Document Review & Production

We run review in four explicit passes, each with its own prompt, model, gate, and reviewer cohort. Privilege is never coupled to relevance — that decoupling alone eliminates the anchoring bias that quietly drives errors in legacy review.

• Pass 1 — Responsiveness Triage. Fast classifier (Claude Haiku) sweeps the full corpus, scoring against your responsiveness criteria with reasoning and cited passages. Confident-negative documents are statistically retired; everything else moves forward.
• Pass 2 — Issue Coding. Mid-tier model (Claude Sonnet) codes against an attorney-approved issue taxonomy auto-derived from the opposing party's RFPs and the complaint. Every issue tag traces back to a specific RFP number.
• Pass 3 — Privilege Review. Dedicated privilege pass with strict prompt and zero-tolerance elusion. Outputs cover attorney-client, work product (fact / opinion), joint defense, common interest, deliberative — plus privilege holder, waiver risk, and redaction scope.
• Pass 4 — Hot / Key Documents. Identifies the small percentage of documents that move the case, with narrative role and strategic-value scores attached.

Progressive filtering also makes the pipeline efficient: the expensive privilege pass only runs against the ~25–30% of documents that survived responsiveness triage, not the full corpus.

Every matter is calibrated on its own data. We draw a stratified, dual-blind-coded control set of 600+ documents (1,000 for low-prevalence matters), require Cohen's kappa ≥ 0.80 across reviewers, and qualify every reviewer with a 30-doc test at ≥ 90% before they touch a coded document.

• Banded confidence. Isotonic regression maps raw model confidence to calibrated confidence. Confident-Negative band must show measured elusion ≤ 2%; Confident-Positive ≥ 95% precision. The grey zone (typically ≤ 20% of corpus) is where qualified humans review.
• Six enforced gates (G0–G6). Each pass blocks downstream work on failure. Statistical sampling, reviewer-kappa monitoring, and recall estimation (Cormack-Grossman Total Recall, ≥ 0.90 with 95% CI lower bound ≥ 0.85) are gates, not ceremony.
• Pass 3 zero-tolerance. 5% of the cleared pool (min 300 documents) is re-reviewed blind. One privileged miss triggers a full Pass 3 re-run and a dual-signed override event.
• Pass 4 production precision. 400-document random sample, Clopper-Pearson 95% CI, target precision ≥ 0.97 with lower bound ≥ 0.93 before anything ships.

Every assertion the review makes is supported by a paper trail engineered to survive a meet-and-confer, a special master, or a Daubert challenge.

• Reproducibility Ledger. A machine-readable JSON record of prompt hashes, model versions, random seeds, calibration model SHAs, thresholds, gate outcomes, and reviewer kappas — retained 7 years.
• Validation Report. An interactive HTML deliverable: reliability diagrams, confusion matrices, per-reviewer kappa, recall / precision / elusion with 95% Wilson confidence intervals, ledger appendix, and case-law citations. Designed to hand directly to opposing counsel.
• Case-law anchored. Built to support sworn declarations citing Da Silva Moore (2012), Rio Tinto v. Vale (2015), Hyles v. NYC (2016), and FRCP 26(g), 26(b)(5), FRE 502(d).
• Reasoned, cited classifications. TAR 1.0/2.0 give you a rank score; we give a reasoned decision with the document spans the model relied on, traceable per document.

Every document in production was reviewed either by licensed counsel or by an AI model whose predictions were statistically validated against a blind, dual-human-coded control set drawn from this matter's own corpus — with recall, precision, and elusion reported as the specific sentences we stand ready to swear to under oath.

Privilege screening runs with explicit rules and learned patterns — counsel-of-record relationships, document types, content signals, communication graph context — and feeds the privilege log automatically. Responsiveness criteria are jointly defined with case strategy at intake, then refined as review reveals what actually matters.

Issue tags are not a generic taxonomy — they are derived from this matter's RFPs, complaint, and case theory by an agent, attorney-approved before any document is coded, and traced back to specific RFP numbers so when opposing counsel asks "did you code against RFP 17?" the answer is a tag, a count, and a sample.

Productions land in the format opposing counsel or the regulator expects: Concordance / Relativity / Eclipse load files with .DAT and .OPT, image stores, native files, and burned-in redactions. Each production package includes:

• Auto-drafted FRCP 26(b)(5)-compliant privilege log (human-finalized).
• Bates numbering with contiguity check (Gate G6) and per-document chain-of-custody export.
• Redaction logs, hashes, and a defensibility memo describing the review protocol.
• The Validation Report and Reproducibility Ledger covering the exact data shipped.

Every metered event in the LexGo AI platform is attributed back to the matter, so review can be invoiced any of three ways — and you can change the meter mid-engagement without rebuilding the audit trail.

• Per document. Predictable unit pricing. Best when the corpus is well-bounded and the pages-per-doc distribution is roughly known.
• Per gigabyte. Volume-based. Best for ingestion-heavy matters with mixed file types, native productions, and large attachments.
• Per token. Pure pass-through of model usage at a published markup, with three model tiers (Fast, Balanced, Deep Analysis). Best for portfolios where matters vary widely or where in-house teams want full transparency into AI consumption.

Volume discounts apply at scale — 5% at 5–9 seats up to 25% at 100+ seats, plus 10% on annual subscriptions and 15% on annual platform fees. A representative 240,000-document review typically completes in 14–24 days versus 48–60 days for traditional linear review at 50 docs/hr/reviewer — roughly a 70% review-time reduction.

Document review prompts are mostly static across a matter — the review protocol, the issue taxonomy, the exemplar set, the calibration thresholds. We engineer review jobs to take advantage of two model-provider optimizations and pass the savings directly to clients.

• Prompt caching. The static blocks of every review prompt — protocol, taxonomy, exemplars, ontology context — are written into the model provider's prompt cache once per matter, then re-used across every document. Each subsequent document only pays for the dynamic part of the prompt (the document text and its metadata), reducing per-document inference cost dramatically while keeping the prompt itself byte-stable for defensibility.
• API batching. Document review is a textbook batch workload — large volumes, no real-time UX dependency, parallel-safe per-document calls. We submit review passes through Anthropic's Message Batches API at the published 50% discount on input and output tokens, with overnight or matter-window turnaround. Urgent matters fall back to interactive endpoints; everything else goes through the batch lane.
• Tiered model routing. Pass 1 (responsiveness triage) runs on Haiku-class models; Pass 2 (issue coding) and Pass 3 (privilege) on Sonnet-class; Pass 4 (hot docs) and any narrative work on Opus-class. Each document only pays for the tier it actually needs.
• Pass-through pricing. When provider rates drop, batch discounts grow, or cache hit-rates improve, those savings show up directly in the next invoice — not retained as margin. Per-case cost reports break down cached vs. uncached tokens, batched vs. interactive calls, and per-pass model spend so the savings are auditable.

The combined effect: a defensible four-pass review at a unit cost meaningfully below what legacy TAR-based review services charge — without giving up the validation rigor that makes the output usable in court.

Pricing starts at $0.05 per document or $6 per 1M input tokens for AI-accelerated review. Per-GB and matter-flat envelopes are also available. Inquire about our additional pricing optimizations — implemented prompt caching and Anthropic Message Batches API integration drive incremental savings that flow straight through to your invoice.

Our pricing re-inverts case economics: lawyering and strategy become the most significant cost of litigation again — not document review.

For decades, document review has dominated case budgets, often eclipsing the cost of the actual legal work the review was supposed to support. By driving review unit cost down by an order of magnitude — without sacrificing defensibility — we put the dollar back where it belongs: with the attorneys building the case.

Every engagement starts with a complimentary scoping call. You receive a written estimate with the pricing model, expected volumes, and a not-to-exceed cap before any work begins — full transparency and predictable pricing before you commit.