverging from source meaning [1][2][3].

This transformation is not inherently incorrect. It is what allows AI systems to organize and interpret complex inputs. However, in doing so, the system often produces something cleaner than what the source actually said.

The resulting output remains coherent and plausible, but it no longer fully reflects the structure and limits of the original material. An output can be factually correct and still misrepresent the meaning of the source by omitting qualifying conditions, overstating conclusions, or compressing nuanced information into a single narrative.

The issue is not whether the output is "right" in isolation. It is whether the output remains faithful to what the source actually supports. Grounding, faithfulness and factual-consistency evaluation increasingly examine parts of this relationship. The remaining problem is whether the complete structure of relevant source meaning survived into the output.

In most workflows, outputs are assessed based on clarity, coherence, and perceived correctness. These qualities make the output easier to use, but they also obscure what may have been lost in the transformation process. Users rely on outputs that appear complete without visibility into what was preserved, what was altered, and what was omitted.

Failure Modes: How Meaning Gets Altered

Meaning loss in AI-generated outputs is not random. It follows consistent patterns that emerge from how systems transform and simplify information. These patterns are often subtle. The output remains coherent and defensible, but key aspects of the original meaning are altered or lost in ways that are difficult to detect.

Across all cases, the pattern is consistent: the output remains plausible, the structure appears coherent, the claims are defensible, and the relationship to the source has still changed. These failures persist because they are not obvious on the face of the output, and because detecting them requires comparing the output against the source rather than reading the output alone.

Why Current Approaches Are Necessary but Incomplete

Existing approaches to improving AI reliability focus on model capability, retrieval quality, and review. Each addresses a real dimension of the problem. What none of them establishes on its own is whether the full decision-relevant meaning of the source survived into the output.

Retrieval-Augmented Generation

Retrieval-based systems improve access to relevant source material by grounding model outputs in retrieved documents or passages. Retrieval improves access to relevant source material. It does not guarantee preservation of every condition, exception, qualification or source relationship. Even when relevant source material is correctly retrieved, conditions may be omitted, qualifiers may be flattened, and relationships between retrieved elements may be altered. Citations do not resolve this. A cited source may contain the necessary information while the output still misrepresents how that information was originally structured. Retrieval improves what the model is given but does not guarantee that the output stays faithful to it.

Evaluation and Benchmarking

Most evaluation frameworks measure factual correctness, similarity to reference answers, and performance on standardized tasks. An output can be factually correct, logically consistent, and highly rated by benchmarks, and still omit critical conditions, overstate conclusions, or resolve uncertainty that was present in the source. Benchmark performance evaluates outcome plausibility. It does not evaluate transformation integrity.

Recent evaluation frameworks such as RAGAS [5], TruLens [6], and G-Eval [7] provide structured ways to assess output quality, grounding, and consistency. These approaches improve observability and benchmarking, but primarily evaluate outputs at an aggregate level and do not explicitly model how meaning is transformed between source and output at the level of individual assertions.

Similarly, methods such as SelfCheckGPT [8] attempt to detect hallucination or inconsistency through self-evaluation. While useful for identifying unsupported content, they do not provide a structured account of how outputs differ from source material when transformations are subtle rather than outright incorrect.

Human Review

Human review is widely used as a safeguard in AI-assisted workflows, but it has structural limitations: it is post-hoc, inconsistent across reviewers, and non-scalable as AI usage expands. Outputs that are fluent, coherent, and internally consistent are difficult to challenge even when they diverge from the source. Human review often approves output that looks correct rather than checking whether it still says what the source said.

Model Improvement

Better models and better prompting can reduce these errors. They do not establish a guarantee that source meaning will always be preserved. More fluent and more coherent output can also be harder to challenge on inspection, which is a reason to make the comparison against the source explicit rather than to rely on the reader noticing.

The Missing Layer: Source Integrity

The preceding sections establish a consistent gap: current systems can generate, retrieve, and evaluate outputs, but they do not determine whether those outputs remain faithful to the meaning and constraints of their source material. This gap is not addressed by improving inputs, increasing model capability, or adding post-hoc review. It requires a distinct layer with a different function.

That layer is source integrity.

Definition

Source integrity is the ability to say what an AI output is actually entitled to claim, measured against the material it was built from. It has four parts: claim support, meaning preservation, material omission and, where more than one source is involved, preservation of the relationships between them. Source integrity complements retrieval, grounding and evaluation by inspecting transformations that matter between bounded source material and generated analysis. It evaluates the relationship between what the source contains and what the output asserts, establishing whether the output stays within the bounds of what the source supports, including its conditions, constraints, and structure.

This shifts evaluation from "does the output sound correct?" to "is the output faithful to what the source actually supports?"

Core Primitives

It works in three parts: break the output into individual claims, trace each claim back to the source, and track what changed on the way. Each assertion in the output is mapped to the specific portions of the source material that support it, establishing not just whether support exists, but whether the structure of the claim corresponds to the structure of the source, including conditions, qualifiers, and relationships.

What the comparison surfaces

System Model: How It Works

It is built as a structured comparison between the source material and the AI output. The goal is not to reinterpret the content, but to evaluate how meaning is preserved or altered during transformation.

Inputs

The system operates on two inputs: the source material, meaning the original content with its conditions, qualifiers, dependencies, and contextual relationships, and the AI output, which is a transformed version of the source intended to be consumed by a user or downstream system.

Processing

Plumb Analysis receives bounded source material and an AI-generated output. It identifies the statements and source information that matter to the comparison, evaluates how the output represents that material, and returns structured findings showing where meaning held, changed, disappeared or was introduced.

The findings describe observable differences rather than internal mechanics: conditions, exceptions, certainty, obligation, scope, timing, numbers, attribution, causality, material omissions and unsupported synthesis.

Outputs

For each assertion, the system provides a classification, a mapping to the relevant source material, and an indication of how the assertion differs from the source where applicable. The output is not a single score or summary. It is a structured account of how the generated content relates to the underlying source material.

This differs from traditional evaluation by looking at what changed between source and output, rather than at output quality alone. The system does not attempt to determine whether an output is useful or well-formed in isolation. It determines whether the output remains within the source bounds defined by the source material.

Real-World Examples

The failure modes described in earlier sections are not theoretical. They occur in common workflows where AI is used to interpret source material and produce outputs that inform decisions.

Earnings Call Analysis

Contract Interpretation

Product Performance Analysis

Implications: Why This Matters at Scale

These failure modes are often framed as issues with individual outputs. That framing is incomplete. The underlying problem is not about individual outputs. It is about how decisions are made when AI systems mediate access to information.

As AI becomes embedded across workflows, it increasingly acts as the layer through which professionals interpret source material. The output is no longer an optional convenience. It becomes the primary input to judgment.

This pattern holds across domains: consulting and advisory work, financial analysis, legal interpretation, enterprise copilots, internal operations. In each case, AI intermediates access to information, outputs are treated as authoritative, source material is no longer directly inspected, and decisions are made on the transformed representation.

As this pattern scales, the impact is cumulative. Small distortions accumulate across decisions, uncertainty is systematically reduced, and critical context is lost incrementally. The result is a class of decision-making that is internally consistent and operationally scalable, but structurally misaligned with the underlying source material it is intended to reflect.

The Future: From Outputs to Governable Autonomy

AI systems are evolving beyond content generation. Their role is expanding from assisting users to informing decisions, and increasingly, to taking action within workflows. As this progression continues, the consequences are no longer confined to interpretation. They begin to affect outcomes directly.

Improvements in model capability increase fluency, coherence, and reasoning quality, but they do not eliminate transformation effects. More capable systems produce more persuasive output and resolve ambiguity more confidently. That raises their usefulness and also raises the chance the output is accepted without anyone checking it.

Autonomous systems depend on the ability to make decisions without continuous human intervention. Without source integrity, decisions cannot be inspected against their source, reasoning cannot be validated after the fact, and system behavior cannot be reliably governed. Increasing autonomy expands system capability, but without accountability, it also expands the scope of untraceable error.

These are two distinct relationships, and they became two products. Plumb Analysis compares a generated representation with the bounded sources behind it. Plumb Boundary compares an evolving workflow state with the authority that workflow was originally given, and returns that result before a consequential step; the workflow decides the consequence. To operate under that constraint, An autonomous system must be able to distinguish between supported and inferred conclusions, retain conditions and dependencies, surface omissions, and maintain a traceable link between source material and action. Source discipline is not a constraint on capability. It is the structural condition under which autonomous systems can be reliably operated.

Conclusion: The Category

AI systems are increasingly used to transform source material into outputs that inform decisions. As this role expands, the relevant question is not only whether these outputs are useful or correct in isolation, but whether they remain faithful to the source material they are derived from.

Existing retrieval, grounding, evaluation and review methods address important dimensions of AI reliability. Source integrity adds a specific question: what happened to the meaning of bounded information as AI transformed it into analysis, and, as AI begins to act, did that evolving state remain within its original authority? This gap becomes more consequential as AI moves from assisting users to shaping decisions and enabling autonomous action.

Source integrity defines the missing layer. It provides a structured mechanism for determining what an AI-generated output is actually supported by, relative to its source. By making this relationship explicit and inspectable, it lets output be judged not only on whether it reads well, but on whether it still says what the source said.

This capability is distinct from existing methods of improving or evaluating AI outputs. It addresses a different problem: not whether the output is well-formed, but whether the output remains within the source bounds of the source it was derived from.

As AI becomes more deeply integrated into professional and operational workflows, the structural significance of this layer increases in proportion to the role AI plays in shaping outcomes.

References

1. Maynez et al., On Faithfulness and Factuality in Abstractive Summarization, ACL 2020.
2. Pagnoni et al., Understanding Faithfulness in Abstractive Summarization, ACL 2021.
3. Ji et al., Survey of Hallucination in Natural Language Generation, ACM Computing Surveys 2023.
4. Gao et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, NeurIPS 2021.
5. Es et al., RAGAS: Automated Evaluation of Retrieval-Augmented Generation, 2023.
6. TruLens, Evaluating LLM Applications, 2023.
7. Liu et al., G-Eval: NLG Evaluation using GPT-4, 2023.
8. Manakul et al., SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection, EMNLP 2023.