Anatomy of a Murder: AI Data Chains, Trusting the Process, & The Case of the Missing Person
AI data chains, trusting the process and not just better models, and the case of the missing person who connects the pieces together.
Every organization has its own version of a murder mystery.
A project dies without one obvious cause. A customer relationship deteriorates. A number changes in a way nobody expected. A decision appears settled, but the record does not make clear how it was reached. A document says one thing, while the people involved remember another.
The “murder” is metaphorical. It is the moment something breaks, fails, disappears, or becomes difficult to explain. The point is not to identify a culprit. It is to understand what happened.
The clues are usually buried somewhere in the material: emails, spreadsheets, meeting notes, drafts, messages, recordings, presentations, images, documents, and the metadata around them. But they are scattered across systems, formats, dates, and people. Some are relevant. Some are not. Some conflict. Some are missing.
That last problem is more important than it sounds.
Much of the conversation about AI assumes that the main challenge is intelligence. Give the model a larger context window. More tokens. More GPUs. More infrastructure. More autonomy. Put enough information into the machine and, eventually, it will understand the whole situation and produce the right answer.
Capability will matter. Better models can read more formats, process more material, and generate better hypotheses. But scale alone does not solve the real problem.
The missing element is often not intelligence. It is the missing data, the missing process, and the missing person.
A language model can read a document, answer a question, and produce a convincing explanation. It can also flatten context, overlook material it was never shown, confuse inference with fact, and state a conclusion with more confidence than the record allows. NIST calls this risk confabulation: generated content that is erroneous, inconsistent, or disconnected from the input, yet presented in a credible form. NIST also warns that generated citations and reasoning can create a misleading appearance of support. NIST, Artificial Intelligence Risk Management Framework: Generative AI Profile. (nist.gov)
A larger black box is still a black box.
The better question is not whether a model can become intelligent enough to solve every mystery. It is whether the process around the model can preserve the connection between the material it receives, the events it reconstructs, the evidence it relies on, and the judgment a person ultimately makes.
The answer should not be a black box. It should be a chain.
Chain of Data
Every mystery begins with a record that is both too large and too incomplete.
There may be thousands of pages of documents, years of communications, multiple versions of the same presentation, numbers that changed without explanation, and relevant details embedded in images or recordings rather than text. The record may include direct materials, later recollections, summaries, opinions, errors, omissions, and contradictions.
The first job is not to generate an answer. It is to establish the corpus of material.
What do we have? Where did it come from? When was it created? Who created it? What version are we looking at? What is original, and what is a later summary? Which materials are connected? What is missing?
This is the Chain of Data.
A mature system begins by collecting and preserving the underlying materials rather than treating them as disposable prompt input. It extracts usable information from documents, images, recordings, spreadsheets, and messages. It retains the source, date, authorship, version, location, and relationship of each item to the larger record.
That work may involve transcription, document processing, metadata capture, version comparison, indexing, entity recognition, and structured extraction. AI can assist at several points. But extracted information must remain tied to the underlying material.
The difference matters.
A conventional AI interaction often begins with a fragment. A user pastes a few paragraphs into a chat window, uploads a file, or asks a one-shot question. The model may produce a useful answer, but it cannot know what was left out, what context was already lost, or whether the material it received fairly represents the larger record.
Even very long context windows do not make this problem disappear. Research on long-context models has found that performance can decline sharply when relevant information appears in the middle of a large input, rather than near the beginning or end. More recent work on retrieval systems likewise finds that context size, ordering, retrieval quality, and model choice interact in ways that do not reduce to simply giving a model more material. Liu et al., “Lost in the Middle”; Gabín, Perez, and Parapar, “Lost in the Evidence?”. (arxiv.org)
A better process starts with the record itself.
The goal is not simply to make information searchable. It is to preserve the links among the data: the email and its attachment, the spreadsheet and the assumptions behind it, the meeting notes and the decision that followed, the draft and the changes made before it became final.
A summary can be useful. It is not a substitute for the record.
In a good murder mystery, the crucial clue is often ordinary. A time on a receipt. A changed phrase in a letter. An object that appears in the wrong place. Something that is missing. It only becomes meaningful because it remains connected to the rest of the story.
Data works the same way.
Chain of Events
Once the material is gathered and structured, a person can begin to ask what happened and what might happen next.
What came first? What changed? Who knew what, and when? Which decisions followed which information? Where do accounts diverge? Which facts fit together, and which do not?
This is the Chain of Events.
It is the effort to reconstruct a coherent account of sequence, change, and consequence from a fragmented record.
AI can be genuinely useful here. It can assemble timelines across messages, documents, data, and recordings. It can compare versions and identify changes. It can surface a communication that preceded a decision. It can identify repeated themes, inconsistent accounts, unexplained gaps, and relationships that deserve closer attention.
These are leads. They are not conclusions.
A timeline does not prove causation. A pattern does not establish intent. A discrepancy does not automatically establish which account is correct. The fact that an explanation is coherent does not mean it is true.
This is where raw AI often fails most attractively. It takes incomplete material and turns it into a finished story. It makes uncertainty disappear because uncertainty is awkward in prose.
The missing element is process.
A mature process distinguishes among:
- What the underlying record directly shows;
- What a participant claimed or remembered;
- What the system extracted, organized, or compared;
- What may reasonably be inferred;
- What remains unknown.
It does not merely retrieve a few passages, generate an answer, and append citations afterward.
That is a harder problem than it appears. Research on retrieval-augmented generation has found that keeping an answer faithful to its supplied context remains an active technical challenge, especially when the model’s learned assumptions conflict with the retrieved record. Zhang et al., “FaithfulRAG”. Other researchers have documented attribution bias and sensitivity in systems designed to link answers back to their source documents. Abolghasemi et al., “Evaluation of Attribution Bias in Generator-Aware Retrieval-Augmented Large Language Models”. (aclanthology.org)
This does not mean AI has no role in the process. It means the role must be designed.
Some tasks should be automated. Extracting text from a scanned page. Transcribing recordings. Finding duplicates. Comparing document versions. Sorting a large record. Identifying dates, entities, and repeated references. Pulling structured fields from forms.
These tasks do not eliminate human judgment. They make room for it.
The mature system does not promise that automation will produce the final answer or predict the probable outcomes. It uses automation to make the record more visible, the sequence more intelligible, and the remaining uncertainty more explicit.
Sometimes the chain of events will support a clear account. Often it will support something narrower: this is the most plausible explanation, these are the facts that support it, these are the competing explanations, and this is what would change the analysis.
That is not a lesser outcome. It is the difference between a persuasive story and an accountable understanding.
Chain of Evidence
A coherent account is not enough.
In any decent murder mystery, the final explanation must account for the chain. It has to explain why the timeline matters, why one fact conflicts with another, and why the seemingly minor detail was not minor after all.
The same principle applies here.
If someone says, “Here is what happened,” the next question should be simple: How do you know?
A mature system should make that question easy to answer.
Every consequential conclusion should remain connected to the material beneath it: the document, message, data point, image, recording, date, version, or other source that supports it. A person should be able to move backward through the process:
- Conclusion: What does the available record support?
- Chain of events: What sequence, comparison, relationship, or inference led there?
- Chain of data: Which underlying sources, versions, excerpts, timestamps, and records support each step?
This is the Chain of Evidence.
It is not a few citations appended after a paragraph has been generated. It is a reconstructible record of how the process moved from raw materials to extracted data, from data to observations, from observations to inferences, and from inferences to a conclusion a person is prepared to stand behind.
The record should show what happened along the way. Was text extracted from a scanned document? Was a recording transcribed? Were two spreadsheets compared? Was data normalized? Did the system identify a pattern across records? What material supports that pattern? What material contradicts it? What assumptions were made? What gaps remain?
This is where the person returns to the center of the system.
The person is not a ceremonial reviewer who clicks “approve” after the machine has completed the real work. They define the question. They determine the relevant scope. They understand the context that is not captured in the record. They decide whether a discrepancy matters. They decide whether an inference is justified. They remain responsible for the judgment and the action that follows.
In short, the person does not merely receive the answer. They make the answer accountable.
The broader concept of provenance offers a useful model. The C2PA Content Credentials standard is deliberately modest about what provenance can do. It can record an asset’s origin and history in a tamper-evident structure. It does not, by itself, determine whether the asset is true, accurate, or factual. (c2pa.org)
That is the right discipline for systems that help people work through evidence.
A chain of evidence is not a truth machine. It does not relieve anyone of judgment. It gives a person the ability to inspect the basis of a conclusion, challenge it, correct it, and understand where certainty ends.
Trusting the Process
This is why the argument is not really about AI.
It is about the trusting the process that surrounds it.
Models matter. Better models will improve the process. But a model is a component. A useful process is an entire system.
That process begins before the model sees anything: with collection, preservation, extraction, organization, and context. It continues when AI helps a person search, compare, synthesize, and test possible accounts of what happened. It ends not with a polished answer, but with a record that makes the answer inspectable.
Without the Chain of Data, the system has nothing reliable to work from.
Without the Chain of Events, the data remains a pile of fragments.
Without the Chain of Evidence, the conclusion cannot be tested.
And without a person responsible for judgment, the process collapses into a machine producing language about a reality it cannot fully see.
This is not an argument against more capable models, better infrastructure, or more automation. Those developments may be necessary. But they are not sufficient.
The mistake is to confuse a more capable answer engine with a more reliable system of understanding.
The Missing Person
The promise of a mature process is not that it will answer every mystery correctly.
It is that it can give a person better tools for working through real uncertainty. It can retain the underlying record, reveal its structure, make competing explanations easier to test, and preserve the basis for every important conclusion.
The standard should be a continuous chain:
Person → Data → Events → Evidence → Human Judgment
Not an answer generated from a fragment. Not a confident summary detached from its sources. Not an automated process that becomes less accountable as it becomes more powerful.
The person still asks the question. The person still tests the story. The person still decides what the evidence supports.
AI may be part of that work. It may be a powerful part.
But the real product is the process that carries a person from the first piece of data to a final, accountable judgment without breaking the chain.
Note: This article was drafted with the assistance of AI and is for informational purposes only. It does not constitute legal or professional advice of any kind.