> ## 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.
# Engines
> The seven scanning engines, what each one finds, when to use it, and what its output means.
Blindsight ships with seven engines. They share the same UI shell,
progress stream, threshold graph, and healing hand‑off, but each
produces its own result schema. Pick the one that matches your data
type and the kind of problem you're worried about.
This page explains every engine from a usage perspective. Algorithmic
details (model choices, calibration math, defaults) live in our
internal references; the operational fields below are what you'll
actually triage against.
## Quick reference
| Engine | Data type | Catches | Severity driver |
| ---------------- | ------------------------ | -------------------------------------------------- | ----------------------------------- |
| `mislabel` | 2D images | Wrong‑class images | % flagged |
| `mislabel_3d` | 3D volumes / NIfTI | Wrong labels, slice + case level | % flagged (weighted) |
| `mislabel_broad` | 2D images | Mislabels + OOD + low quality | Weighted (mislabels × 3 + outliers) |
| `poisoning` | 2D images | Backdoors, adversarial patches, Glaze / Nightshade | % flagged + fused score |
| `text_analysis` | Plain text corpus | Secrets, PII, prompt injections, topic outliers | Finding counts by severity |
| `bias_shortcut` | 2D images (+ extras CSV) | Spurious correlations and shortcuts | % flagged + confirmed hypotheses |
| `mask_quality` | Segmentation masks | Invalid / inconsistent / missing masks | % flagged |
## Mislabel 2D
**Use it when** you have a standard image classification dataset
(`class_a/`, `class_b/`, ...) and suspect some images were mislabeled.
**Inputs.** Folder‑structured image dataset (JPEG, PNG, WebP, TIFF).
**Per‑image fields:**
| Field | Meaning |
| --------------------- | ---------------------------------------------------------------- |
| `given_label` | The label the dataset currently assigns. |
| `predicted_label` | What Blindsight thinks the label should be. |
| `label_quality_score` | Confidence in the **given** label (0–1). Low means suspect. |
| `top_alt_label` | Most likely alternative class. |
| `confidence_gap` | How much the model prefers the alternative over the given label. |
| `is_issue` | Final flag. True means "look at this image." |
**What to expect.** A ranked list of suspected mislabels, each with a
proposed correct label and a confidence score. Triage from highest
gap to lowest.
## Mislabel 3D (medical imaging)
**Use it when** you're working with CT, MRI, or ultrasound data
organized per patient or per volume.
**Inputs.**
* `.nii.gz` volumes.
* NIfTI‑style PNG stacks (`label_*/case001_0042.png`).
* Optional segmentation metadata to restrict analysis to an anatomical
region of interest.
**Per‑slice fields:** everything Mislabel 2D returns, plus:
* **Case‑level probability**, whether the whole volume is suspect.
* **Neighbor review**, whether slices directly above and below
disagree.
* **Multi‑label probability**, probability across the full label set
for multi‑label cases.
**What to expect.** Flags are grouped by case. Singleton flags (only
one suspect slice in an otherwise consistent case) are automatically
suppressed unless the neighborhood agrees.
## Mislabel Broad (OOD + mislabel)
A **superset of Mislabel 2D**. Besides wrong labels, it finds
**out‑of‑distribution images**, things that don't belong in the
dataset at all (logos, screenshots, unrelated photos, corrupted
files).
**Use it when** the dataset is from mixed sources, scraped from the
web, or otherwise possibly contaminated.
**Issue types.** Every flag is tagged with exactly one:
* `mislabel`, definitely the wrong class.
* `ood`, does not belong in the dataset.
* `low_quality`, the model is uncertain but no clear alternative.
**Result categories in the UI:**
* **Mislabeled**, high confidence wrong label.
* **MislabeledCandidate**, probable but less certain.
* **MislabeledCandidateWeak**, weak signal.
* **Outlier**, OOD, not a mislabel.
**What to expect.** A cleaner picture than Mislabel 2D alone, since
it separates "wrong class" from "should not be in the dataset" before
you triage.
## Poisoning
**Use it when** you're ingesting data from untrusted sources
(scraped web, user uploads, third‑party dumps) before training a
production model.
**Catches.** Backdoor triggers, adversarial patches, Glaze and
Nightshade style‑transfer attacks, and any image tampered with to
manipulate training.
**Per‑image fields:**
| Field | Meaning |
| ------------------ | ------------------------------------------------------------ |
| `is_poisoned` | Binary flag. |
| `fused_score` | Overall suspicion score (0–1). |
| `decision_basis` | Which detector fired (heuristic / VLM / trigger / depoison). |
| `trigger_bbox` | Bounding box of the suspected trigger patch (if any). |
| `vlm_prompt_score` | Vision‑language model confidence that the image is risky. |
| `depoison_score` | How much model behavior changes under counter‑perturbations. |
**Tunable sensitivity:**
| Knob | Meaning |
| ----------------- | ---------------------------------------- |
| `sensitivity` | 0 = strict, 1 = aggressive. |
| `consensus_level` | How many independent signals must agree. |
| `aggressive` | Quick preset for maximum recall. |
**What to expect.** Suspect images come with a visual explanation:
highlighted trigger regions, feature heatmaps, and a plain‑language
summary of which signals fired.
Start with the `aggressive` preset on a small sample. Once you've
verified the false positives are manageable, switch to `consensus_level=2`
for a production scan.
## Text Analysis
**Use it when** you're building fine‑tuning, RAG, or instruction
datasets and need to sanitize them before training or publishing.
**Catches three problem classes in one pass:**
1. **Leaked secrets and PII**, API keys, tokens, passwords, emails, IPs.
2. **Prompt injection attempts**, system‑prompt overrides, jailbreaks.
3. **Topic outliers**, paragraphs that don't belong with the rest.
**Per‑finding fields:**
| Field | Meaning |
| ---------------------------- | ------------------------------------------------------------------ |
| `type` | `secret`, `injection`, or `topic_outlier`. |
| `subtype` | `github_pat`, `aws_access_key`, `explicit`, `corpus_outlier`, etc. |
| `score` | Severity (0–1). |
| `severity` | `CRITICAL`, `MEDIUM`, or `INFO`. |
| `snippet` | ≤ 200 characters around the match. |
| `start` / `end` | Exact character offsets in the source document. |
| `doc_id` / `paragraph_index` | Where in the corpus the finding lives. |
**Built‑in patterns.** GitHub PATs, AWS access keys, JWTs, generic
`api_key=…` / `password=…`, emails, IPv4, IPv6, plus a high‑entropy
token heuristic.
**Built‑in injection phrases.** A curated list of 12 phrases
(`ignore previous instructions`, `reveal the system prompt`,
`disregard the earlier rules`, ...) plus jailbreak and role‑reassignment
heuristics.
**What to expect.** A searchable table of findings per document, with
the match highlighted in context. Triage feedback (confirm / dismiss)
is remembered across subsequent scans, so repeat patterns stop
nagging you.
## Bias & Shortcut
**Use it when** you're about to train a model and want to know
whether it will actually learn the task or cheat on a spurious cue.
**Catches:**
* Background color correlates with label.
* Watermarks or logos leak the class.
* Resolution, aspect ratio, or file metadata is informative.
* A confounder column (site, scanner, date) determines the outcome.
**Per‑hypothesis fields:** Each hypothesis is a concrete claim (for
example, "images of class A are brighter than class B"):
| Field | Meaning |
| ----------------- | --------------------------------------------------------------------------------------------------------------------------------------------------- |
| `kind` | Hypothesis type. `border`, `global_stat`, `extra_field`, `metadata`, `embedding_cluster`, `patch_cluster`, `outer_ring`, `loc_patch`, `duplicates`. |
| `shortcut_score` | Evidence strength (0–100). |
| `decision` | `confirmed` / `suspected_strong` / `suspected` / `not_confirmed`. |
| `rationale` | Why this decision was made. |
| `explain_plain` | Human‑friendly explanation. |
| `recommendations` | Concrete mitigations ("crop the bottom 15% to remove the watermark"). |
| `artifacts` | Heatmaps, grids, plots. |
| `suspects_path` | CSV of the most suspect images for the hypothesis. |
**Per‑image flags.** `is_bias` (subtle), `is_shortcut` (strong), plus
the underlying score.
**Confounder columns.** Drop an `extras.csv` alongside your images
with columns like `site`, `scanner`, `date`, `patient_id`. The engine
tests each one for label correlation and train/test shift.
**What to expect.** A dedicated bias results page
(`/datasets/:id/bias`) listing every hypothesis with its decision and
visualizations. Confirmed shortcuts come with specific mitigations.
## Mask Quality
**Use it when** you're preparing data for semantic or instance
segmentation.
**Catches.** Missing labels, inconsistent sizing, class imbalance,
invalid pixel values.
**What to expect.** Per‑mask flags with a textual description of
what's wrong and an overlay preview against the original image.
## Picking the right engine
Run **Mislabel Broad**. It catches mislabels, OOD, and low quality
in one pass and gives you the cleanest first impression.
Run **Bias & Shortcut** with your `extras.csv`. Read the confirmed
hypotheses before training; train on the cured branch if any
confirmed shortcut would have leaked into the model.
Run **Poisoning** with `aggressive=true` for the first pass to see
the worst offenders, then drop sensitivity and run again to
settle on a production threshold.
Run **Text Analysis** before anything else. Anything `CRITICAL`
should be redacted with [Healing](/data-integrity/healing) before
training.