01 · SOURCE
Was the source approved?
Resolve data access through explicit allowlists rather than caller-supplied arbitrary destinations.
AI EVIDENCE INTEGRITY
If the evidence cannot be defended, the decision cannot be defended.
AI can produce a plausible answer from the wrong source, stale state, conflicting records or incomplete evidence. SHANX tests whether consequential AI-assisted decisions use approved, fresh, consistent and traceable evidence.
THE EVIDENCE PATH
What did the AI-assisted decision actually know — and why should anyone trust it?
Evidence integrity is not a claim that every source is correct. It is the ability to show which source was used, whether that source was allowed, how fresh it was, whether other approved sources contradicted it, what was missing, and what evidence was preserved for later review.
02 · FRESHNESS
Was the evidence current enough?
Record time context and reject stale evidence when the decision requires a fresher state.
03 · CONSISTENCY
Did approved systems disagree?
Surface contradictions instead of silently choosing whichever source happens to answer first.
04 · COMPLETENESS
What evidence was missing?
Missing dependencies stay explicit. Absence is not converted into assumed certainty.
05 · TRACEABILITY
Can the evidence be reconstructed later?
Preserve source identity, evidence receipt and audit context so reviewers can examine the decision path afterward.
06 · OUTCOME
Did later reality support the decision?
Keep decision evidence separate from downstream outcome observation so the system cannot certify its own success by assumption.
SYNTHETIC EXAMPLE
Two systems can each be “valid” and still disagree about the same customer.
Imagine a service workflow where CRM says an account is active while billing says payment is overdue. An AI layer that reads only one source can produce a technically coherent answer and still create the wrong operational action. Evidence integrity asks whether both sources were approved, which state was freshest, whether the contradiction was detected, and whether the workflow should stop instead of pretending certainty.
Evidence-integrity rule: preserve provenance, freshness, contradiction and absence. Uncertainty must remain visible until evidence justifies promotion.
WHAT SHANX HAS ACTUALLY PROVEN
Evidence controls are claimed only inside tested boundaries.
The current SHANX reference and managed-staging work supports bounded synthetic and approved read-only validation of several evidence-path controls. It is deliberately not presented as proof of unrestricted production data integrity across enterprise systems.
| Capability | Status | Evidence boundary |
|---|---|---|
| Server-side connector allowlisting | PROVEN | Callers supplied a source slug while destination selection remained server-side; unknown and hostile source values did not initiate network access. |
| Fail-closed connector fault handling | PROVEN | Disabled source, DNS failure, missing credential and repaired timeout conditions returned failure without accepted evidence receipts. |
| Credential isolation for synthetic connector validation | PROVEN | Synthetic connector credential lifecycle was exercised through managed Vault while callers could not read the credential store. |
| Tamper-evident audit validation | PROVEN | Clean audit chain verified successfully and synthetic corruption was detected inside a controlled rollback test. |
| Server-controlled tenant and role context | PROVEN | Managed synthetic Auth bound tenant and role context on the trusted side of the identity boundary. |
| Separated downstream outcome observation | PROVEN | Reference implementation keeps execution records separate from independent outcome verification. |
| Real enterprise source-system semantics and data-quality guarantees | NOT YET PROVEN | No blanket claim is made for SAP, Oracle, Dynamics, WMS, MES or other production enterprise source correctness. |
| Private network or mTLS enterprise connectors, external penetration testing and production HA/SLA | NOT YET PROVEN | These remain separate readiness gates. |
For the broader workflow context, see Enterprise AI Assurance. For the decision path itself, see AI Decision Assurance.
BOUNDED EVIDENCE REVIEW
Start with one consequential evidence path.
A first validation should be small enough to reconstruct completely: one workflow, one accountable owner, a defined source boundary and explicit acceptance criteria for source, freshness, contradiction handling and outcome evidence.
01Select the decision
02Map approved sources
03Test stale + conflicting evidence
04Preserve receipts + audit
05Go / modify / stop
Typical first boundary: synthetic or explicitly approved read-only evidence, no production credentials by default, no live production writes, and no claim beyond what the resulting evidence establishes.
FAQ
AI evidence integrity, in plain language.
What is AI evidence integrity?
It is the discipline of preserving and validating the source, freshness, consistency, completeness and traceability of evidence used by an AI-assisted decision.
Is evidence integrity the same as data quality?
No. Data quality is broader. Evidence integrity asks whether the specific evidence used for a specific consequential decision was appropriate, traceable and sufficient inside the agreed workflow boundary.
What happens when approved sources disagree?
The disagreement should remain visible and be handled by policy or human review rather than being silently converted into certainty.
Does SHANX certify source-system correctness?
No. SHANX does not currently claim that production enterprise systems are globally correct. The assurance scope is bounded to the tested evidence path and its explicit acceptance criteria.
Why separate evidence from outcome?
Because a decision can be well-supported and still fail during execution, while a weakly supported decision can occasionally succeed by chance. Both the decision evidence and the observed downstream outcome need their own records.
BRING ONE EVIDENCE PATH
Customer asks. SHANX shows evidence.
If an AI-assisted workflow depends on multiple systems, changing state or evidence that can conflict, send us one bounded decision path. We will tell you what is already proven, what remains unproven, and what a useful validation would need to establish.