Sign in
Enter your email and we'll send you a magic link — no password.
You are signed in, but this account is not an editor — there is nothing here for it. Your key and history are on the dashboard.
Rarest clauses
How unlike the rest of the Baseline each clause is, as a percentile within its document type. A shortlist and never a verdict: it nominates clauses for judgment, it does not pass one. Rare is not the same as harmful — that is still yours to decide, and the neighbours below each score are how you check it.
Scoring is per document type, because a percentile is a position within one: pick a type to score it. The pass fetches nothing and spends no tokens — it reads the clauses and embeddings the crawl already stored — and a new run sits beside the last rather than replacing it.
How much Baseline is enough to score
A score is the mean distance to the 15 nearest clauses in other companies’ documents of the same type, so the number that decides whether a run means anything is companies crawled for that type — not companies admitted, and not clauses in total.
- 0–1 companies of the type: nothing to compare against — Fewer than two companies have a document of this type crawled, so no clause has an eligible neighbour and no run is written at all. This is not a thin ranking — it is not a ranking.
- 2–19 companies of the type: a smoke test — It will score, but rare means unlike those few. Read it as proof the pipeline works end to end, not as a shortlist worth an afternoon.
- 20–99 companies of the type: usable — The scale the trial run was measured at — 20 documents and about 1,700 passages — and the first point at which the top of the list was worth reading.
- 100 or more companies of the type: what this was designed for — A hundred or more companies of the type. Distances stop moving much as the pile grows past here.
Two things to read on the run itself. Clauses reported as not embedded under this model were ranked against nothing and did not count — the embeddings pass has versions left to finish, and they are listed under this panel with the button that finishes them. And the Baseline is counted three ways on every run, so a ranking that looks thin can be checked against the pile it came from rather than guessed at.
Documents with no document type
Rarity is a position within a document type, so a document carrying none is scored by nothing and never reaches the ranking below. These are documents admitted before the classifier learned to read a marker after a dash or an underscore — nothing re-derived the column when it did.
This rewrites one column from an address already on file. It fetches nothing, re-cuts nothing and scores nothing: the clauses do not change, they become visible to a ranking that could not see them. Addresses carrying no marker at all are listed and left alone — typing those is a judgement over the documents themselves, not something to guess from a URL.
Versions cut under older rules
Merging a rule into the cutter filters every future crawl and leaves every clause already stored exactly as it was — and the top of the ranking below is made of those. This is the walk that re-cuts them from the canonical copy we already keep: it fetches nobody’s website and spends no tokens, and it re-embeds what it produces as it goes.
One click queues the whole pile and the work continues in the background, so this page can be closed. It is safe to press again — a version already cut under the current rules is skipped. Only versions a run would actually read are walked: current versions of admitted companies’ documents that have a document type.
Versions with clauses left to embed
Embedding is a pass over stored clauses: it fetches nobody’s website and spends no tokens. A crawl that ran out of its embedding budget, or ran while the embedding service was down, leaves clauses with no vectors — and a clause with no vector is ranked against nothing, so finishing these is what makes the run below cover the whole pile. Only versions a run would actually read are listed: current versions of admitted companies’ documents.
Documents waiting on a harm score
Rarity nominates and a model judges. The ten rarest clauses of each document go to a model together with the whole agreement they came from, because harm is contextual: “we may terminate at any time for any reason” costs a reader nothing in a free service and a year of prepaid fees in an annual contract. What comes back is a signed score and one line saying why.
Signed is the point. Negative is a clause that is good for the reader, zero is one that does nothing to them either way, and positive is a trap — so a trademark notice ranks first for rarity, scores zero, and never reaches a Hall. Empty is not zero. A clause with no score is one nobody has judged, and every clause scored before this pass existed is in that state.
Ten clauses per document, and per document rather than per type: a percentile is a position within a type, so a type-wide cut is monopolised by whichever documents are unusual overall and a company whose single worst clause sits at the 80th percentile is never looked at. This is the one pass here that spends tokens — one model call per document — and it fetches nobody’s website: the agreement it reads is the copy we already keep.