> ## Documentation Index
> Fetch the complete documentation index at: https://docs.blindsight.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Healing

> Turn a scan's findings into a cured copy of the dataset. Drop outliers, relabel mislabels, redact secrets, and keep a full lineage trail.

Healing takes a scan's findings and produces a **cured copy of the
dataset** with the bad samples either fixed or removed. You choose
which fixes to apply; every action is captured in the audit trail.

## When to heal

* After a `mislabel` or `mislabel_broad` scan, to relabel suspects and
  drop OOD samples.
* After a `poisoning` scan, to drop tampered samples before training.
* After a `text_analysis` scan, to redact secrets and remove
  injection paragraphs from a fine‑tune corpus.
* After a `bias_shortcut` scan, to apply the recommended mitigations
  (cropping, dropping, rebalancing).

## What the toggles do

| Toggle                 | Effect                                                      |
| ---------------------- | ----------------------------------------------------------- |
| **Fix mislabeled**     | Rewrite labels according to the engine's predictions.       |
| **Remove outliers**    | Drop OOD or outlier samples entirely.                       |
| **Remove poisoned**    | Drop samples flagged by the poisoning engine.               |
| **Remove low quality** | Drop borderline low‑confidence cases.                       |
| **Confidence floor**   | Only act on findings above the chosen confidence threshold. |

Toggles are independent. You can drop poisoned samples without
touching mislabels, or relabel mislabels without dropping anything.

## How to heal

<Steps>
  <Step title="Open the source">
    Start from either the dataset detail page (**Heal**) or the scan
    detail page (**Heal from this scan**). Healing from a scan
    pre‑fills the dialog with that scan's findings.
  </Step>

  <Step title="Pick what to apply">
    Flip the toggles you want. Set a **confidence floor** if you only
    want to act on the highest‑confidence findings.
  </Step>

  <Step title="Pick the output">
    Healing always writes a result. You can choose:

    * **New child dataset**, the original stays untouched. A lineage
      edge `healed_from` links the child to the source.
    * **New branch on the source**, the cured contents live as a
      branch you can switch between.

    A zip download is always offered alongside.
  </Step>

  <Step title="(Optional) Auto‑rescan">
    Toggle **Auto‑rescan after healing** to immediately run the same
    engine on the cured output. Useful for verifying the heal moved
    the dataset out of `CRITICAL`.
  </Step>

  <Step title="Submit and watch">
    Healing runs as a background job with its own progress stream.
    The result appears in the dataset's lineage graph.
  </Step>
</Steps>

<Tip>
  When healing for the first time on a noisy dataset, set the
  confidence floor high (e.g. 0.9) and inspect the cured branch
  before relabeling everything. Lower the floor incrementally once
  you trust the engine's calls.
</Tip>

## Healing for text

Text healing is different from image healing because the unit of repair
is the snippet, not the file.

| Action                  | What it does                                                                                                   |
| ----------------------- | -------------------------------------------------------------------------------------------------------------- |
| **Redact secrets**      | Replace every detected secret span with `[REDACTED_<type>]` in place. The original document remains otherwise. |
| **Strip injections**    | Remove paragraphs flagged as injection attempts.                                                               |
| **Drop topic outliers** | Remove paragraphs flagged as off‑topic.                                                                        |

Redacted documents keep their original `doc_id` and offsets so the
audit trail can reconstruct what was changed.

## What you get back

Every healing run produces:

* A **new dataset or branch** with the cured contents.
* A **zip download** of the result.
* An **action record** in the audit trail (who triggered it, what
  toggles were set, which scan it descended from).
* A **lineage edge** linking the cured output back to the source
  dataset.
* *(Optional)* a follow‑up scan if you enabled auto‑rescan.

## Common workflows

<AccordionGroup>
  <Accordion title="Clean up a public dataset before training">
    1. Run `mislabel_broad` and `poisoning`.
    2. Heal with **Remove outliers**, **Remove poisoned**, and
       **Fix mislabeled** at confidence floor 0.85.
    3. Enable **Auto‑rescan** and confirm the result is `HEALTHY` or
       `UNHEALTHY−`.
    4. Train on the cured branch.
  </Accordion>

  <Accordion title="Sanitize a fine‑tune corpus">
    1. Run `text_analysis`.
    2. Heal with **Redact secrets** and **Strip injections**.
    3. Re‑run `text_analysis` to confirm zero `CRITICAL` findings
       remain.
    4. Export the cured corpus.
  </Accordion>

  <Accordion title="Iterate on a noisy labelling effort">
    1. Run `mislabel`.
    2. Heal **only the top 10%** of suspects (high confidence floor)
       into a new branch.
    3. Have a reviewer compare the original and cured branches side
       by side.
    4. Lower the confidence floor for the next pass once the
       reviewer signs off.
  </Accordion>
</AccordionGroup>

## What healing does **not** do

* It does not modify the original dataset unless you explicitly pick
  "main branch" as the output target.
* It does not run a fresh scan automatically (you have to opt in via
  **Auto‑rescan**).
* It does not delete the source scan, so you can re‑heal with
  different toggles at any time.