A silently corrected hallucination isn't fixed. It's erased.
Some automated correction tools patch a hallucination or contradiction and quietly move on, leaving no record of what conflicted, what won, or why. SAGE's two-tier mechanism resolves provable conflicts automatically and logs them, and escalates genuinely ambiguous conflicts to a human whose ruling is recorded and inherited by every connected tool. Never a silent patch.
A hallucination gets caught, quietly patched, and the conversation moves on as though nothing happened. That's an increasingly common design pattern: treat a wrong claim the way a spellchecker treats a typo. Flag it, fix it, keep going, and make sure nobody on the other end ever has to see the mess.
It's a useful trick for grammar. It's a stranger promise to make about the fact your AI just gave a staff member about their leave entitlement, because a silently corrected wrong answer and a right answer look identical from the outside. Only one of them comes with a reason.
The trail that disappears
With some automated correction tools, invisibility is the design, not an oversight. Because the fix is built to be invisible to the user, it's usually invisible to everyone else too. There's no separate record of what the AI first claimed, what replaced it, or which source settled the disagreement. Just the corrected answer, delivered as though it had always been the only one. The session ends, the log rolls over, and the only evidence a conflict ever existed was a flicker nobody was watching for.
That works fine for a typo. It's a bad outcome for a claim that reached a resident's care plan or an employee's payslip. The correction and its reasoning are exactly what someone will need to reconstruct later, not the bare fact that a correction happened.
What auditors actually ask
Nobody auditing a regulated AI deployment asks whether the final answer happened to land correctly. They ask what the two conflicting facts were, which one the system decided was current, and on what basis. A silent auto-correction has nothing to offer any part of that question, because "what conflicted, what won, and why" was never designed to outlive the moment it got resolved.
Two tiers, not one
SAGE decides, before resolving anything, whether that resolution needs to be checkable, and treats every resolution as worth keeping either way. Where lineage, supersession, or authority settle a conflict outright (a version that explicitly supersedes the one before it, a source the organisation has already ranked higher for that topic), SAGE resolves it and logs the resolution: which relationship closed it, what evidence supported that, and what the superseded fact used to say. Where nothing settles it, two sources equally current, neither outranking the other, SAGE doesn't pick one. It escalates to a person, writes their ruling back as its own lineaged fact, and every connected tool inherits that ruling the next time it asks.
Not a patch. A decision, on the record either way.
The record outlives the fix
The distinction that matters isn't automatic versus human. Plenty of SAGE's resolutions happen without a person in the loop, by design, because provable conflicts don't need anyone's judgment. The distinction is whether the correction leaves behind something an auditor can pick up eight months later: what was believed, what replaced it, which rule or which person decided, and when. Resolved automatically or resolved by a ruling, that reasoning is retired, not deleted. Still there to check.
The honest version of contradiction detection was never going to be the quiet one. It's built on the assumption that someone, eventually, will ask what happened, and that the answer needs to already be sitting there waiting.
Loriq builds SAGE, the governed memory engine. Talk to us.