Skip to content
EdgeLex

AI & Intelligence · Clause Benchmark

Benchmark every clause against what your firm has actually agreed to.

Included with EdgeLex · Works in EdgeDocs and Word

EdgeDocs — Permit Application Requirements Excerpt · Lex Document Expert · Clauses
FileHomeInsertDrawLayoutReferences✦ Lex

“Project facts of record: the Riverbend Commerce Center site improvements dis…”418 chars · Esc

✦ Ask LexAdd to promptNotePinCompare
Fact noteChronology entryDeposition questionsIssue statementRule checkDatesMatter Intelligence

City Stormwater Permit — Narrative Requirements Excerpt

The application narrative must address: A. a description of the site and proposed improvements, including total disturbed area; B. the measures preventing sediment discharge during construction; C. the permanent stormwater controls proposed; D. the party responsible for maintenance after completion.

Project facts of record: the Riverbend Commerce Center site improvements disturb approximately 18,500 square feet for parking-field regrading and a new bioretention area; silt fencing and inlet protection are specified during construction; Riverbend Commerce Center LLC holds fee title and will maintain permanent controls under its property-management program.

The problem

“Is this market?” usually gets answered from memory. Contract-review AI compares against generic standards or a corpus you can't see — not against what your firm has actually negotiated, accepted, and signed.

How a benchmark earns the right to be cited

Observed. Planned. Snapshotted. Cited. Overseen.

Follow one clause from a verified span to a number Lex is allowed to cite — screens from the real product, shown with sample benchmark data.

01 · Observed

Every observation cites the actual document text.

Clause positions are extracted as source-grounded observations — each one tied to a verified span of the agreement it came from, each one waiting at Approve / Needs review / Reject. Unverifiable extractions are flagged, never trusted; unapproved observations never enter a benchmark.

Clause Benchmark — Master Services Agreement · Harborview / Cobalt IT

Source

Master Services Agreement — Cobalt IT (draft v4)

Version 1fbf86e1-5e92-45b4-9f3b-…

Benchmark result

Limitation of liability — your draft caps at 24 months’ fees. Across 61 comparable agreements (same type, CA law, your side of the paper), 68% cap at 12 months. Your firm standard: 12 months.

Outlier vs. cohort Snapshot q0.87 · Aug 12

Benchmark percentages come only from a saved cohort snapshot — Lex can inspect loaded observations, but no number is cited without one.

02 · Planned

Coverage is managed like a docket.

The planner tells you what's READY for a snapshot, what's gone STALE past the refresh threshold, where review would unlock evidence, and where coverage is too weak to say anything — with a recommended action per clause type.

Benchmark Coverage Planner — what's ready, stale, weak, or blocked
Coverage PlannerMin sample: 5Stale days: 180Run Planner

READY

7

Enough evidence, no current snapshot

STALE

3

Snapshot should be refreshed

NEEDS REVIEW

12

Review can unlock evidence

WEAK COVERAGE

4

Backfill or source work needed

Recommended benchmark actions

Indemnification61 approved observations · no current snapshotCreate snapshot
Limitation of LiabilitySnapshot is 212 days old (stale threshold: 180)Refresh snapshot
Non-solicitation3 observations — below minimum sample of 5Backfill from archive
Governing Law14 observations awaiting review — review unlocks evidenceOpen review queue

03 · Snapshotted

A benchmark is a persisted, scored artifact.

Each snapshot records the question, the clause, the rate, the denominator, and a backend quality score. And when the sample is too small, the snapshot is suppressed — the honest answer is no number, not a misleading one.

Snapshot Library — persisted, scored, and honest about sample size
QuestionClauseRatenQualityStatus
Is a 12-month liability cap market?limitation_of_liability68%610.87completed
Do we take mutual indemnification?indemnification74%480.91completed
Is CA governing law standard for us?governing_law92%1270.94completed
Do our MSAs include non-solicits?non_solicitation3suppressed

Suppressed means the honest answer: three observations is a coincidence, not a benchmark. Below the configured minimum sample, no percentage exists for Lex to cite.

04 · Cited

“Is this market?” — answered from your own deals.

Ask about a draft and Lex answers from the persisted snapshot: the rate, the cohort, your firm standard, and where the draft sits — with the basis attached. What it can't support, it says plainly is suppressed.

Lex — “is this market?” answered from your own deals
Is the 24-month liability cap in the Cobalt IT draft market for us?

It’s an outlier for this firm. Across 61 comparable agreements — same contract type, California law, your side of the paper — 68% cap liability at 12 months’ fees, and 12 months is your approved firm standard. The draft’s 24-month cap sits in the top decile of caps you’ve accepted.

Basis: persisted snapshot limitation_of_liability · n=61 · quality 0.87 · generated Aug 12 — every underlying observation cites a verified span and passed lawyer review.

Not answered: the draft’s non-solicit — that benchmark is suppressed (3 observations, below your minimum sample of 5). No number rather than a misleading one.

05 · Overseen

The program at a glance.

The admin overview keeps the whole benchmark program honest: how many snapshots are persisted, how many are complete enough for Lex to cite, how many are suppressed below sample size, and the firm's position and outcome mix across its negotiation history.

app.edgelex.com — Clause Benchmarks · governed clause intelligenceMC

Persisted snapshots

12

Available for governed benchmark reports

Completed

9

Percentages may be cited by Lex

Suppressed

3

Below configured minimum sample size

Avg quality

0.86

Backend-scored snapshot strength

Benchmark Position Mix

Most recent snapshot distribution

● Preferred 54% · ● Fallback 31% · ● Red-line 15%

71%
Average numerator rate

Generated from persisted benchmark snapshots

Outcome mix: Accepted 44% · Accepted With Modification 37% · Rejected 19%

The difference

Your deals. Not a vendor’s market data.

“Market” from a vendor dataset is somebody else’s negotiations, anonymized past usefulness. EdgeLex benchmarks against what your firm has actually negotiated and signed — filtered by contract type, governing law, and your side of the paper — with lawyer approval standing between machine extraction and firm standard.

The same philosophy, applied to models →

Span-grounded

Every position cites a verified span of the actual agreement — flagged when it can't.

Review-gated

Machine extractions never silently become firm standards. Approval is one audited click.

Suppression-honest

Below minimum sample, the benchmark says nothing. No number beats a misleading one.

Capabilities

What it does

Source-grounded clause observations

Every clause position is extracted as an observation with a verified source span — a citation back to the exact document text. An observation whose span can't be verified is flagged, never trusted.

Benchmarks from your negotiation history

Compare each counterparty position against the firm's own executed agreements. Benchmark cohorts keep the comparison to comparable deals — contract type, governing law, and which side of the paper you were on.

Standards and outliers, instantly

See where the deal in front of you departs from what your firm has historically accepted — one-way versus mutual confidentiality, late-fee positions, liability caps tied to fees paid.

Lawyer-approved, machine-honest

Every observation carries its provenance: Approved (lawyer-reviewed and confirmed), Needs Review, Machine Extracted, or Rejected. Machine extractions never silently become firm standards — approval is one click in the workbench, and it's audited.

Readiness gates that fail closed

A document isn't benchmark-ready until its observations are reviewed and its source spans verify. Pending reviews and unverified spans block benchmarking — the tool tells you exactly what's blocking and why.

Version-pinned to the DMS

Benchmarks attach to an exact document version in your DMS. When the draft moves, the benchmark shows stale and re-runs extraction on demand — you're never reading analysis of a version that no longer exists.

In the flow of drafting

The Clause Benchmark workbench rides alongside EdgeDocs and the Lex chat (/clause) — review observations, approve positions, and redline in EdgeDocs or native Microsoft Word without leaving the document.

Positions, fallbacks, and the clause library

Benchmark results sit next to the firm's pre-vetted clause library, so choosing the fallback for this deal happens with your firm's actual accepted positions — and the deal's context — in front of you.

In your control

The benchmark corpus is your firm's own executed agreements, in your DMS, in your deployment — cloud or self-hosted. Your contracts are never used to train models unless you choose to: with the Lex Training Center, the firm can fine-tune its own model on its own curated work product, in its own infrastructure. Every approval that shapes a firm standard is on the audit trail.

See EdgeLex on your own terms.

We'll walk through self-hosting, model control, and your firm's workflows.