Methodology & evidence

Inside a correction cycle: from detection to verified persistence

October 24, 2026 7 min read
Kamel Malek Director of SetRecord.ai

When we say "correction" in Deep Narrative Intelligence, we do not mean a single act of publishing a counter-statement. We mean a cycle with five stages — detection, provenance tracing, correction, propagation, and persistence verification — each of which can fail independently, and each of which must succeed for the correction to hold. This article walks through the full cycle, so you understand what happens between "we found an error" and "the error is gone and stays gone."

The cycle at a glance

A correction cycle begins when a DNI interrogation panel flags a factual error or narrative distortion in a model's answer. It ends when a subsequent measurement cycle confirms that the error no longer appears across the model panel, in both search and no-search modes, and has not reappeared for at least 90 days. Between those two points, five stages must be executed in sequence, and each stage has its own failure modes that the programme must guard against.

Stage 1: Detection

Detection is the output of the interrogation panel: a specific model, a specific query, produced a specific answer containing a claim that does not match the verifiable record. The detection is not a vague "the model said something wrong" — it is a structured record: the query, the model, the mode (search/no-search), the claim, the correct fact, and the source of the correct fact. Without this structure, the correction cannot be targeted. The most common detection failure is imprecision: flagging a "negative tone" without identifying the specific claim that drives it, which makes downstream stages impossible to execute.

Don't know what AI says about you today?

A SetRecord audit queries supported AI models about your organisation and measures the gap with verifiable facts — within five business days.

Request an audit

Stage 2: Provenance tracing

Once a distortion is detected, the next question is: where did the model get this? Provenance tracing uses indirect methods to infer the likely source. If the error appears in no-search mode, it is in the training data; if it appears only in search mode, it is in a retrievable web source. Comparing across models with known training cut-offs can narrow the source further. The output of this stage is a hypothesis about which source or source class the model is drawing from — a specific article, an aggregator entry, a regulatory filing that has been misread. The most common provenance failure is stopping at "the model is wrong" without identifying the source, which makes correction a shot in the dark.

Stage 3: Correction

The correction is the creation or update of an authoritative source that contradicts the distortion with evidence. This is not a press release saying "we are right and the model is wrong"; it is a structured, dated, verifiable piece of content that a model's training pipeline or retrieval system can ingest. It must be precise (citing the exact correct fact), authoritative (published on a channel the model trusts), and structured (machine-readable where possible). The correction is co-developed with the client's subject-matter experts to ensure it does not overstate or introduce new inaccuracies. The most common correction failure is vagueness: a correction that says "our company is doing well" when the specific error was "the model said our revenue declined" is useless because it does not contradict the specific claim with a specific fact.

Stage 4: Propagation

Propagation is the process by which the correction enters the model's answer. For web-search mode, propagation can be fast — days to weeks — because the retrieval system can pick up the new source quickly. For no-search mode, propagation requires the model to be retrained, which happens on the lab's training cycle (typically 3–6 months). During the propagation stage, the programme monitors whether the correction has begun to appear in the model's answers. A propagation failure occurs when the correction is published but the model does not pick it up — often because the source is not structured or authoritative enough for the model's pipeline to prioritise it.

Stage 5: Persistence verification

Persistence is the final and most important stage. It asks: has the correction held? A measurement cycle is run 30, 60, and 90 days after the correction was first confirmed. If the error reappears at any point, the cycle restarts from provenance tracing — the correction did not hold, and a new hypothesis about the source is needed. If the error does not reappear across 90 days and across both search and no-search modes, the correction is marked as persistent. SetRecord treats persistence verification as the defining metric of programme success: a correction that does not persist is not a correction — it is a temporary suppression, and the programme is not done until persistence is verified.

See what AI says about you

A SetRecord audit checks what supported AI models say against your verifiable facts — within five business days.