# Drift

Handle changes in inputs, models, policies, evidence chains, and attacker behavior over time.

Source URL: https://trymighty.ai/docs/concepts/drift

Drift means risk changes after the first scan.

Your product may receive new file types. A model may change. OCR may produce different text. A policy may become stricter. Attackers may adapt. A reviewer may add new evidence.

When that happens, do not rely on an old scan as if the workflow is unchanged.

Drift flow: detect content, model, policy, or evidence change, then rescan or review before trust.

## Types Of Drift

| Drift type | What changes | Example |
| --- | --- | --- |
| Input drift | Users submit new formats, larger documents, more images, or different language. | A claims flow starts receiving multi-page PDF packets instead of one damage photo. |
| Modality drift | The same workflow shifts from text to image, PDF, audio transcript, or extracted fields. | A support chat adds file attachments. |
| Model drift | The model, prompt, tools, system message, or output format changes. | A chat assistant starts generating claim recommendations instead of summaries. |
| Policy drift | The product changes tolerance, routing, or review standards. | Public output moves from `standard` to `strict` data sensitivity. |
| Evidence drift | Derived text or summaries diverge from the original item. | OCR misses hidden instructions in a PDF, then an AI summary treats the extracted text as clean. |
| Adversary drift | Attackers learn where the controls are and change their payloads. | A prompt injection moves from chat text into an image or PDF layer. |

## When To Rescan

Rescan or send to review when:

- A file is transformed into OCR text, extracted fields, or a summary.
- A model output will be shown to a user or used by automation.
- A user edits and resubmits evidence.
- A new model, prompt, tool, or agent policy ships.
- Your routing policy changes from tolerant to strict.
- A reviewer adds evidence or marks a result as disputed.
- A scan returns `indeterminate` and the workflow receives new material.
- A batch runs over older evidence after a policy update.

## What To Keep Stable

| Field | How to use it during drift |
| --- | --- |
| `session_id` | Keep the chat, claim, case, customer, batch, or agent run together. |
| `scan_group_id` | Reuse it for the original item and derived scans from that item. |
| `request_id` | Use a new value for each new scan request. |
| `metadata` | Record model version, prompt version, workflow version, reviewer ID, and policy version when available. |

Use a new `scan_group_id` when the material is a new item. Reuse the existing group when the material is derived from the same item.

## Drift Routing

| Drift event | Recommended route |
| --- | --- |
| New OCR output from a scanned PDF | Scan OCR text with the same `scan_group_id`. |
| Model output from a scanned prompt | Scan output with `scan_phase=output` and the input `scan_group_id`. |
| New model or prompt version | Rescan public output paths and compare `WARN` or `BLOCK` rates. |
| New policy version | Rescan high-risk pending items or route old scans to review. |
| New evidence in a claim | Scan the new item with a new group, then attach it to the same `session_id`. |
| Weak or conflicting signals | Route `indeterminate` to review or request more evidence. |

## Drift Metrics

Track these fields over time:

- `action` rate by workflow.
- `WARN` and `BLOCK` rate by model version.
- `indeterminate` rate by modality.
- `risk_score` distribution by workflow.
- Review outcome versus Mighty action.
- False positive and false negative feedback from reviewers.
- SCU usage by workflow after mode changes.

These metrics help you decide when to adjust `mode`, `focus`, `profile`, `data_sensitivity`, async usage, or review routing.

## Common Mistakes

- Treating an old `ALLOW` as valid after OCR, model output, or policy changes.
- Creating a new session for each derived scan, which breaks audit history.
- Reusing a `scan_group_id` for unrelated evidence.
- Measuring only blocked scans and ignoring `WARN` or `indeterminate`.
- Letting model output into a public response because the original prompt was already scanned.

## AI-Agent Prompt

### Add drift handling

```text
Add Mighty drift handling to this product.

Requirements:
- Identify where content changes after the first scan.
- Rescan OCR output, extracted fields, model output, agent output, and edited evidence.
- Keep related derived scans on the same scan_group_id.
- Keep the wider chat, claim, case, batch, or agent run on the same session_id.
- Add metadata for model version, prompt version, workflow version, and policy version when available.
- Route indeterminate results to review or request more evidence.
- Track action rate, WARN rate, BLOCK rate, indeterminate rate, and review outcome over time.

Acceptance criteria:
- Old ALLOW results are not reused after meaningful content, model, or policy changes.
- Tests cover rescanning derived output.
- Review history can show how a result changed over time.
```