LLMs produce fluent answers, not defensible ones. Introducing Investigation Provenance, Answer Debt, and Kill Threshold Governance: the three primitives enterprises need before AI conclusions can be trusted in regulated decisions.
Every enterprise using AI to answer hard questions (due diligence, clinical pathway evaluation, fraud signal validation, exposure estimation) faces a structural problem that no current governance framework names or resolves. The problem is not that AI gives wrong answers. The problem is that AI gives answers with no accountable record of how they were reached.
A wrong answer with a traceable evidence record is a recoverable error. A wrong answer with no record is a liability. The distinction matters in every regulated industry, in every board-level AI review, and in every context where a human signature follows an AI conclusion.
This article introduces three original constructs that close this gap. Investigation Provenance is the traceable chain from question to evidence to claim that makes an AI conclusion auditable. Answer Debt is the organizational liability that accumulates when AI conclusions circulate without that chain. Kill Threshold Governance is the discipline of committing stopping criteria before an investigation begins, so results cannot be silently cherry-picked after the fact.
These constructs are grounded in an emerging open-source implementation called Zorp (aviskaar/zorp, MIT, pre-alpha), which applies pre-registration discipline borrowed from clinical trial methodology to AI-driven investigation. The patterns it demonstrates are not product-specific. They describe an architectural layer that enterprise AI is currently missing.
NIST AI RMF [1] addresses risk categorization, output monitoring, and human oversight at the organizational level. ISO/IEC 42001:2023 [2] addresses AI management system requirements. Both are governance frameworks for AI systems. Neither addresses the investigative accountability layer: the question of whether an AI conclusion is traceable to the evidence that produced it, and whether that evidence was gathered under conditions that prevent post-hoc selection.
This is not a minor omission. It is the precise gap that creates enterprise liability in regulated decisions. Consider three common scenarios:
In each case, the answer is plausible. In each case, the answer is not defensible. The distinction becomes material when a regulator asks, a court asks, or a risk committee asks: show me what this conclusion is accountable to.
Answer Debt is the organizational liability that accumulates when AI conclusions circulate without a traceable evidence record. Like technical debt, it is invisible until it is called in. Unlike technical debt, it may be called in by a regulator, a court, or a board rather than an engineering team. Answer Debt compounds when AI conclusions are used as inputs to subsequent decisions, because the provenance gap propagates forward through every downstream use of the original conclusion.
Investigation Provenance is the traceable chain from a research question to the evidence gathered to the claims made in the final artifact, such that every claim in the output resolves to a row in the evidence record rather than to a paragraph an agent wrote about itself. An AI conclusion has Investigation Provenance if and only if: (1) the hypothesis was registered before evidence gathering began, (2) every attempt was recorded including those that conflicted or failed, and (3) the final artifact was drafted from the evidence record alone, not from the agent's general knowledge.
The clinical trial literature has formalized this requirement for decades. Pre-registration, which means committing a hypothesis, primary metric, and stopping criteria before a trial begins, exists precisely to prevent post-hoc selection of favorable results [3]. An unregistered trial that reports only the analyses that worked is not evidence. It is a narrative. The same principle applies, with equal force, to AI investigation.
The distinction matters in practice because AI agents have a structural incentive toward confident answers. They are optimized for fluency and completion, not for surfacing contradiction or registering failure. Without a pre-registration constraint, an agent can silently iterate through framings until it finds one that produces a satisfying answer, then report that answer as if it were the result of a principled investigation. The answer may be correct. There is no way to know.
Standard AI output chains run from prompt to response. Investigation Provenance requires the reverse: the output is assembled from the evidence record, not from the model's generative capacity applied to the question. This is not a prompt engineering challenge. It is an architectural one. The model must be handed only what the track recorded and instructed to cite only those figures.
Kill Threshold Governance is the discipline of committing, before any evidence is gathered, the metric and the threshold value that would falsify the investigation's hypothesis. A kill threshold is: (a) set by a human, not proposed by the agent; (b) committed to a tamper-evident record before the first attempt; (c) applied automatically when crossed; and (d) recorded as the reason the investigation terminated. Kill Threshold Governance makes cherry-picking structurally impossible rather than merely discouraged.
The key insight is that a kill threshold must be committed before evidence exists, not after. An organization that decides its stopping criteria after seeing preliminary results has not constrained the investigation. It has selected it. The distinction is the difference between a pre-registered clinical trial and a retrospective analysis: both are legitimate research methods, but only the first produces evidence that regulators and courts treat as primary.
For enterprise AI, the governance implication is direct. Before any AI investigation is used to inform a regulated decision, a human must commit in writing: what metric will be measured, which direction constitutes improvement, and what magnitude of result would indicate the hypothesis is wrong. That commitment must be hash-verified so it cannot be quietly revised after the investigation runs.
| Capability | Standard LLM Pipeline | RAG Pipeline | Investigation Provenance Architecture |
|---|---|---|---|
| Pre-registration | None | None | Required |
| Failed attempt recording | None | None | Every attempt |
| Kill threshold | None | None | Human-set, hash-verified |
| Human checkpoint gates | Optional | Partial | Between every stage |
| Evidence record schema | None | Log files only | 6-table typed store |
| Claim-to-row traceability | None | Source docs only | Every claim |
| Regulatory defensibility | Low | Moderate | High |
Not every enterprise AI use case requires full Investigation Provenance. The framework below identifies the four variables that determine when it is architecturally required versus operationally advisable.
If a human will sign a document, approve a decision, or take a regulatory action based on the AI conclusion, Investigation Provenance is required. The signature creates accountability. Accountability requires a traceable record.
If the AI conclusion could be subpoenaed, audited by a regulator, or examined in litigation, Investigation Provenance is required. The question is not whether the conclusion is correct. The question is whether the method that produced it can be defended.
If the action taken on the AI conclusion cannot be undone (a contract executed, a patient pathway changed, or a position taken), Investigation Provenance is required. Reversible decisions can be corrected when provenance gaps are discovered. Irreversible decisions cannot.
If the AI conclusion will serve as an input to a subsequent AI investigation, Investigation Provenance is required at the first step. Answer Debt propagates forward. A conclusion without provenance that becomes an input to a second investigation produces a second conclusion with compounded provenance debt.
A bank uses AI to assess whether a product change triggers regulatory reporting obligations. The AI returns a conclusion that it does not. That conclusion influences a board decision. A regulator later disagrees. Without Investigation Provenance, the bank cannot show what evidence the AI weighed or what contradictory guidance it found and discounted. Kill Threshold Governance would have required the legal team to commit in advance the threshold at which ambiguity requires human counsel review, making the decision process auditable rather than opaque.
A health system uses AI to evaluate whether published evidence supports a care pathway change for a specific patient population. The AI produces a recommendation. The pathway is changed. An adverse outcome occurs. Without an evidence record, the institution cannot demonstrate which studies were weighed, which were discounted, and what the pre-committed threshold was for recommending the change. Investigation Provenance makes every step of the reasoning auditable by a medical board or a plaintiff's counsel.
A biotech uses AI to evaluate which of four drug targets has prior art weak enough to justify a program. The AI recommends one target. The program is initiated. Prior art surfaces in due diligence eighteen months later. Without pre-registration, there is no record of what prior art the AI found, what it weighted, or what threshold of prior art strength would have changed the recommendation. Kill Threshold Governance applied at the outset would have made the stopping criterion explicit and the conclusion defensible.
An internal audit function is asked to verify that AI-assisted conclusions used in a board presentation are accurate. With a standard LLM pipeline, the audit function finds the output but not the method. With Investigation Provenance in place, the audit function can trace every claim in the board presentation to a row in the evidence record, verify that the kill threshold was committed before the investigation ran, and confirm that failed attempts were recorded alongside successful ones. The audit is possible. Without the architecture, it is not.
Identify every AI-assisted conclusion currently influencing regulated decisions. For each, document: who signed off, what evidence exists, whether the method is reproducible. This is your Answer Debt audit. The output is a prioritized list of provenance gaps by risk level.
For all Tier 1 use cases (downstream signature, regulatory exposure, or irreversibility), implement pre-registration: hypothesis, metric, and kill threshold committed before any AI investigation begins. Start with a Git-based record. Add hash verification. Enforce via the human checkpoint gate before evidence gathering starts.
Build or adopt a six-table evidence record store: tracks, preregistrations, experiments, metrics, checkpoints, validations. Require that every AI conclusion delivered to a human decision-maker resolves to a row in this store. Integrate with existing audit and compliance infrastructure.
Organizations implement logging systems that look like evidence records but do not enforce pre-registration. The logs capture what the agent did, not what the agent committed to doing before it started. Kill Threshold Governance requires the threshold to be committed before evidence exists, not reconstructed from logs after the fact. The distinction is architectural, not cosmetic.
Four human checkpoints per investigation is operationally demanding. Organizations under time pressure will be tempted to collapse checkpoints into a single final review. This is the highest-risk shortcut. The checkpoint before evidence gathering begins, which validates whether the question is worth investigating and whether the pre-registration is correctly formed, is the one that cannot be skipped; all others can be delegated to a trained reviewer. The first cannot.
An organization that adopts Investigation Provenance for new investigations but does not address existing AI conclusions in circulation will inherit Answer Debt from its past. Conclusions produced without provenance that continue to inform decisions accumulate liability regardless of what the new architecture does. The provenance inventory in Phase 1 exists to quantify this inherited debt and prioritize its resolution.
Investigation Provenance is not required for every AI use case. Applying it to content generation, summarization, or internal Q&A creates overhead without proportionate governance benefit. Reserve it for the four trigger conditions: downstream signature, regulatory exposure, irreversibility, and multi-step conclusion chains. Scope it tightly at the outset so it remains operationally sustainable.
Zorp [4] (zorp.dev, aviskaar/zorp on GitHub, MIT licensed) is the only open-source implementation of this architecture currently available. It is pre-alpha and describes itself honestly: two of six behavioral contracts resolved as compatible, four were inconclusive because the reference runtime failed its own eligibility floor. The team publishes all six results, including the ones that did not resolve. That is itself a demonstration of the principle the tool implements: a system that practices Kill Threshold Governance cannot selectively report only its successes.
What makes Zorp significant beyond its current feature set is its architectural premise. It is the first openly available system that treats investigation provenance as a first-class concern, not as a log artifact or an audit afterthought. Its six-table evidence record schema (tracks, preregistrations, experiments, metrics, checkpoints, validations) is a concrete implementation of the abstract requirement this article names. Its human checkpoint gates between every capability stage are the operational embodiment of Kill Threshold Governance. Enterprise AI teams evaluating evidence accountability architectures should treat Zorp's schema and checkpoint model as a reference design, even if they build proprietary implementations.
The patterns Zorp demonstrates are not specific to any one platform or research domain. They describe the missing accountability layer that sits between an AI agent and a human decision-maker in every regulated enterprise context. No existing governance framework (NIST AI RMF, ISO/IEC 42001, EU AI Act) formally requires this layer. That will change. Organizations that build it before regulators mandate it will have audit-ready AI pipelines when their peers are still designing them.