How Rationale reads your artifacts

1. Ingestion

Artifacts arrive from a connected source or a direct upload. Rationale fingerprints each one, so a trivial edit (a typo fix, reformatting) doesn't trigger a full re-read - only meaningful changes do. Before anything else happens, each document is screened locally for sensitive content - flagged documents are held for your review and are never sent to an AI provider unless you approve them.

2. Understanding

The document is segmented and read for its reasoning: the decisions being made, the problem each one answers, the evidence offered, the assumptions it rests on, the alternatives that were considered, and the tradeoffs that were accepted. This is where Rationale differs from search - it extracts why, not just what.

3. Curation

Extracted knowledge is resolved against what the map already holds. The people, systems, and projects mentioned are matched to known entities (so "the billing service" and "billing-svc" are one thing, not two), duplicate documents are recognized and set aside before they can inflate anything, and each decision tracks how many distinct documents support it.

Your judgment is part of curation too. You can verify, correct, or delete anything Rationale extracted - a correction outranks what re-reading the document says, and deletions are remembered, so a future sync never resurrects what you removed. AI output becomes trusted organizational knowledge only after it survives your review, never automatically.

4. The map

Decisions become connected objects: linked to the evidence that informed them, the decisions they depend on, and the ones they replaced. When a new artifact changes the picture - a decision is revisited, an assumption breaks - Rationale re-reasons the affected part of the map. Superseded decisions are never deleted; they keep their reasoning and their place in the timeline, so you can always see what your organization believed before, and why it changed.

5. Provenance

Every claim in Rationale traces back to the artifact it came from, and every decision shows the support behind it - single source, corroborated by distinct documents, or verified by a person. That vocabulary is deliberate: Rationale shows measured support instead of a numeric confidence score, because a percentage would imply a precision the extraction doesn't have. Nothing is asserted from thin air: you verify against the original document, not against a summary. That's also why nothing you add is ever used to train AI models - the map is your organization's asset, not ours. More in our privacy policy.