Method
This page describes exactly what the tool does, so you can judge whether its numbers mean anything for your case.
The question being answered
“Is this original?” is not answerable. Nobody can search all human knowledge, and no absence of evidence establishes that a thought is unprecedented. So the tool answers a narrower question that is tractable: how far does this idea sit beyond what someone could reach cheaply?
“Cheaply” is operationalised as two baselines — the accessible idea frontier. Everything the tool reports is a statement about position relative to those two baselines, and nothing more.
The pipeline
Fourteen durable steps. Each commits its results before the next begins, so a crash, timeout or redeploy resumes rather than restarts.
- 1Freeze reviewed fingerprint
- 2Generate search queries
- 3Run public web searches
- 4Normalise and deduplicate results
- 5Compare sources with the fingerprint
- 6Generate model-baseline samples
- 7Extract atomic claims
- 8Generate embeddings
- 9Cluster related claims
- 10Compare idea with baseline clusters
- 11Evaluate coherence, specificity, testability, grounding
- 12Calculate deterministic scores
- 13Generate the narrative report
- 14Mark the analysis complete
Source relations
Each retrieved source is labelled by its relation to your idea, not by keyword or embedding similarity. Similarity shortlists candidates; it never decides equivalence. The number beside each relation is how strongly a source of that kind suppresses an originality claim for the facets it covers.
| Relation | Suppression weight |
|---|---|
| Essentially identical | 1.00 |
| Same position, different wording | 0.90 |
| Same conclusion, different mechanism | 0.55 |
| Same mechanism, different conclusion | 0.55 |
| Known components combined differently | 0.35 |
| Adjacent idea | 0.15 |
| Contradictory idea | 0.05 |
| No meaningful match | 0.00 |
Weights are combined with a rank-decayed noisy-OR, so one decisive match saturates the facet while a pile of weak matches accumulates slowly and never reaches certainty.
The model-accessibility ladder
Level 0 — Produced directly by a basic prompt
rarity 5
A first, ordinary prompt returned this.
Level 1 — Produced through repeated basic sampling
rarity 20
Repeated sampling at the same basic prompt surfaced it.
Level 2 — Produced only after requesting unconventional alternatives
rarity 40
It appeared only once the model was asked for unconventional options.
Level 3 — Produced only through multi-step reasoning
rarity 60
It appeared only when the model was asked to reason in steps.
Level 4 — Reconstructable only after supplying key assumptions
rarity 80
It appeared only after the model was handed your key assumptions.
Level 5 — Not produced within the tested baseline
rarity 95
No sample in this baseline produced it. That is a statement about these samples, not about the model's knowledge.
Level 5 caps at rarity 95, not 100
Why the numbers are deterministic
Language models are used for judgement calls they are good at: labelling how a source relates to a claim, splitting prose into atomic claims, rating an argument against a described rubric. They are never asked to produce a score.
Instead, the evaluator returns observations — ordinal ratings, counts, relation labels — and pure TypeScript functions convert those into 0–100. The same evidence always yields the same numbers, and every report can therefore show its full arithmetic.
Originality Estimate =
0.30 × public-search deviation
+ 0.25 × model-baseline rarity
+ 0.20 × mechanism & structural novelty
+ 0.15 × consequential difference
+ 0.10 × specificity
Qualified Originality =
Estimate × coherence factor × testability factor
where factor(x) = 0.70 + 0.30 × (x / 100)
Handling untrusted content
Retrieved pages are written by third parties and are treated as data, never as instructions. Excerpts are stripped of markup, control characters and zero-width smuggling before storage, then fenced inside labelled blocks with a random delimiter when passed to a model.
A source caught issuing instructions to the analysis system is classified as no-match, excluded from facet matching, and flagged in the evidence explorer — visible, but unable to influence a score.
What would change the answer
- A different search provider or index. Each covers a different slice of the web.
- More baseline samples. The frontier is stochastic; more sampling finds more of it.
- A deliberate prompter. Our prompts are ordinary on purpose. Someone working toward your idea would reach it far sooner.
- Sources outside the open web. Books, paywalled journals, theses, patents and non-English work are never searched.