Text-layer reconstruction · workspace instrument

Take a fabricated review apart —
then rebuild the honest one underneath.

Paste any review. The analyser scores it against the text-layer signals documented in this workspace, marks the exact spans that make it read as manufactured, and then runs the transform backwards: persuasion removed, recount structure restored, and the facts only the reviewer could know left as explicit gaps.

14 signals · text layer only deterministic · no network calls weights illustrative
01 — Analyser

Paste a review

Analysis runs as you type or paste. The score covers text only; the real systems weigh behaviour and arrival patterns far more heavily.

Comparison corpus for duplication (optional)
Load a sample
—
Awaiting input · no text scored
— of 14 signals computable
Paste a review to see which families fire.
Structure
0%
Lexicon
0%
Discourse
0%
Stance
0%
Authorship
0%
Duplication
0%

Why it reads as fabricated

Ordered by contribution to the score. Every quote is taken from your text.

No analysis yet. Load a sample, or paste a review above — the findings appear here and the spans light up on the right.

The evidence, in place

persuasion inflation essay connective
Highlighted spans appear here once there is text to inspect.
—tokens
—sentences
—type–token
—past markers
02 — Reverse transform

The reconstruction

The same content, run backwards through the fabrication signals. Editable gaps are facts the tool refuses to invent for you.

rewrite risk — no rewrite yet
Rewritten review

The reconstruction appears here after analysis.

What changed, and why
No edits yet.
Word-level diff
The diff appears here after analysis.
restored removed
03 — Method

How the score is built

One formula, three stages. Everything is a deterministic function of the text you paste — no model, no sampling, no network.

R = 100 · Σ(cᵢ · wᵢ · riskᵢ) / Σ(cᵢ · wᵢ) cᵢ = 1 only where the signal is computable

1 · Decompose

Each sentence is measured on six axes: structure (length, rhythm, repetition, punctuation), lexicon (promotion, superlatives, evaluative density), discourse (specificity, temporal orientation), stance (rating–sentiment fit, friction), authorship (machine-writing cues) and duplication against a corpus you supply.

2 · Normalise by coverage

A ten-word review cannot support a lexical-diversity measurement. Rather than scoring that as safe, the signal is excluded and the score is computed over what remains, with coverage reported beside it. Short texts therefore lose precision, not truth.

3 · Invert

The rewrite applies the inverse of each signal that fired: persuasion is replaced with a reuse statement, superlatives deflate to plain adjectives, essay connectives are stripped, and the recount frame is rebuilt around explicit gaps for dates, objects, friction and resolution.

04 — Transforms

The reverse transform, signal by signal

Each rule exists because the corresponding signal fired. This is the operational answer to “what would this review look like if it were real?”

SignalWhat it caughtReverse transformExample
LEX-01Boilerplate persuasionReplace only the recommendation phrase with the reviewer's own reuse decision; the rest of the clause survives.I would highly recommend them to anyone → I'd use them again.
LEX-02Superlative stackingDeflate to a plain adjective. A real thing described plainly is more credible than the same thing praised maximally.absolutely fantastic → good
LEX-03Intensifier inflationDelete intensifiers outright; they carry no information the adjective lacks.so professional → professional
ORT-01Thin textKeep only sentences with content, then restore the recount frame that carries length honestly.Great service. Fast delivery. → context + outcome sentences
ORT-04Uniform rhythmSplit the longest sentence so lengths vary; templated prose marches in step.one 28-word sentence → 18 + 9 words
ORT-08Emphatic punctuationNormalise exclamation marks and all-caps shouting into statements; real acronyms are kept.GREAT SERVICE!!! → Great service.
DIS-01Evaluative, not recountingRecast into the past tense of an event that happened, not a verdict that exists.The service is great → It arrived on time and worked.
DIS-02No specificsInsert explicit gaps for date, object, friction and resolution — never invented values.the team → the engineer on [12 March]
STA-01Frictionless praiseRequire the small imperfection that nearly every genuine account contains.+ The one thing that went wrong was […]
STA-02Rating ≠ sentimentReconcile the star rating with the valence of the words, in whichever direction is wrong.5 stars, complaining text → friction added
AIG-04Machine-writing cuesStrip essay connectives, add contractions, remove the “throughout the process” register.Overall, I would… → I'd…
DUP-01Corpus duplicationFlag near-duplicate phrasing; the transform cannot fix it, because the fix is a different author.max Jaccard 0.59 → reported, not rewritten
05 — Catalogue

Every signal the page computes

IDs are stable across the workspace and match spec/signals.csv. Weights are illustrative reconstructions, not Trustpilot's.

IDSignalMeasurementWeight
ORT-01Token lengthBanded token count; short text is a prior, not a finding.0.09
ORT-04Sentence rhythmCoefficient of variation of sentence lengths. Uniform prose is a template signature.0.06
ORT-05Lexical diversityType–token ratio, suppressed below 20 tokens where it is mechanically high.0.05
ORT-07Trigram repetitionShare of repeated 3-token shingles inside the review.0.05
ORT-08Emphatic punctuationRepeated exclamation marks and all-caps words.0.04
LEX-01Promotional languagePhrase matches from a lexicon of calls to action and sales idiom.0.10
LEX-02SuperlativesDensity of maximal adjectives per 100 tokens.0.08
LEX-04Evaluative vs concreteRatio of judgements to facts — numbers, names, times, domain terms.0.05
DIS-02Event specificityNumbers, currency, clock times, dates, named entities, domain terms; saturates at five.0.07
DIS-01Temporal orientationPast-tense recount markers against present/future evaluation markers.0.05
STA-02Rating ↔ sentimentExpected valence from stars against observed valence from a negation-aware lexicon.0.08
STA-01Frictionless praiseHigh rating, long text, zero negative tokens and no explicit reassurance. Genuine accounts almost always carry one cost.0.05
AIG-04Machine-text cuesContractions, essay connectives, rhythm uniformity, first-person absence. Deliberately weak.0.04
DUP-01DuplicationMax Jaccard over 3-token shingles against a supplied corpus.0.10
06 — Limits

What this cannot tell you

A high score is not an accusation, and a low score is not a clearance. Both readings misunderstand what text can carry.

Text is the weakest layer. Published research on the same n-gram method shows 89.6% accuracy on crowdworker fakes, 67.8% on real filtered reviews, and 52–54% — chance — when a model trained on one is transferred to the other. Adding behavioural features recovers roughly 14 F1 points.

Base rates decide the false-positive count. At a 1-in-1,000 fabrication rate, a detector with 95% sensitivity and 90% specificity is wrong about 99 times for every 100 reviews it flags. No text-only score escapes that arithmetic.

The rewrite is stylistic, not forensic. It removes the markers of manufacture and restores the shape of a recount. It cannot make an invented experience true — that is why every fact it needs is left as an editable gap rather than filled in.

Non-native and terse writers pay for it. Short reviews, translated reviews and reviews from writers with limited English trip length, diversity and specificity signals at higher rates. Any deployment that treats these as fabrication evidence inherits that bias.

89.6%same method, crowdworker fakes
67.8%same method, real filtered reviews
52–54%transferred between the two — chance