Measured.
Not promised.
Most legal-AI vendors quote one accuracy number averaged across categories. We publish recall per contradiction type against the CLAUSE legal benchmark, including where our pipeline still struggles. Then we show you exactly what a sprint delivers, before you ever book a call.
One accuracy number is a marketing artifact.
Recall varies by type. A pipeline that catches 95% of inconsistencies and 22% of structural defects is not “95% accurate.” It is two different things wearing one number. We publish the breakdown.
- 01Inconsistency0.83STRONGEST
Two sections of the same contract contradict each other on a quantitative or factual point: payment term, notice window, liability cap.
EXAMPLE§4.1“Fees are payable for twelve (12) months.”vs§9.3“On termination, Customer is liable for six (6) months of remaining Fees.” - 02Ambiguity0.78STRONG
A single clause is readable two contradictory ways. Often a numeric reference (“twelve months of Fees”) without a clarified denominator, or a defined term used in an undefined sense.
EXAMPLE§11.1“…limited to twelve (12) months of Fees paid hereunder.”vsn/aReadable as 12× monthly OR last 12 months of annual fees. Material delta on a €200k contract. - 03Omission0.56MODERATE
A required provision, schedule, or annex is referenced but missing from the delivered document. Frequently sign-blocking under DORA Article 30.
EXAMPLE§5.6“Vendor shall maintain the authorized Subprocessors list in Annex B.”vsAnnex BReferenced annex absent from delivered PDF. - 04Terminology drift0.44WEAK
A defined term is renamed or redefined mid-document: “Service Levels” in §2 becomes “SLA” in §9 with subtly different obligations. Easy for human reviewers to miss after 30 pages.
EXAMPLE§2“‘Service Levels’ means the metrics defined in Annex A.”vs§9.4“SLAs are measured monthly; breaches trigger service credits.” - 05Structural0.22 / 0.60WEAKEST
Self-contradictory section hierarchies and broken cross-references: “see §3.4(c)” when §3.4(c) does not exist. The hardest type in CLAUSE; we publish two numbers because pair-mode and single-section analysis perform differently.
EXAMPLE§3.4“See sub-clause 3.4(c) for notice requirements.”vsn/aSub-clause 3.4(c) does not exist. Cross-reference broken.
NOTE · Precision baseline before playbook calibration: 0.21. Calibration to your severity rules and required-clauses list typically improves precision to the 0.55-0.70 range, depending on how strict your rubric is. Numbers above are recall against CLAUSE labels at delivery configuration.
Five elements per finding.
Missing one? It gets dropped.
- CITATIONVerbatim quote, section reference, PDF page. No paraphrasing.
- EXPLANATIONTwo to three sentences, plain language. What contradicts what.
- SEVERITYMapped to your playbook’s rubric, not our opinion.
- ACTIONDrafted recommendation. Your lawyers override; we provide the start.
- OWNERSuggested decision-maker. Editable in the delivered XLSX.
What we're not good at, said out loud.
“If a vendor won't show you this table, ask them why not.”
Read it the way your auditor would.
A person answers
for these numbers.
Every number on this page is mine to defend. I configure the pipeline, run the validation pass, and sign the findings memo on every sprint. There is no junior team behind a curtain.
That's also why the weak numbers are published next to the strong ones. You're going to find out anyway. Better from this page than three weeks into an engagement.
One page. Cleared by your InfoSec before signing.
- RESIDENCY
- Azure EU · Frankfurt region · ephemeral processing tenants
- ENCRYPTION
- TLS 1.3 in transit · AES-256 at rest · per-engagement keys
- ISOLATION
- Multi-tenant isolation · per-engagement vector store · per-org chunk scoping
- RETENTION
- 30-day default · 7-day on request · delete-on-demand · contractual
- AUDIT LOG
- Tenant-scoped · every action recorded · exportable on request
- GDPR
- Pre-signed DPA covers processor obligations · SCCs where applicable
- ROADMAP
- ISO 27001 audit scheduled Q1 2027 · SOC 2 Type II following
- WE DON'T
- Train on your data · share with third parties · retain after deletion
Judge the output,
not the pitch.
Browse real findings from an anonymized sprint: citations, severities, owners, redlines. If it looks like something your team would actually use, book 20 minutes and we'll scope yours.
OpsSolved does not provide legal advice or regulatory representation. We deliver evidence-backed findings and review artifacts; your legal and compliance teams validate findings and make final decisions.