Data integrity does not guarantee meaning integrity
A document can remain unmodified while its meaning changes in the way it is represented, summarized or interpreted downstream.
Checksums, signatures and immutable storage are valuable because they show that recorded bytes remained intact. But many decision failures occur without any corruption of the underlying data.
The change happens in representation: which details are selected, which context is omitted, how uncertainty is expressed and which interpretation becomes dominant.
The chain from evidence to decision
A typical evidence-sensitive workflow contains several representational transitions rather than one direct step from source to outcome.
- Evidence
- Record
- Summary
- AI analysis
- Decision
- Review
Evidence: the original basis
Evidence is the source material available to the process: documents, measurements, records, statements, observations or other artefacts.
At this stage, meaning is still closely tied to the original context. The evidence may contain ambiguity, contradiction, uncertainty or qualifications that later stages compress or reorganize.
A reliable evidence architecture therefore starts by preserving identity, version and provenance before interpretation begins.
Record: structure already introduces selection
Turning evidence into a structured record is already a transformation.
Fields are selected. Categories are assigned. Free text may be normalized. Some information becomes first-class data while other information remains in notes or disappears entirely.
Nothing needs to be technically incorrect for this step to change what later systems can see.
Summary: compression creates semantic risk
Summaries are useful precisely because they remove information. That also makes them one of the most important semantic boundaries in a decision chain.
A summary may preserve the main conclusion while losing uncertainty, chronology, dissenting evidence, exceptions or conditions attached to the source.
The result can be factually consistent with the original material while supporting a different interpretation.
AI analysis: interpretation becomes explicit
When AI analyses a record or summary, the system moves from representation toward interpretation.
Model behaviour may depend on prompt structure, retrieved context, configuration, model version and system state. The same source evidence can therefore support different analyses under different conditions.
For reconstructability, the analysis should remain linked to the exact evidence basis and relevant configuration that produced it.
Decision: interpretation becomes consequence
A decision converts analysis into an operational consequence.
At this stage, the important question is not only what the AI produced, but how that output was used, weighted or overridden by human judgment and policy.
A strong decision trail distinguishes source evidence, machine analysis, human interpretation and the final authority responsible for the outcome.
Review: later reconstruction can introduce another interpretation
Review is often treated as a neutral reconstruction of what happened. In practice, reviewers may see a different subset of records, summaries or system outputs than the original decision-makers.
If the original evidence basis and transformation chain are unavailable, the review may evaluate the record of the decision rather than the information that actually produced it.
That creates a second-order semantic risk: the review can become an interpretation of an interpretation.
Where semantic drift can occur
Semantic drift does not require malicious alteration. It can emerge naturally whenever information changes representation.
- omission of contextual detail
- loss of uncertainty or qualification
- reordering of evidence
- normalization into categories
- summarization
- translation
- model interpretation
- human interpretation
- policy mapping
- later reconstruction from incomplete records
Why hashes alone cannot detect semantic drift
A cryptographic hash can prove that a particular artefact has not changed. It cannot prove that a later summary preserved the important meaning of the earlier artefact.
If both the source and summary are individually intact, their hashes may both verify successfully while the relationship between them remains semantically weak.
This is why evidence integrity and transformation verification should be treated as related but separate problems.
Preserve transitions, not only artefacts
A reconstructable system records more than a collection of documents. It preserves the relationships between them.
For every material transition, the system should be able to identify the input, output, actor or process, relevant configuration and the type of transformation that occurred.
That makes it possible to inspect where meaning may have changed rather than merely confirming that individual files exist.
Semantic preservation is not absolute
No technical system can guarantee preservation of every possible human interpretation.
The practical objective is narrower: define which semantic properties matter to the workflow and preserve enough evidence to evaluate whether those properties survived a transformation.
For one workflow, chronology may be critical. For another, uncertainty or source attribution may matter more.
The verification boundary should state which properties are being evaluated rather than claiming universal semantic equivalence.
A useful technical model
A stronger evidence chain treats each transition as an explicit object of review.
- identify the source artefact
- record the derived artefact
- record the transformation type
- record the participating identity or system
- capture materially relevant configuration
- preserve provenance and lineage
- record integrity evidence
- declare which semantic properties matter
- produce a structured verification outcome
Meaning should be reviewable at the point of change
The most useful place to detect semantic change is the transition where it occurs, not months later at the end of the process.
If a summary removes a qualification, that relationship should be inspectable between source and summary. If AI analysis introduces an interpretation, the evidence basis and configuration should remain visible at that stage.
This reduces the need to infer meaning retrospectively from incomplete downstream records.
Start with one transformation chain
A practical assessment can begin with one real workflow and map every transition from source evidence to final decision.
At each step, ask what information entered, what representation left, what changed, who or what performed the transformation and whether the relationship can later be reconstructed.
The result is often more revealing than examining the final AI output alone.