What is CheckIt?

.docx only

CheckIt surfaces objective signals in student writing—entropy, lexical variety, sentence-level stylometry, and document metadata—so you can spot outliers, compare against year-level expectations or your cohort baseline, and target feedback. It informs moderation; it doesn’t replace professional judgement.

1. Upload

Drop a student draft or submission to analyse structure, vocabulary, and metadata.

2. Review Flags

High entropy, uniform sentences, atypical lexical spread, or abrupt style shifts are highlighted with explanations.

3. Act with Context

Use indicators alongside drafting evidence, task design, and teacher judgement in moderation.

Moderation
Quickly locate paragraphs with unusual complexity or vocabulary patterns.

Academic integrity
Combine text signals with revision/time metadata to focus follow-ups.

Cohort baselines
Build class/school baselines (≥10 revisions > 2 hrs) and flag > mean + 1σ.

Privacy
Files are processed transiently; references are detected and excluded.

Note
Indicators ≠ verdicts. 30 MB file size limit.

Accepted: .docx files, or a .zip that contains only .docx files. 30MB file limit.

+ Exclude student response template text (optional)

Use documents with ≥10 revisions and >2 hours editing time as entropy baseline

Upload Student Template

Year Level (optional):

Unspecified 7 8 9 10 11 12

Entropy/corpus baselines use pre-AI student work and are cohort-dependent. Use alongside professional judgement.
No uploaded files are retained or stored by LeaveLens. The files are temporary and removed after processing. No record/log of the analysis is retained.

Drag & drop files here

or click to select files

Quick Guide for Teachers

Signal	What it may indicate	Why it matters	Suggested follow-up
Entropy (bpw) High entropy	Complex/unusually unpredictable wording; beyond human expected range for selected Year Level, or above cohort baseline (mean + 1σ).	May reflect text stitched from varied sources or machine-like lexical spread.	Ask student to explain word choices; request short rewrite on a similar prompt in class.
Lexical TTR/HLR/OOV outliers	Low lexical variety, low hapax (few unique one-offs), or very low OOV vs corpus.	Uniform vocabulary and low novelty can be AI-like or indicate pasted content.	Probe drafting process; compare with earlier work; set an in-class paragraph task.
Lexical JS divergence extremes	Very low (near-clone) or very high (nonsense/overly strange) distribution vs corpus.	Can signal copying or artificially varied/odd token distribution.	Check for sources; discuss planning notes and research trail.
Stylometry Sentence uniformity	Zero/very low variance in sentence lengths; identical token sets across sentences.	Machine text often has “even” sentence rhythm and repetition patterns.	Ask for outline/rough draft; do a timed in-class rewrite to compare style.
Formatting Em-dashes / hidden runs / font shifts	Excess em-dashes; zero-width spaces; hidden text; abrupt font family/size changes.	Common markers of pasted or manipulated content; may mask edits.	Request the working file with tracked changes; discuss drafting habits.
Metadata Low revisions / low time	Very few revisions or minutes; creation and last save close together.	Suggests paste-in or minimal drafting on device.	Ask for planning artefacts; require iterative drafts or version history.
Baseline Cohort benchmark	Paragraphs exceed your cohort’s high-effort baseline (mean + 1σ).	Flags writing that is atypically complex for your school/class context.	Re-contextualise task expectations; check if support/extension is needed.
Spelling US→AU variants	American spellings (e.g., organize) where AU style is expected.	AI defaults to US English unless steered; also seen in pasted sources.	Reinforce AU conventions; ask student to justify style choices.
References Ignored sections	Bibliography/URLs detected and excluded from scoring.	Prevents inflated complexity from citations and formatting noise.	No action; verify citations are genuine and appropriately used.

CheckIt analyses complexity and the likelihood that writing exceeds age-appropriate expectations, as well as potential AI-assist signals. It does this using a blend of entropy measures, corpus-based baselines, and lexical/stylometric features.

1) Entropy (bits per word)

What it is: Entropy approximates how unpredictable the next word is within a paragraph. We report a bits-per-word figure derived from n-gram distributions.
Why it matters: Human writing at typical school levels usually falls into a predictable band. Very high entropy can indicate unusually complex, eclectic or stitched text, while very low entropy can signal repetitive or templated language.
How we use it: For each paragraph, we compare entropy against human expected ranges (by Year Level & Text Type when selected) and optionally against a baseline built from your own high-effort documents (≥10 revisions >2 hrs).

Note: We ignore references and normalise hidden spacing so results aren’t skewed by bibliography or formatting tricks.

2) Corpus baselines

Reference corpus: We load a level-appropriate corpus to anchor vocabulary statistics (type distributions, out-of-vocabulary rate, etc.).
Your baseline (optional): If enabled, we compute a school/class baseline from uploaded human-effort docs (revision/time floors). Paragraphs exceeding mean + 1σ on entropy vs this baseline are flagged as “outside expected range”.

3) Lexical & stylometric checks

Lexical variety: Type-Token Ratio, hapax (singletons) ratio, and out-of-vocabulary rate benchmarked to the corpus. Extreme values can indicate machine-like uniformity or stitched content.
Readability & sentence variation: We examine sentence-length spread (burstiness) and compare stylometric features across the document to detect uniformity anomalies.
Formatting cues: Excessive zero-width spaces, hidden runs, abrupt font changes or pasted chunks (e.g., em-dash patterns) may suggest non-native composition.

4) Interpreting the score

Non-metadata score: Uses only paragraph-level text signals (entropy, lexical, stylometric, formatting) — helpful when revision metadata is sparse.
Overall score: Blends text signals with document metadata (revision count & time spent). Low edit time and very few revisions can increase the score.
Flags vs verdicts: Flags are indicators, not proofs. Use flagged paragraphs to guide professional judgement and follow-up with students.

Tip: Select a Year Level to apply a tighter “human expected” range. Enable Use baseline to compare against your cohort’s authentic high-effort writing.

Example: When is a paragraph flagged?

A paragraph may be flagged if its bits-per-word is above the Year-Level human range and above your baseline mean + 1σ, or if multiple lexical/stylometric indicators fall outside typical human bounds simultaneously.