Your own private AI,
without the trade-off.
Your firm gets its own copy of the model, on a machine only your firm uses. Not slower. Costs less to run. Zero setup.
What goes wrong with legal AI.
58 to 88%
of legal answers were wrong when general AI was used raw.
- Cites cases that do not exist
- Answers out of the wrong matter
- Never says "I don't know"
- Priced per seat, at enterprise rates
- Your files sit on someone else's servers
Private does not mean second best.
Your model, with and without Sidian Chambers, the AI control layer on top of it, against each lab's best on 1 September 2026.
GPQA Diamond, graduate-level science questions. Higher is better
- The model on its own
- With Sidian Chambers
Legal accuracy scores are coming soon.
A model small enough to run on one machine, within two points of Google's and OpenAI's best, and above Anthropic's best.
One fee for the whole firm.
Set by headcount, never per seat or per use. A small model keeps it low.
Yours
- The environment
- The model, your firm's own copy
- Full control
You see everything that runs on it.
Sidian
- Setup, zero on your side
- Securing and maintaining the machine
- Sidian Chambers, on top
No AI engineers to hire.
One machine, rented through Sidian, used by your firm alone.
The part that stops a wrong answer.
Three problems, three controls. Sidian Chambers is all three.
- Every answer cites a document you can openGrounding
- Only that matter's files are openedRouter
- "I don't know" instead of a guessGovernance, at a threshold you set
All three run on every request, over the files you connect, read-only. Governance is built to stop on an uncertain answer rather than write it.
Answers you can open.
The same request, put to a generic AI tool and to your firm's AI.
Same request to both"What documents were produced in the data breach incident?"
A generic AI tool, with no access to your files
These exhibit references cannot be confirmed
It never saw the case file, so these exhibit numbers could be right or invented.
Someone would have to check every exhibit by hand before filing.
Your firm's AI, connected to your real documents
Answers from the connected case document set
Exhibit 14, page 3; Exhibit 7, page 1; Exhibit 22, page 5Approved source · Doe v. TechCorp, case no. 3:22-cv-0451Each exhibit it cites is a document in the case you can open.
Matters and citations are invented for illustration; Mata v. Avianca is the only real case.
Show exactly what left.
Most of the work never leaves the firm. When something does, you can still answer for it a year later.
- You approve the sendNothing leaves on the model's say-so.
- DataGuard removes what your rules say should not leaveNaming the rule that fired, and the passage.
- The record is tamper-evidentAny later edit shows. A record, not a recollection.
Logged for anything sent through a connector you approved: Box, Egnyte, SharePoint, email. Sidian runs it in its private cloud, on SOC 2 audited infrastructure.
Your model. Your rules.
Two ways to get it. Yours either way, with Sidian Chambers on top.
- We set one up for youPick an open-weights model. Sidian installs your firm's copy on a machine rented for you, and runs Chambers on top.
- Bring your own modelAlready running an open-weights model, one your firm can download and keep? Sidian puts Chambers on top.
The machine is only your firm's and closed to the internet, so your data and your model stay sovereign. Nothing dials out except connectors you approve.
One difference that matters.
Both are good tools. What differs is where your data goes.
| Harvey | Spellbook | Your private AIDelivered by Sidian | |
|---|---|---|---|
| Where your data goes | Processed on Harvey's cloud infrastructure | Sent through GPT‑5, Claude and other outside models | Stays on one machine dedicated to your firm, rented through Sidian or in your own cloud or building. Closed to the internet except the connectors you approve. Your firm keeps data sovereignty |
| When the AI is not sure | Not documented as a separate, visible control | Not documented as a separate, visible control | It says so and holds for your review rather than committing, at a strictness threshold your firm sets |
| What you can show about one privileged file | Not documented publicly at the file level. Harvey states it does not train on customer data and requires zero data retention from its model providers | Not documented publicly at the file level. Spellbook states it holds zero-data-retention agreements with OpenAI and Anthropic | A tamper-evident record of every file that left through a connector you approved, naming the rule that applied and what it changed |
Five more rows, with their sources
| Who checks the answer | Not documented as a separate, independent check. Harvey states its system hallucinates less than the underlying foundation model, but does not publish how that check happens or how well it holds up | Not documented as a separate, independent check | Each step is checked against your documents as the answer is written, and one that drifts is corrected then, rather than by a second pass over a finished draft |
|---|---|---|---|
| Whose model it runs on | OpenAI, Anthropic and Google models routed on Harvey's platform. Harvey has also announced Harvey Tenet, a research preview of its own model post-trained on an open-weight base, on 20 August 2026 | The third-party frontier models Spellbook connects to | The open-weights model your firm chooses. The firm keeps its own copy, whoever runs the machine it sits on |
| Hallucination rate, independently published | Not published at the task level, per public reporting. Harvey has published its own internal, claim-level figures (October 2024), which are not an independent audit | Not published, per available public sources | No independently audited rate, and the benchmark published above is not one either. Answers come with their source, and one your files do not support is held rather than written |
| Risk of an ungrounded frontier model on legal questions | A 2024 Stanford study found general-purpose AI models, used directly on real legal questions with no independent source check, gave an incorrect answer 58 to 88% of the time depending on the model. This describes the risk of using a frontier model raw. It is not a published number for Harvey or Spellbook's actual product | Your own files outrank the model's training data, and an answer drifting off them is checked as it is written. This is a substantial reduction in that risk, not a claim to have removed it | |
| If you ever switch tools | Export terms are not documented publicly. The models stay with Harvey and its providers | Your contracts stay in your own Word files. The models stay with Spellbook's providers | Your model and your data leave with you |
Neither publishes an equivalent to Sidian Chambers: Router, Grounding and Governance as separate, visible controls. The table compares what each company documents publicly, not what happens inside either one.
Sources (15)
- GPQA Diamond is the hardest subset of GPQA (Rein et al., 2023), a graduate-level multiple-choice science benchmark on which PhD-level experts scored 65% and skilled non-experts with unrestricted web access scored 34%
- The Gemini 3.1 Pro Preview, GPT‑5.6 Sol Pro, Claude Fable 5.1 and Qwen3.6 27B scores are OpenRouter's own evaluations from its 19 September 2026 run, one fixed question set through each provider's live endpoint (openrouter.ai/benchmarks/gpqa-diamond). Each of the first three is its lab's highest-scoring model on that run among those publicly available on 1 September 2026. GPT‑6 Astra (released 3 September, 94.4) is outside that date, and GPT‑5.5 ties GPT‑5.6 Sol Pro at 93.8. OpenRouter keeps only its latest run, so these are 19 September scores for the models available on 1 September
- The technical specification behind Sidian Chambers requires deployment "100% locally on firm hardware or Virtual Private Clouds (VPCs)", with no logs and no document content held outside the firm. When Sidian operates the machine, it provisions one for that firm alone
- The 92.4% with Chambers on Qwen3.6 27B and the 148x size comparison come from the September 2026 technical overview for the control layer behind Chambers. The 92.4% has not yet been independently verified. Qwen's own model card reports 87.8 for Qwen3.6 27B under its own test settings; the chart uses OpenRouter's 80.1 so that every model on it was scored in the same run
- Harvey does not publish pricing. Third-party reporting (The Legal Prompts, 18 July 2026) describes it as priced per seat, for firms of 50 to 200 attorneys
- Spellbook does not publish pricing either: "Our pricing is structured around the number of team members on your license" and "Book a demo to talk through your team's needs and we will provide customized pricing", quoted from spellbook.com/pricing, checked 20 September 2026
- Harvey added Anthropic and Google models alongside the OpenAI models it already ran on 13 May 2025, "integrated through their respective cloud providers (AWS Bedrock, Google Vertex)", quoted from harvey.ai/blog/expanding-harveys-model-offerings
- Harvey announced a research preview of Harvey Tenet, its first post-trained model, on 20 August 2026, post-trained together with Fireworks on Moonshot AI's open-weight Kimi K3 base (harvey.ai/blog/post-training-update-harvey-tenet)
- Spellbook: "It uses GPT-5, Claude, and other LLMs to help legal teams draft and review contracts", quoted from spellbook.com's own page description, checked 18 September 2026
- Harvey: SOC 2 Type II, ISO 27001, ISO 27701 and ISO 42001 are listed alongside GDPR and CCPA, and ethical-wall sync is documented at harvey.ai/security, checked 18 September 2026
- Harvey states "we don't use inputs, outputs, or uploaded documents to train underlying models" and requires Zero Data Retention from model providers (harvey.ai/security); Spellbook states it has zero-data-retention agreements with OpenAI and Anthropic (spellbook.com/security). Both checked 17 September 2026. Neither publishes a per-file record of what left a firm, which is what the privilege row compares
- Harvey: "Harvey's models hallucinate at a lower rate than the foundation models", from its own internal claim-level evaluation, not an independent audit (harvey.ai/blog/biglaw-bench-hallucinations, 7 October 2024)
- Neither Harvey nor Spellbook has published an independently audited, task-level hallucination rate as of this writing
- The 58 to 88% figure is Dahl et al.'s "Large Legal Fictions" (2024) finding for general-purpose models used raw on legal questions, where a hallucination is an incorrect answer. The widely cited Stanford HAI and RegLab legal-AI study from the same year tested Lexis+ AI, Westlaw AI-Assisted Research and Ask Practical Law AI with GPT-4 as a baseline. Neither tested Harvey or Spellbook
- Mata v. Avianca is a real, widely reported sanctions matter involving fabricated case citations from ungrounded AI use, referenced here generically as a well-known industry example
See it on your own matters.
Fifteen minutes with Ben, Sidian's founder. Bring a routine question, no client details, and watch the answer come back with its source. Tell us your headcount and we'll give you your firm's price.
15 minutes, no commitment.