XIO Compliance Brain

AMD hackathon submission: the MI300X / Qwen 2.5 72B Triad Review demo lives on branch feat/amd-mi300x-vllm. Start with the 90-second judge demo, then inspect the release assets.

Canadian legal workbench. A lawyer drops a document, picks what kind of review they want, gets structured AI-assisted output grounded in Canadian law, verifies every citation, and exports a filed-work-product-ready DOCX — with a full audit trail from the first prompt to the final file.

What a lawyer does (5 steps)

1. Go to /matters/new and pick the task type. Eight to choose from, from securities OM review to PIPEDA privacy review to court AI-use disclosure memo to contract redline.
2. Drop the document. PDF, DOCX, or plain text. The app classifies it, chunks it, creates a matter, and routes it to the right reviewer persona.
3. The review streams in with structured [c1] citations. Every citation carries source-locker metadata (jurisdiction, source type, authority date, confidence). The footnote list lands already colour-coded — green ✓ verified against the seed corpus, blue ↗ canlii with a click-through to verify manually, or red ✕ not found for citations to investigate.
4. Request approval when the review looks right. Approval binds to the SHA-256 of the output text — any edit invalidates a prior approval, so the signoff is tied to the exact artifact.
5. Export DOCX (or redline DOCX for contract-redline matters). Export is blocked until the binding approval is in place. Exhibits appendix carries the source-locker metadata for every cited authority.

Every step writes to the hash-chained audit log: query → retrieval → generation → verdict → approval-requested → approval-granted → export.

Task types

Task	What it produces
om-review	OM compliance report: required-disclosures checklist, gap memo, risk flags, resale-restriction check, post-filing checklist
kyc-gap-check	KYC/AML gap report against NI 31-103 Part 13 + PCMLTFA + FINTRAC guidance
marketing-signoff	Flagged-claims table + required-disclosure checklist + revised-language suggestions for pitch decks, one-pagers, social posts
response-memo	Point-by-point draft response to an OSC / CIRO / FINTRAC / AMF deficiency or inquiry letter
court-ai-disclosure	AI-use disclosure memo for filed court materials — Federal Court / Ontario SC / ABKB / BCSC practice-direction compliant
pipeda-check	PIPEDA conformance review (Schedule 1 principles + s. 10.1 breach regime + Quebec Law 25 / AB PIPA / BC PIPA)
missing-authority-scan	Citation-risk audit over an existing matter's output — flags low-confidence, stale, wrong-jurisdiction, and uncited assertions
contract-redline	Inline redline of an uploaded draft → DOCX with strikethrough deletions, underlined insertions, footnoted rationale

Canadian source packs

Every matter is scoped to named authority bundles the wizard shows as chips before the review runs. Five packs ship by default:

Pack	Scope
ca-securities-ontario	NI 45-106 + companion, NI 31-103 Part 13, NI 81-102 Part 15, OSC Rule 45-501, Securities Act (Ont) s. 130.1, CSA notices, OSC/CIRO/FINTRAC deficiency patterns (~120 items)
ca-federal-aml	PCMLTFA s. 6.2, FINTRAC Guideline 6G, FINTRAC examination patterns
ca-consumer-protection	Pan-Canadian — federal + every province + multi-provincial harmonization
ca-privacy	PIPEDA Schedule 1 + s. 10.1 + regs, OPC guidance, Quebec Law 25, AB PIPA, BC PIPA (9 items)
ca-court-ai	Federal Court + Ontario SC + ABKB + BCSC AI practice directions, LSO + CBA guidance (7 items)

The matter wizard routes each task type to the right pack(s). GET /api/source-packs?taskType=... returns the live selection.

Key guarantees

• Every citation has structured provenance. { authorityId, section, quote, jurisdiction, sourceType, authorityDate, confidence, pinpoint }. The reviewer's UI renders pill badges for each. The DOCX export's Exhibits appendix carries them through.
• Every review's citations are auto-verified. The server runs the offline-corpus + CanLII URL-heuristic verifier after each review and streams the results as a verifications SSE frame. The matter page lands pre-coloured.
• Every export is hash-gated. /api/matters/[id]/export and /api/matters/[id]/export-redline return 403 approval-required unless an approved approval binds to sha256(output). Admin override via X-Approval-Override: <reason> (audited).
• Every persona refuses to refuse. When retrieval is partial, reviewers mark rows [NEEDS VERIFICATION] and keep producing work product. Meta-refusals ("I can't review because the corpus is incomplete") are explicitly disallowed.

Quickstart

pnpm install
pnpm --filter @compliance-ai/web dev

Open http://localhost:3000/matters/new, pick a task type, drop a file.

For persisted production mode:

export DATABASE_URL=postgres://user:pass@localhost:5432/compliance_ai
pnpm --filter @compliance-ai/db db:migrate
psql "$DATABASE_URL" -f packages/db/rls/policies.sql
pnpm --filter @compliance-ai/web dev

Preview mode is intentional when DATABASE_URL and NEXTAUTH_SECRET are unset: the app uses in-memory stores and a synthetic preview session for anonymous demos.

Provider modes

The agent runner supports three provider modes:

• Hosted preview: LLM_PROVIDER=openai, OPENAI_API_KEY, optional OPENAI_MODEL.
• Private/local: LLM_PROVIDER=ollama, OLLAMA_BASE_URL, optional OLLAMA_CHAT_MODEL.
• Legacy: LLM_PROVIDER=anthropic, ANTHROPIC_API_KEY.

If LLM_PROVIDER is unset the runtime picks OpenAI when OPENAI_API_KEY exists, Anthropic when only ANTHROPIC_API_KEY exists, otherwise local Ollama.

Railway

Railway uses:

• railway.json healthcheck: /api/healthcheck
• nixpacks.toml build: pnpm --filter @compliance-ai/web build
• root start command: pnpm start, which binds Next to 0.0.0.0

Minimum hosted preview variables:

LLM_PROVIDER=openai
OPENAI_API_KEY=sk-...
OPENAI_MODEL=gpt-5.4-mini
NEXTAUTH_URL=https://compliance-ai-preview-production.up.railway.app

Add DATABASE_URL and NEXTAUTH_SECRET when persistence and real auth are required.

Verification

pnpm install
pnpm test            # 533 tests across agents / cognition / web
pnpm -r typecheck
pnpm --filter @compliance-ai/web build
pnpm smoke:demo      # full route + endpoint smoke (source-packs, verify, export-redline)

pnpm smoke:demo expects a running app at http://127.0.0.1:3000, or set DEMO_BASE_URL to test a deployed URL.

HTTP surface

GET   /api/healthcheck                          cognition + db + auth + provider + version + uptime
GET   /api/agents                               persona registry + timeline participants
GET   /api/source-packs                         list packs; ?taskType= and ?lane= filter
POST  /api/quick-review                         upload + classify + create matter
POST  /api/matters                              create matter from wizard
GET   /api/matters/[id]                         matter detail
POST  /api/matters/[id]/review                  SSE stream: thinking → prose → citations → verifications → verdict
POST  /api/matters/[id]/chat                    SSE stream: matter-scoped follow-up
POST  /api/citations/verify                     offline-corpus + CanLII URL verification for a citation batch
POST  /api/matters/[id]/export                  DOCX with exhibits appendix — gated on approved hash
POST  /api/matters/[id]/export-redline          DOCX redline (strikethrough/underline/footnote) — gated on approved hash
GET   /api/matters/[id]/handoff                 MatterContextBundle + DOCX handoff pack
GET   /api/matters/[id]/transcript?fmt=jsonl    persona events, chronological
GET   /api/matters/[id]/graph                   pure evidence graph
POST  /api/approvals                            reviewer creates / approves / rejects a request bound to an output hash

Monorepo layout

apps/web
  src/app/matters/new                 wizard (8 task types + source-pack chips + presets)
  src/app/matters                     matter workspace, timeline, graph, verification badges
  src/app/api/source-packs            lane-filtered pack listing
  src/app/api/citations/verify        batch verifier endpoint
  src/app/api/matters/[id]            review, chat, export, export-redline, transcript, graph, handoff
  src/lib/auto-verify.ts              server-side auto-verify helper
  src/lib/redline-diff.ts             inline diff token parser

packages/agents                       8 task-specific personas + retrieval plans + judge loop + citations
packages/cognition                    seeded authority corpus, hybrid retrieval, source-locker, source-packs, verifier
packages/approvals                    hash-bound approval state machine
packages/chat-structure               registry, mentions, typed tools, transcripts
packages/ingest                       PDF/DOCX/TXT parse, chunk, classify
packages/db                           Drizzle schema, migrations, RLS
packages/frameworks                   framework metadata (SOC 2, GDPR, EU AI Act) for the infosec surface
packages/prioritizer                  risk-queue scoring

scripts/smoke-demo.mjs                end-to-end smoke

Design decisions

• Source locker on every citation: jurisdiction + type + date + confidence are first-class fields, not prose. Without structured provenance, a lawyer can't verify AI output at scale.
• Multi-query retrieval plans per task type: each reviewer has 10–18 short targeted BM25 queries, unioned and deduped, so the persona sees the full authority set its checklist demands. Replaces the single-pass-over-document-text retrieval that kept hitting "corpus is incomplete" refusals.
• Never-refuse persona clauses: reviewers produce partial output with [NEEDS VERIFICATION] flags rather than refusing when retrieval misses.
• Auto-verify on review completion: citations are checked against the seed corpus + CanLII URL heuristics as a follow-up SSE frame, so the user never has to click "Verify citations" before trusting the output.
• Hash-bound signoff: approval binds to sha256(output), so post-approval edits invalidate the approval. Export is gated server-side, not UI-only.
• Named source packs: the authority corpus is organised into named bundles keyed to matter lane. A lawyer sees exactly which packs will be in scope before running the review.

License

Apache-2.0