# Sovereign Desk **Private on-premise AI for Colorado law and CPA firms.** Document review, intake triage, and drafting that run on hardware inside the firm's own office. Client files never leave the building, and the software produces the log that proves it. A RIG offering. Built and supported by Mike Rodgers, Denver CO. --- ## What is in this directory | File | What it is | |---|---| | `app/` | The working application — plain HTML/CSS/JS, no build step, no dependencies | | `app/index.html` | Application shell, four tabs | | `app/styles.css` | All styling | | `app/app.js` | Rules engine, triage engine, drafting engine, fleet client, egress ledger | | `app/samples.js` | Sample corpus: 5 documents, 4 intake emails, 10 conflict records, 4 templates, 19 clauses | | `landing.html` | Sales page with a live iframe of the real app | | `icon.svg` | Hand-authored 512×512 mark | | `icon-256.png` | 256px raster export | | `GTM.md` | ICP, trigger events, 5-email sequence, 3 LinkedIn posts, objections, 10 prospect types | | `PRICING.md` | Tiers, replacement math, COGS, margins, capacity ceiling, discount policy | | `README.md` | This file | --- ## How to run it The app is static. Two ways to open it. **Recommended — local web server** (needed for the fleet calls, because Ollama echoes a specific `Origin` header and a `file://` origin is `null`): ```bash cd ~/Developer/rig-offerings/sovereign-desk python3 -m http.server 8080 --bind 127.0.0.1 # then open: # http://127.0.0.1:8080/app/index.html <- the workspace # http://127.0.0.1:8080/landing.html <- the sales page ``` **Double-click** `app/index.html` also works. The rules engine, triage, drafting, ledger, and all sample data function normally over `file://`. The blackwell node (vLLM, `Access-Control-Allow-Origin: *`) still answers; the Ollama nodes may not, because of the origin echo. No install step. No npm. No bundler. That is deliberate — a firm's IT contractor has to be able to read every line of this. --- ## What the app actually does ### Document Review Five sample documents load on open: a Colorado commercial lease, an unexecuted vendor MSA, a mutual NDA, an engagement letter, and an IRS CP2000 notice. Paste or upload your own (`.txt`, `.md`, `.eml`). A **deterministic rules engine** runs in ~1–20 ms with no model involved and flags: - governing law and whether it is Colorado - mandatory arbitration and its venue, jury-trial waivers - auto-renewal clauses **and the notice window in days** (it reads both sides of the sentence, so "not less than one hundred eighty (180) days prior … shall automatically renew" resolves to 180) - indemnity with no limitation of liability - one-way vs. mutual indemnity, and which party carries it - assignment permitted without consent - data-boundary language: subprocessors, out-of-state processing, "to improve its models" - cure and notice periods, insurance limits, annual escalation with the compounded five-year effect - unfilled placeholders (`[STATE]`, `____`, `TBD`) - hard deadlines, with days-until computed against today's date - missing signature block Each finding carries the **literal quoted span**, is severity-weighted into a 0–100 risk score, and clicking it scrolls to and highlights that text in the document. Extracted facts (parties, dollar figures, dates, governing law, cure periods) are pulled out separately. **"Deep read on fleet"** then sends the document to a real GPU node and returns a four-section attorney review. This is a live network call to actual hardware. ### Intake Triage Paste an inbound client email (or click one of four samples). Locally, before any model runs: - classifies practice area against a weighted keyword model across 6 areas, shows the runner-up and the matched signals - extracts a stated deadline ("21 days", "due March 3, 2026") and computes days remaining - parses incident age from natural language ("14 months ago", "3 years and 2 months ago") - runs the **Colorado limitations clock** — CADA/EEOC 300-day charge windows, C.R.S. 13-80-101 three-year contract and motor-vehicle periods, 13-80-102 two-year negligence, 13-80-103.5 six-year liquidated debt, CP2000 and SNOD procedural windows — and marks each open / tight / critical / barred - runs a **conflict check** against the firm client list and flags current clients and adverse parties with the RPC 1.7 prompt - produces a ranked next-actions list **"Draft reply on fleet"** generates an intake response on your own hardware. When a conflict is detected the prompt is changed so the reply refuses to discuss facts — that behavior is verified working (the Ortega sample hits Front Range Capital Partners LP, an existing client). ### Drafting Four templates (engagement letter, demand letter, records request, mutual NDA) over a 19-clause library. Assembly is **deterministic string merge** — no model, no hallucination surface. "Polish on fleet" is optional and clearly labelled as the model layer. Copy or download as `.txt`. ### Sovereignty - live latency probe of all four fleet nodes, with a routing selector - **request ledger**: every outbound call the app makes, classified LAN or WAN by RFC1918 + CGNAT 100.64/10 + loopback matching. Exportable as JSON. - egress byte counter in the header, red the instant anything non-private is contacted - session data inventory and the written data-boundary policy --- ## What is real vs. what is stubbed ### Real — verified, not asserted - **The app runs.** Opened in Chrome, all four tabs exercised via Playwright, **zero console errors, zero page errors.** Screenshots taken of every tab. - **The rules engine is real code.** It found 10 findings and a 100/100 risk score on the vendor MSA, including the Delaware governing law, the uncapped indemnity, the 90-day auto-renewal window, the San Francisco arbitration venue, the blank insurance limit, and 7 unfilled placeholders — each with its quoted span. It found the 180-day auto-renewal window in the lease. - **The fleet calls are real.** Measured in-browser during verification: `blackwell 100.67.126.117:8000` — doc review 4,403 B up, HTTP 200 in 1,479 ms; intake reply 1,369 B up, 200 in 1,060 ms. All four nodes probed 200 at 9–24 ms. - **The triage engine is real.** The Ortega sample classifies as Tax controversy (signal 57), urgency Same-day, extracts the 21-day deadline, anchors the CP2000 window to it, and returns 2 conflict hits including a current client. - **The drafting engine is real.** Produces a complete, correctly merged demand letter and engagement letter with live field edits. - **The egress ledger is real.** It wraps every `fetch()` the app makes; nothing bypasses it. - **The sample corpus is real text**, generated on the fleet (blackwell + rig-96gb + rig-36gb + rig-28gb) and hand-checked — not lorem ipsum. The MSA was hand-authored to carry the exact defects the engine is built to catch, which is why it scores 100. - **The landing page renders.** Screenshotted at 1500×1050 across every section, zero page errors, and the demo section iframes the actual application rather than a picture of it. - **The public sourcing channels in `GTM.md`** were each fetched live and return HTTP 200. ### Stubbed, simulated, or not built - **No persistence.** Documents live in browser memory. Reload destroys everything. That is a demo-honest choice, not a shipped-product choice — a real install needs local disk storage with matter-level access controls, and that is Vault-tier work. - **No authentication, no user accounts, no per-matter permissions.** Single-user demo. - **No DMS integration.** No NetDocuments, iManage, or Clio connector. Paste, drop, upload only. This is stated plainly in the landing page FAQ rather than hidden. - **No PDF or DOCX parsing.** Plain text only. A production install needs `pdf.js` and a `.docx` extractor; both are straightforward, neither is built. - **The conflict list is 10 fictional records**, not a real client database. - **The Colorado limitations table is a triage aid, not legal advice**, and is labelled as such in the UI. It covers six practice areas at the headline level. It is not exhaustive, does not model tolling, discovery-rule accrual, or minority, and would need a Colorado attorney's review before a firm relied on it. - **The fleet nodes are RIG's hardware, not a customer's.** In a real install this is one appliance in the client's office. The four-node fleet in the demo is what makes the routing selector interesting; a customer sees one node. - **No installer, no provisioning automation, no monitoring agent.** Setup today is manual work by Mike. That is fine at 2 accounts and breaks at 15. - **Apollo prospect list was NOT pulled.** The live key in `com.rig.env.APOLLO_API_KEY` is on a Free plan — `mixed_companies/search` returns HTTP 403 `API_INACCESSIBLE` and `organizations/enrich` returns 422 insufficient credits. `GTM.md` documents this and uses manually verified public sources instead. No fabricated prospect list was written. - **Nothing has been sent.** No emails, no LinkedIn posts, no outreach of any kind. All GTM copy is drafted and staged only. Gate-D applies. - **The three LinkedIn posts were drafted on the fleet and then rewritten by hand**, because the model versions invented client counts ("I've helped dozens of firms") and a war story that never happened. The published versions contain only arithmetic and cited authority. ### Known rough edges - The rules engine is regex-based. It will miss clauses phrased unusually and can double-count a cure period that appears twice. It is a first-pass tool and the UI says so. - Entity extraction over-captures on documents with many capitalized headings. - `max_model_len` on the blackwell vLLM node is 16,384 tokens, so documents are truncated to 9,000 characters for the deep read. Long agreements need chunking, which is not built. - No mobile layout for the app (the landing page is responsive; the workspace is desktop-only by design). --- ## Fleet used to build this | Node | Endpoint | Model | Used for | |---|---|---|---| | blackwell | `100.67.126.117:8000` (vLLM) | Qwen3-Coder-30B-A3B-Instruct | sample lease, intake emails, LinkedIn drafts, live doc review | | rig-96gb | `100.102.142.84:11434` | qwen3-coder:30b | sample NDA | | rig-36gb | `100.89.143.27:11434` | qwen3-coder:30b | sample intake batch | | rig-28gb | `100.76.209.22:11434` | qwen3-coder:30b | engagement letter + CP2000 notice | All four are on the private network (100.64/10 CGNAT). Verified reachable and CORS-open from a browser at build time. --- ## License and ownership Not open source. This is a commercial RIG offering. Customer installs include a source and model-weight escrow clause — see `PRICING.md` §1.