Methodology Overview
SpecShift uses a structured, observable-only review process to examine whether an AI workflow's claims, evidence, and visible behavior align.
The methodology is designed to support engineering, product, and risk teams before broader deployment. It does not require model internals, source code, private chain-of-thought, or hidden system access to begin.
The review is bounded by the agreed workflow scope and produces structured observations for human decision-makers.
High-Level Review Process
1. Define the workflow
Identify one concrete AI workflow, its intended purpose, and the claims being reviewed.
2. Define the review boundary
Establish which observable materials are available, what is outside the review scope, and what questions the review is intended to answer.
3. Review observable evidence
Examine the workflow's visible claims, evidence, outputs, timestamps, handoffs, and completion behavior.
4. Apply bounded bounded synthetic review scenarios
Review how the workflow responds when evidence becomes incomplete, stale, conflicting, or requires escalation.
5. Document candidate discrepancies
Record situations where observable claims appear stronger than the reviewed evidence or workflow behavior.
6. Return structured review artifacts
Provide candidate-discrepancy memos, workflow observations, and recommended engineering review questions.
7. Human review remains in control
SpecShift does not approve deployment. Engineering, product, and risk teams determine what to test, revise, narrow, or escalate.
Methodology Boundaries
SpecShift reviews observable claims, evidence, workflow behavior, and candidate discrepancies within an agreed review scope.
The methodology is not designed to inspect model weights, private chain-of-thought, hidden prompts, source code, production infrastructure, or other protected implementation details unless explicitly included within the agreed review scope.
The purpose of the methodology is to produce structured review artifacts that support engineering, product, and risk teams—not to certify AI systems or approve deployment.