Blind spots
What this page is — the capability that compares the draft in front of you against the documents your own organisation has already finalised, and tells you what yours is missing that they all have.
What it is for — so "does this look like the agreements we normally sign?" stops being a question only your most experienced lawyer can answer.
The problem it solves — a draft missing a clause every comparable agreement carries is only noticed by the one person who has read them all.
Route: /org/papers/documents/<document>/ai → Understand → Blind spots ·
Permission: Use AI drafting/extraction/Q&A for documents. ·
Utility: papers_blindspot · Cost: 450 credits estimated, 1,800 maximum.
1. What it is
Blind spots benchmarks a draft against your gold-standard corpus — the documents of the same family that your organisation has already taken all the way to finalised. It samples passages from that corpus and asks one question: what do our finished agreements consistently contain that this draft does not?
Every gap it reports is tagged with where the expectation came from, and that tag is the whole point of the capability:
corpus_support | Means | How much weight to give it |
|---|---|---|
corpus | Your own finalised documents of this family reliably carry this | High. This is your house style, evidenced |
benchmark | Any market-standard document of this kind would carry it | Moderate. Generic, but rarely wrong |
both | Your corpus and market practice agree | Highest. Hard to argue this one away |
That distinction is what separates it from its nearest sibling:
| Checks against | Answers | |
|---|---|---|
| Blind spots | Your own finalised corpus, sampled live | "This is not how we write these" |
| Missing clauses | A generic clause checklist | "This is not how anyone writes these" |
| Cross-document consistency | The other documents in this contract family | "This contradicts the MSA it sits under" |
Running missing clauses and blind spots on the same draft is not redundant work. The first catches the omission that is obvious to the market; the second catches the omission that is only obvious to someone who has read your last forty NDAs.
2. Why you would use it
The expensive gaps are never the ones a checklist finds. A missing governing-law clause gets caught by anyone. What does not get caught is the indemnity carve-out your team has quietly added to every services agreement for two years because of one bad experience in 2024 — and which is absent from the draft a new joiner produced last week, because nobody wrote the reason down.
That knowledge lives in your finished documents and nowhere else. Blind spots reads it back out.
Concretely, it earns its keep three ways:
- It transfers institutional knowledge without a person. A reviewer who has never seen your back catalogue gets the benefit of it on their first draft.
- It gives you an argument, not just a flag.
corpus_support: corpusmeans you can tell the counterparty "this is our standard position", and be telling the truth. - It gets better as you work. Every document you finalise enriches the corpus it benchmarks against. A capability that improves because you used the product is rare; this is one.
Early in a project's life there is little finalised corpus, so most findings come back
benchmark — market practice only. That is the capability being honest rather than the capability
failing. The corpus proportion rises as your finalised set grows.
3. What you provide
Nothing. There are no inputs to fill in and no documents to select — this is a one-press capability.
| Input | Required | Notes |
|---|---|---|
| — | — | No parameters. The comparison corpus is chosen for you |
The corpus is not selectable, and that is deliberate: letting a reviewer pick which finalised documents to benchmark against would let them pick the ones that make the draft look finished.
4. What it reads automatically
| Read from the document | Used for |
|---|---|
| Draft text — the full body under review | The thing being benchmarked |
| Corpus excerpts — passages sampled from your finalised same-family documents | The standard it is benchmarked against |
| Title & family | Selects which corpus to sample. A "Mutual NDA" is compared to your NDAs, not your MSAs |
| Parties & key field values | The filled schema values — amounts, dates, named parties |
The corpus is drawn from your organisation's indexed documents, which means a document only joins the gold standard once it has been finalised and indexed. A document finalised five minutes ago may not be in the sample yet.
5. What you get back
A single array, blindspots[], ordered highest severity first. An empty array is a real result:
it means the draft is genuinely complete against both your corpus and market practice.
| Field | Type | Values | Means |
|---|---|---|---|
item | string | — | Short name of the missing clause, definition or schedule |
kind | enum | clause, definition, schedule, annexure, other | What sort of thing is missing |
severity | enum | high, medium, low | high is reserved for gaps whose absence shifts money or liability |
why | string | — | One sentence on why the gap matters for this document |
corpus_support | enum | corpus, benchmark, both | Where the expectation came from — see §1 |
Read severity and corpus_support together. A high / both finding is the one to fix before
the draft leaves the building. A low / benchmark finding is often a deliberate choice your
organisation has already made.
6. Worked example
A contracts manager runs Blind spots on a services agreement drafted from a new supplier's paper.
Input: none — the document is open, and Blind spots is pressed under Understand.
Result — one of four findings returned:
Missing definitions section
kind: definition·severity: mediumwhy: Terms like "Confidential Information" and "Services" are used throughout but never defined, leaving their scope arguable.corpus_support: both
What the manager does with it. both means this is not a matter of taste — their own finalised
agreements define these terms and any market-standard agreement would. The scope of
"Confidential Information" is exactly what a dispute would turn on, so the finding goes into the
first round of comments rather than being noted for later.
The three other findings in the same run were graded low / benchmark, and were left alone: the
draft omits a force-majeure clause that this organisation deliberately drops from short engagements.
The capability flagged it; the human overrode it. That is the intended division of labour.
7. Running it
- Open the draft in Orbit Papers.
- Open the AI Assistant drawer, or go to the document's AI workspace.
- Choose Blind spots under Understand.
- The result appears in its own tab and is saved to the document.
Because results belong to the document rather than to your session, a colleague opening it tomorrow sees this run rather than paying 450 credits to ask the same question. Use Re-run when the draft has changed materially — a re-run is a fresh charge.
Blind spots can also be set to run automatically — on submit, on each counterparty round, or at finalize — from the document type's AI Config tab. See AI configuration.
8. The admin contract
| Must be true | Where | What happens if it is not |
|---|---|---|
| Your role holds Use AI drafting/extraction/Q&A for documents. | Role editor | The AI Assistant button does not appear at all |
The papers_blindspot utility is active | Orbit AI Flow → utilities | "This utility is currently disabled" |
| Your organisation's AI master toggle is on, and this utility is enabled for it | /org/ai-utilities | The capability is absent from the Understand group with no error |
papers.ai_monthly_credit_cap is not yet reached | System Config | "monthly AI credit cap reached: n of n credits used this month" |
| Finalised documents of the same family exist | Your own repository | The run still succeeds, but every finding comes back benchmark — there is no corpus to compare against |
For a document whose confidentiality is confidential:
- Automatic runs are blocked entirely — default deny. A platform administrator must set the
System Config key
papers.ai_confidential_cloudto the literal valueallow. The block is silent to the user and appears only in the server log. - Manual runs are not blocked. If a local model key is configured, the run is forced onto the
on-prem model and audited on the document timeline as
ai_local_routed; content never reaches a cloud API.
So a confidential document can return a Blind spots result you pressed for, while its type's auto-run policy quietly does nothing.
9. Don't confuse this with…
| Missing clauses | Generic checklist. Blind spots benchmarks your own corpus |
| Compliance & policy check | Checks against rules you configured. Blind spots infers the standard from documents |
| Fairness / balance | Judges the terms that are present. Blind spots is about what is absent |
| Ask the repository | Answers a question across the corpus. Blind spots benchmarks one draft against it |
10. Troubleshooting
| Symptom | Cause |
|---|---|
Every finding says corpus_support: benchmark | You have no finalised documents of this family yet, or they are not indexed. The result is market practice only |
blindspots[] came back empty | A real answer — nothing is missing against corpus or market. Check the draft is not near-empty first |
| The capability is missing from Understand | Your role lacks Use AI drafting/extraction/Q&A for documents., or the utility is off for your organisation (§8). The document type does not gate it — the type's AI Config controls auto-run only |
| It found nothing on a long agreement | The document may be a scanned PDF with no extracted text. Run OCR first |
| Nothing happened on submit, though auto-run is configured | The document is confidential and cloud AI is default-denied (§8) |
| "monthly AI credit cap reached" | The organisation's monthly Papers AI spend cap is exhausted. It resets next month |