Provenance Is Evidence: The Missing Layer in AI Accountability
The AI followed the rule. But who supplied the facts?
A deterministic AI policy can execute perfectly and still produce a questionable result if the labels, assumptions or scores entering that policy came from an undocumented, interested or unreliable source.
That is why provenance should not be treated as an administrative footnote. Provenance is evidence about the evidence.
Source, method, version and reviewer context can change how much confidence a label deserves even when the number itself does not change.
AI governance is not only about rules. It is also about inputs.
Consider a simple policy:
render_allowed = restoration_value >= attention_cost
If attention cost is declared as 5 and restoration value is declared as 1, the result is deterministic:
attention_cost = 5 restoration_value = 1 RESULT = SUPPRESS
The policy executed correctly. But one critical question remains: where did the values 5 and 1 come from?
Reproducibility answers whether the same inputs and same rule produce the same result. Provenance helps answer where those inputs came from.
Preserve the path from source to decision.
Each step can be preserved separately. That creates a clearer audit trail and makes it harder for assumptions to disappear inside a single unexplained score.
The same label can carry different levels of confidence.
Product Owner
A company or designer may understand the interface deeply, but can also have a direct commercial interest in the outcome.
Independent Reviewer
A reviewer separated from the product owner can add valuable independence and reduce direct conflicts of interest.
Measured Proxy
A technical or behavioral proxy can improve repeatability, but its definition, instrumentation and interpretation still matter.
Model-Generated Label
Another AI system can assign a label, but the model, prompt, version and confidence should be preserved.
User Declaration
A person may explicitly state a preference or intent. That can be meaningful without being universally generalizable.
Unknown Source
A label with no preserved source should be treated differently from one with documented provenance.
Identical values can represent very different evidence.
| Value | Source | Method | Potential concern |
|---|---|---|---|
| 5 | Product owner | Internal declaration | Possible incentive to shape the outcome |
| 5 | Independent reviewer | Documented evaluation | Reviewer judgment still remains subjective |
| 5 | Automated model | Classifier output | Model, prompt and version may affect the label |
| 5 | Unknown | Not preserved | Little basis for evaluating trustworthiness |
The label value did not change. The evidentiary context did.
Make the label inspectable before it reaches policy.
This does not prove that the label is objectively correct. It makes the source and method visible enough to challenge.
Reproducibility tells you whether execution can be repeated.
If a system receives the same inputs and applies the same deterministic rule, the result should reproduce.
INPUTS attention_cost = 5 restoration_value = 1 POLICY restoration_value >= attention_cost EXPECTED SUPPRESS REPRODUCED SUPPRESS STATUS PASS
That is useful evidence. But it still does not answer whether the labels themselves were well sourced, unbiased or appropriately defined.
Strong governance needs both.
The receipt should preserve more than the final number.
An AI Decision Receipt™ can record the policy input and attach a reference to the provenance record behind that input.
DECISION RECEIPT element: checkout-upsell attention_cost: value: 5 provenance_ref: prov_89b4 restoration_value: value: 1 provenance_ref: prov_a72c policy: restoration_value >= attention_cost decision: SUPPRESS receipt_hash: SHA-256: 98fa41...
This creates a chain from the interface element, to the label, to its source, to the rule, to the final decision.
Provenance can expose who benefits from the label.
A product owner assigning labels to its own interface does not make those labels automatically invalid. Product owners often possess useful domain knowledge.
But if the party assigning a label also benefits from the resulting policy decision, that relationship is relevant evidence.
Declared interest
Preserve whether the label source owns, sells or benefits from the evaluated system.
Independent review
Where useful, preserve whether a second party reviewed the same artifact.
Disagreement
Conflicting labels can be preserved rather than collapsed into a false appearance of certainty.
If a model assigns the label, the model becomes part of the evidence.
AI systems themselves can be used to classify, score or interpret interface elements. If that happens, provenance should preserve enough information to identify the model context.
source_type: ai_model model: evaluator-model-x model_version: 2026-09-01 prompt_version: provenance-labeler-v4 temperature: 0 output: attention_cost = 5 confidence: declared_model_confidence = 0.82
A model-generated confidence score should not be confused with objective truth. It is another generated output whose origin should remain visible.
Governance gets stronger when uncertainty is allowed to remain visible.
Systems often feel pressure to replace uncertainty with a number. That can create a false impression of precision.
If the origin of a label is unknown, a transparent system can preserve:
source: UNKNOWN method: UNKNOWN confidence: NOT ESTABLISHED policy_action: REVIEW_REQUIRED
Provenance increases transparency. It does not create automatic truth.
A perfectly documented source can still be wrong. An independent reviewer can still be biased. A measured proxy can still be poorly designed. A model can still hallucinate. A user declaration can still change.
Provenance should therefore be understood as a transparency and accountability mechanism, not a universal truth detector.
It shows origin.
It helps identify where a label or assumption came from.
It enables challenge.
Reviewers can question the source, method and conflicts.
It preserves context.
Future reviewers can see more than a detached number.
Meaning + provenance + policy + receipt creates a stronger chain of evidence.
1. IDENTIFY What interface element is being evaluated? 2. LABEL What declared or modeled values are assigned? 3. PRESERVE PROVENANCE Who or what supplied each label? Which method and version produced it? 4. APPLY POLICY Execute the explicit rule. 5. PRESERVE DECISION Record the result. 6. HASH ARTIFACTS Preserve integrity fingerprints. 7. REPRODUCE Allow independent reviewers to run the same decision again.
This architecture does not require pretending that every label is objectively measurable. It requires being honest about what is declared, modeled, measured, inferred or unknown.
Do not trust the provenance. Inspect it.
ĀML™ — ĀRU Meaning Language™ is presented as a working open-source research prototype. Developers, researchers and skeptics are encouraged to inspect the policy, inspect the inputs, question the provenance and independently reproduce the result.
Explore the ĀML™ governance stack.
— Daniel Jacob Read
Comments
Post a Comment