Legal work fails in specific ways: missed exceptions, unsupported recommendations, contradicted clauses tagged as silent, and fluent analysis that restates the rule without applying it. Those classes stay separate. Span overlap is not a legal-correctness verdict. A correct label with the wrong evidence span is a different error than a wrong label with a good span.

Public packs are calibration. LegalBench is a 162-task collaboration with per-task licenses; this report uses published 2023 category results and does not collapse them. LegalBench-RAG is held out until a source-by-source license chain is approved. CUAD and ContractNLI are extractive and NLI diagnostics. A Harvey run would still need sealed matters, a written playbook, and qualified review.

Guha et al. 2023 report balanced accuracy by reasoning category, not a winner score. GPT-4, retrieved May to August 2023, is 82.9% on issue-spotting and 59.2% on rule recall. Rule conclusion is 89.9%; interpretation 75.2%; rhetorical understanding 79.4%. GPT-3.5 and Claude-1 are close to each other on several categories and far from GPT-4 on issue-spotting. That pattern is why a diligence record keeps categories apart.

Rule-application is graded twice: correctness and analysis. GPT-4 is 82.2% correct and 79.7% with analysis. GPT-3.5 is 58.5% and 44.2%. A system that emits the right outcome without the inference is a different product failure than a system that gets the outcome wrong. Hearsay slices in the same paper are small (n 12 to 29) and still useful as a map: GPT-4 lifts non-verbal hearsay from 33% to 75% and in-court statements from 57% to 93%, while both models remain weak when the statement is not offered for its truth (25% vs 45%). Those are model diagnostics from 2023, not a 2026 product ladder and not a Harvey result.

CUAD asks where a clause lives across 41 commercial categories. Hendrycks et al. 2021: DeBERTa-xlarge has 47.8% AUPR and 44.0% precision at 80% recall. RoBERTa-large has a higher AUPR (48.2%) and a worse precision at 80% recall (38.1%). Those two numbers answer different questions. Averaging them would manufacture a ranking that neither metric supports. At 80% recall, even the stronger precision figure still means roughly two irrelevant hits per relevant clause.

ContractNLI adds a three-way label: entailment, contradiction, not mentioned, plus evidence spans. Span NLI BERT with DeBERTa v2 fine-tuned on CUAD reaches 89.2% NLI accuracy, 85.9% entailment F1, and 40.5% contradiction F1. The dangerous class is the weak class. A diligence memo that treats silence and contradiction as the same miss will look better than it is.

The general-model-only thesis is the one Harvey left. In June 2026 Gabe Pereyra said Harvey was building a legal foundation-model series on open weights, inspired by Cursor's Composer, so it could serve frontier intelligence cheaper and let firms own specialized models. Winston Weinberg told Artificial Lawyer the firm was running proofs of concept that encode how a practice group — and a longstanding client — actually works. The trade press named the reversal: post-training open source is coming back after the market had treated it as pointless.

Harvey Tenet, published 20 August 2026, is the first artifact of that turn. Harvey says it is a Kimi K3 base post-trained with Fireworks by asynchronous reinforcement learning on synthetic data, public legal data, and human expert data, and that no customer data was used. The company calls it a research preview. It has not published weights, a model card, or an API. The base model is open-weight. Tenet is Harvey's checkpoint. Harvey II, two days earlier, said that until then Harvey had relied on general-purpose models and that Tenet was the first model post-trained end-to-end for legal reasoning.

The scoreboard is Harvey's Legal Agent Benchmark. Harvey reports that Tenet completes almost twice as many held-out LAB tasks as base Kimi K3 and 20 percent more on LAB: Contracts, lifting all-pass rate by 9 and 2 percentage points, with state-of-the-art on LAB: Contracts and second on LAB. It says those gains transferred, untrained, to Mercor's APEX Agents corporate-law subset and Crosby's Redline Bench. LAB is Harvey's own long-horizon legal-agent bench. Those cells are declared vendor results, often Harvey-run, and they do not stand in for legal correctness on a client matter.

The specialist stack is the rest of the research surface: an RLM diligence harness with Baseten on datarooms Harvey says can reach 80 million tokens; Review Table post-training with Applied Compute; and firm-knowledge memory with Engram on a Qwen3.8-27B. Harvey says the point is twofold — frontier legal intelligence on open weights, and a path for two firms to end up with different models because their work shaped them. The second claim is a switching-cost hypothesis. It is not an observation of exclusivity, export, or a measured quality increment when firm history is added and nothing else changes. A PE buyer still has to ask whether a Chinese open-weight base is an acceptable supply and trust fact for the firm's clients.

The failure mode that matters for Harvey is publishing a LegalBench category score, or a Harvey LAB cell, as if it were legal correctness on a client matter. Playbook, retrieval, routed frontier API, and Tenet have to be changed one at a time or left unproven.

If only the product is observable, publish outcome and price. Leave model-level, post-training, and orchestration-level stories as unproven. Native PDF or long-context dumping is a different treatment from retrieval with evidence spans.

Current list prices bound the model component only. An economic-tier model such as GPT-5.6 Luna or Claude Haiku 4.5 can be the right cell if it clears the same quality and critical-miss gates as a frontier cell. That is an efficiency finding, and it requires the frozen ladder, not a preference for the newest name.

Full cost per accepted memo still needs failed attempts, retrieval, partner or associate correction, and any embedded legal-engineering time, with a cap treated as censored. Harvey describes Legal Engineers; intensity, margin, accepted-without-edit rate, and escalation burden are missing. A low token bill with high implementation or review load is not cheap automation.

Tenet's cost story is vendor-declared. Harvey says open-weight token prices are lower and that reward shaping cut tokens used for the same quality. Review Table research claims better citation quality at roughly one-tenth the cost per cell versus the strongest baselines they ran. Those cells are Harvey's. They are not full cost per accepted memo on a client matter.

Matter documents are client- or firm-owned. Harvey says customer data and content are not used to train shared models or improve shared products. Without the target's executed agreement, retained-feature settings, and customer-specific model terms, lawful reuse remains unknown and no cross-client matter-data flywheel may be claimed.

Public CUAD-style extraction and ContractNLI already show that clause location and contradiction are model tasks. Direct substitution is a generic legal stack inside the firm's DMS, not a new corpus. Microsoft, DMS vendors, and research terminals can attach generation to seats the firm already pays for. Workflow embed inside firm process can raise switching costs. That remains a structural hypothesis without retention, export friction, and services-dependence observations.

Harvey still describes a multi-model platform that routes across Anthropic, Google DeepMind, and OpenAI. Tenet sits beside that architecture as a declared post-trained checkpoint, not as a replacement that has been isolated. Routing, retrieval, playbook, and Tenet have to be changed one at a time or left unproven.

Public materials also describe ownership, no-training of customer data, provider-retention, provider-control, and embedded legal-engineering practices. Harvey's Tenet post says the same no-customer-data rule for the research run. Those are useful leads, not approved trust or service claims. Executed contracts, legal correctness, privilege handling, service intensity, and switching still need separate evidence.