Service Introduction

rawctx is an answer evidence layer that gives evidence of what official criteria, sources, and model execution contexts the AI responses to customers are based on. Enterprises can keep and verify answers from customer-facing chatbots, AI analysts, and copilots as question hash, answer hash, semantic refs, and prof bundle rather than just logarithms. rawctx does not guarantee the truthfulness of the answers, as it is not the layer to judge the model's outcome. Instead, it shows what criteria the answers were bound to, have not changed since, and can be verified by external anchors and witnesses. This makes it easier for security, law, customers, and auditors to review AI answers and obtain the evidence needed to approve them.

---

The question we're trying to solve

When companies introduce customer-facing AI, answers are generated quickly, but it is difficult to explain what criteria and grounds the answers were made later. If the same KPI or policy is interpreted differently for each document, database, or prompt, and there is no standard version, source, or model execution context left at the time of the answer, reproducible audits are impossible during security review, legal response, and customer issue processing. It is difficult to objectively prove whether the answer has not changed since then or on what basis it is actually tied to with only internal logs. The problem rawctx is trying to solve is not to judge on behalf of AI, but rather a structural void that does not leave AI answers to customers as evidence that companies can verify, explain, and submit later.

---

A solution to a problem

Rawctx solves the problem by leaving the evidence and execution context that disappear after AI answers are generated as verifiable evidence. First, it versions the company's KPI definitions, policies, and document sources in a semantic package and records what criteria the AI referred to when answering them in the answer audit log. Then, the original text of the question and answer is sealed with a hash, and the execution context, such as the model and settings, and input/output hash, is grouped together to create a prof bundle. This evidence is connected to external verification paths such as Merkle, STH, OpenTimestamps, and Rekor to check for changes and confidence. As a result, security, legal, and customer representatives can review and approve AI answers as submissionable reports without re-explaining them or manually tracking them.