Run the 5-category validator: frontend bindings, API contract, code hygiene, schema sync, README currency.
Validator and pre-flight gates are live. Modules 1–4 continue to run on the legacy UI while we migrate screens.
Run the 5-category validator: frontend bindings, API contract, code hygiene, schema sync, README currency.
Five-category code health check. Run before every deploy. Results saved to validator_reports/.
Press Run validator to start.
—
Upload research records or case files. Doc-type controls obfuscation policy: research files are indexed with real names; case files are obfuscated before embedding.
Point the server at a local folder or network path. Files are ingested server-side — no browser upload needed. Uses the doc type selected above (Research = no obfuscation, Case = full obfuscation).
Every conversation is saved as a brief you can name, search, and export. Give this session a name or use a quick-start below.
Ctrl+Enter to send · answers grounded in your document library
Upload input files for a matter. Choose a category so files land in the correct folder. All files are PII-redacted before any AI sees them — you review and approve the redacted version first.
Research briefs are private to you — no other user or firm can access them. Add answers from Converse with AI, then export to Word.
Review and correct extracted metadata. Values flagged as garbage will be nullified, update values will be corrected. Click any value to edit inline. Use Learn to teach the extractor to never store a pattern again.
This module is being migrated into the new mobile-first SPA. The working version remains available in the legacy UI.
Continue working with this module on the legacy interface:
Top-level case folders. Each case holds court dates, briefs, and attached documents.
Promote filed briefs to the research library? Briefs that scored red on the hallucination meter will be blocked automatically — review and override only if necessary.
Corpus-wide health check. Detects semantic collapse (chunks bunching toward a centroid) and source bias. Complements the per-answer hallucination meter shown on every Ask query.
First run takes a few seconds. Results are not persisted.
—
No route matches —. Try the dashboard.
logs/initial_admin.txt.
Minimum 8 characters with at least one letter and one digit. Changing your password will sign you out of all other sessions.
Minimum 8 characters with at least one letter and one digit.
All law firms on the platform. System Admin only.
Create a new tenant + first Firm Admin. The temporary password will be returned once — share it securely with the new Firm Admin.
Share this temporary password with the new Firm Admin (one-time view):
Back to tenantsTemporary password (one-time view):
Back to usersPhase 3: query log audit trail, hallucination distribution, and cache-readiness signal.
Loading…
Programmatic access for your firm. Keys are shown once at creation — store them securely.
Copy this token now. It will not be shown again.
Public legal research, shared across all firms. Any signed-in user can contribute. Documents go through PII checks then become searchable to everyone.
Briefs submitted for promotion to the shared research library. Approve to publish; reject to keep private. The hallucination guard already blocks red briefs — this is the human-vetting layer.
Seven signals continuously checked for semantic collapse and hallucination drift. Two are hard stops (citation drift > 40%, thumbs-down rate > 15%). The fabricated-citation check is zero tolerance.
Before using BriefProposal, you must acknowledge the four Model Rules of Professional Conduct that apply to AI-assisted legal research:
BriefProposal is a research tool. It does not provide legal advice. You remain responsible for every word you file. Acceptance is recorded with your user account and timestamp; this acknowledgement cannot be bypassed.
The router classifies every query at intake and predicts which model tier should serve it. In Phase 4b this runs as a passive observer — predictions are logged but routing is unchanged. When ≥200 predictions accumulate, Phase 4c calibration becomes feasible.
Multi-turn research conversations preserved with audit. After every third turn, a Haiku summary replaces the conversation history in the LLM prompt — the full history stays in the audit log.
Run a single-advocate trial: pick a case type and role, ask a research question, see the cost estimate, then run. Phase 7b will add a second advocate; Phase 7c will add a judge.
Actual API provider costs broken down by activity type, user, model, and day. Every LLM call — metadata extraction, RAG queries, courthouse trials, distillation, embeddings — is recorded here with actual token counts. Use this as your audit trail and cost-to-feature traceability.
Multi-round adversarial trials: For opens → Against rebuts → For replies → Judge scores. Up to 10 rounds per batch, extend up to 50 total. Free mode for now (Directed mode lands in Turn 4).
No judge profiles yet. Extract a profile from your research library to get started.
Profiles are auto-extracted from research documents that contain judge name metadata.
| Summary | Jurisdiction | Case Type | Side | Effectiveness | Won |
|---|
| Chunk ID | Reason | Preview | Quarantined At |
|---|
| Run ID | Model | Pass Rate | Severity | Date |
|---|
Check if a party name conflicts with existing firm clients or opposing parties in your case database.
| Party | Result | Type | Checked At |
|---|
| Feature Key | Display Name | Category | Default | $/use | $/month | Active Firms | Actions |
|---|
| Firm | Feature | Category | Request Notes | Requested At | Actions |
|---|
After uploading documents, run these three steps in order to make them searchable. Each step picks up where the previous left off — safe to restart if interrupted.
Extracts text from PDFs, Word docs, and text files.
Research documents: text kept exactly as-is — judge names, party names, citations fully preserved.
Case files: PII obfuscated before storage.
Pulls structured metadata (court, judge, year, area of law, outcome, key holding) out of each research document so retrieval can filter on these fields. Combines heuristic regex with an LLM pass and flags documents where the two disagree for human review. Run after Step 1; required before Step 2 for filtered retrieval.
Splits each document into parent and child chunks for precise retrieval. Run after Step 1 completes.
Sends chunks to the embedding model and stores them in ChromaDB. After this step, documents are fully searchable via Converse with AI.
Documents that could not be processed by the text extraction worker. Each row is an attempt — re-uploading or retrying produces a new entry, so you can see the full history per document.
Documents where research-metadata extraction either failed completely or was flagged uncertain (heuristic and LLM disagreed on key fields). Edit fields inline in the Uncertain table, or bulk-accept to clear the flag.
Search all documents. View and edit extracted metadata inline. Hover any cell to see full value.
| File | Docket # | Cause # | Subject | Court | St | Year | Judge | Petitioner counsel | Respondent counsel | Area | Outcome | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Loading… | ||||||||||||
Upload examples of filed complaints and briefs so the AI learns your firm's preferred structure and format. Templates are firm-wide — shared across all matters.