Methodology & Sources
Where the numbers come from.
Pleadly publishes two kinds of claims on its marketing pages: properties of the system (verifiable from the codebase) and industry context (verifiable from third parties). This page lists each claim with its source. Aggregate customer-outcome claims are explicitly out of scope until pilot volume is large enough to report honestly.
Industry Context
Third-party citations
Firm utilization rate
38%
Clio 2026 Legal Trends Report (firm data analyzed: 2025) — average lawyers capture roughly 3 of 8 daily hours as billable, with the remainder lost to admin and non-billable work.
Regulatory Environment
Why these properties exist
Pleadly's architectural choices map to specific regulatory developments. Each item below is cited so readers can verify the source independently.
California SB 574
Senate Jan 2026
Passed the California State Senate January 2026; requires lawyers to verify AI-generated materials and correct false or hallucinated outputs, and restricts confidential information in public AI tools.
COPRAC proposed Rule 1.1
Mar 2026
California State Bar's Committee on Professional Responsibility and Conduct approved proposed amendments March 13, 2026 requiring lawyers to independently review, verify, and exercise professional judgment regarding generative AI outputs. Comment period closed May 4, 2026; adoption status tracked separately.
Heppner (S.D.N.Y.)
Feb 2026
United States v. Heppner — Judge Jed Rakoff ruled that communications with publicly available AI services may lack reasonable expectation of confidentiality and lose attorney-client privilege / work-product protection. See the full analysis on the Security page.
NCSC AI guidance
2026
National Center for State Courts published practitioner guidance: 'never trust, always verify' framing for legal AI use, with explicit warning against submitting AI-generated content to courts without independent review.
ABA Formal Opinion 512
Jul 2024
ABA Standing Committee on Ethics and Professional Responsibility issued the first formal ethics opinion on generative AI in law, addressing competence, confidentiality, communication, fees, and supervisory duties. Pleadly's alignment is documented on the /legal/aba-512-alignment page.
Sources: Reuters — SB 574 · California State Bar — COPRAC · Reuters — Heppner · NCSC practitioner guide · ABA Opinion 512
System Guarantees
Architectural claims
These properties are enforced in code, not aspirational marketing commitments. Each is independently auditable in the codebase.
Private infrastructure
100%
All LLM inference runs on local hardware (Ollama / on-premise GPU). Case content is never transmitted to a third-party AI provider.
Cloud AI calls per case
0
The Control Plane does not install Anthropic, OpenAI, or Google AI SDKs. Document content is processed only by the local Intelligence Plane.
HMAC-signed transport
SHA-256
All Control ↔ Intelligence Plane requests are HMAC-signed with a 5-minute timestamp window to prevent replay attacks.
For the full architecture, see the Security & Compliance page.
Quality Methodology
How a demand earns delivery
Every generated demand is scored by a 5-dimension quality grader before it can reach an attorney's inbox. The scoring rubric and thresholds are fixed in code; firms cannot dial them down.
Citation coverage floor
≥ 80%
Minimum share of factual assertions in a demand that must link to a source citation. Below this floor, the demand cannot auto-deliver.
Overall quality floor
≥ 0.85
Composite score across citation coverage (35%), numeric/date integrity (25%), liability coherence (20%), tone (10%), and weakness coverage (10%) required for auto-delivery.
Grounding threshold
≥ 0.7
Extracted facts scoring below 0.7 grounding (fuzzy match to source text) are auto-capped at 0.5 confidence and flagged for attorney review.
OCR confidence tiers
4 tiers
Excellent (≥90%) safe for assertions; Good (75–89%) spot-check required; Fair (60–74%) context only; Poor (<60%) excluded from demand text.
Attorney review enforcement
Hard gate
No source = no export. Demands with unsupported claims cannot leave the system (HTTP 400); demands that have not been attorney-reviewed cannot leave the system (HTTP 403). Enforcement occurs at the database state machine, not the UI layer, with ABA Model Rule 5.3 alignment documented separately.
Reliability Methodology
How tenant safety is enforced
Tenant-isolation tests
57
SQL-based RLS policy tests covering cross-org isolation, case-level access, paralegal scoping, and shared-state leakage. Run in CI on every change.
Intelligence circuit breaker
5 / 60s
After 5 consecutive failures the EVO-X2 endpoint is marked OPEN; a probe request is allowed every 60s to test recovery.
Document deduplication
SHA-256
Per-case SHA-256 fingerprinting rejects duplicate uploads with HTTP 409 before extraction starts.
Job idempotency
Per-key
Pipeline jobs carry idempotency keys so retries cannot create duplicate demand drafts.
Customer Outcomes
What we are not claiming yet.
Aggregate customer claims — average hours saved per demand, settlement-value lift, paralegal cost displaced — require a pilot population large enough to be statistically honest. Pleadly will publish these figures here, with methodology and sample size, once that threshold is met. Until then, marketing pages reference only system properties (above) and cited industry context.
Tracked metrics in progress: median attorney revision count per demand, time from intake to first reviewable draft, citation-coverage distribution on shipped demands.