Observable-only review means reviewing what an AI system claimed, what it actually did, and what evidence supports those claims—without requiring model internals, source code, private chain-of-thought, or hidden system access.
SpecShift examines the visible record: claims, workflow behavior, evidence, handoffs, and outputs. When those do not line up, we produce a structured candidate-discrepancy memo for human review.
This approach is designed to help engineering, product, and risk teams evaluate AI workflows while respecting operational boundaries and protecting proprietary systems.
Why Observable-Only Matters
Many organizations cannot—or should not—share source code, model internals, private prompts, or proprietary system details with an outside reviewer.
Observable-only review focuses on the evidence that is already visible: what the AI claimed, how the workflow behaved, what information it relied on, and what outputs it produced.
This allows teams to perform structured review while respecting operational boundaries and maintaining control of sensitive systems.
Observable-only review is not intended to replace testing, monitoring, or engineering evaluation. It complements those activities by asking a different question:
Do the system's claims, evidence, and visible behavior line up?
What Observable-Only Review Produces
Observable-only review is designed to produce structured review artifacts for human decision-makers.
Depending on the engagement, outputs may include:
candidate-discrepancy memos
evidence-handling observations
workflow boundary observations
handoff and escalation observations
design-review questions for engineering teams
recommended next review steps
The purpose is not to replace engineering judgment. The purpose is to give engineering, product, and risk teams clearer evidence for deciding what to review, test, narrow, or escalate before broader deployment.