Inside the Originality Gate: Inspired By, Not Copied

Every piece of content that reaches a learner on this platform passes a publication gate that scores it against its own cited references. Here is how the gate works — shingles, similarity, verbatim runs — and why its verdicts are sealed into a hash-chained provenance trail.

The rule

Content published on Go2LearnAI must be inspired by its references, not copied from them. That sentence is easy to put in a policy document. Making it executable is the interesting part.

How the check works

When an author submits a package for publication, the originality engine compares the candidate text against every reference the author cited:

1. k-shingling. Both texts are broken into overlapping token windows (shingles). Two texts that share ideas but not phrasing share few shingles; near-copies share many.
2. Jaccard and containment similarity. Jaccard measures overall overlap; containment catches the case where a short candidate is lifted wholesale from inside a long reference — a case plain Jaccard under-reports.
3. Longest verbatim run. Independent of set similarity, the engine finds the longest uninterrupted token run shared with any reference. A paragraph-length verbatim run is a copy even if the rest of the document is original.

The result is a verdict with three levels: original (publish), derivative (publish with a visible citation), plagiarised (blocked, with the offending spans highlighted so the author can rework them).

Why the verdict is sealed

A gate you cannot audit is a gate you have to take on faith. So every verdict is sealed into an append-only provenance chain: each record carries a SHA-256 digest over the previous record's digest plus its own canonical content. Editing, deleting, or reordering any historical verdict breaks the chain at that link, and a verification pass says exactly where.

Two details matter for privacy: the chain stores a hash of each reference's text, never the text itself — provenance proves what a piece was checked against without republishing the source material. And the chain has two real backends (in-memory and Postgres) that seal byte-identical digests, so a local development chain and a production chain are verifiable by the same math.

What this buys learners

A child reading a lesson here is reading something a machine checked for honesty and a human approved for publication — and both facts are provable after the fact. That is what we think "trust" should mean in an education product: not a brand promise, a verifiable property.