The search got sharper, not noisier.
This pass used the authenticated GitHub CLI/API session and read repository metadata, README/architecture material and open issues. The token stayed in the local keyring; no credential is published here. The 36 records below are a curated follow-up set, deliberately separate from the original 127-screened corpus.
Decision unchangedLegalise remains the closest whole-case reference: matter → selected-source AI → citations → named sign-off → hash audit → verifiable export. It is an evaluation release, not a live dependency. The new repos make the missing product surface, ingestion workers, portal/API, approval durability and failure tests much more concrete.
Case authority
Build custom on the VPS
Provider cases, people, category/form editions, evidence lineage, gaps, approvals, declarations, submission events and restore semantics stay in our PostgreSQL/application layer.
Product surface
Borrow the workspace patterns
Mike, draftLegal, AI WorkDeck, Judicex and doc.haus show matter navigation, document libraries, citations, redlines, templates, versioning and a conversational rail.
Workers + control
Activate only by fixture
LaunchStack, the intake and approval demos, Docket, Temporal and Aulite are probes for ingestion, waits, identity, audit and model-egress—not competing authorities.
The gold finds.
“Fit” is fit to Harrison’s provider-application case, not popularity. A high score means the repository teaches a workflow or boundary we can test; it does not mean it is safe to fork.
| Repo | Stars / licence | What it actually does | Use / verdict |
|---|---|---|---|
| Legalise | 24 · MIT | Matter-first loop: selected-source AI/skills, source anchors, named sign-off, hash audit and verifiable export on a self-hosted stack. | Fit 5/5 · God-source blueprint and synthetic spike; never live unchanged. |
| Mike | 4,167 · AGPL | Legal workspace with matters, document library, assistant and tabular-review workflows, citation verification, Word add-in and Ollama path. | Fit 5/5 · feature reference; inspect citation and upload failures first. |
| LaunchStack | 886 · Apache-2.0 | Ports-based TypeScript engine for ingestion, OCR, RAG, graph, LLM providers, storage and durable background jobs. | Fit 5/5 · strongest new worker/host boundary; packages not yet on npm. |
| Docket | 0 · AGPL | Sovereign cases, customer portal/SSO, on-behalf-of API, scoped service accounts, webhooks, hash-chain audit and full backup/restore. | Fit 5/5 · portal/API/audit pattern; service-desk semantics stay out. |
| AI WorkDeck | 79 · AGPL | One legal/document workspace with tracked DOCX changes, OCR, tabular/due-diligence views, plugins/MCP and evidence pane. | Fit 5/5 · best workspace reference; provenance is still roadmap. |
| draftLegal | 14 · AGPL | Seven LangGraph agents for extraction, portfolio RAG, redline, approvals, e-sign and obligations on Postgres/Redis/Elasticsearch/MinIO/Gotenberg. | Fit 5/5 · feature benchmark; demo credentials and AGPL need a hard boundary. |
| Judicex | 51 · Apache-2.0 | Separate citable sources from operational notes; version sources; fail closed with grounded/limited/abstain answer states and workflow packs. | Fit 5/5 · strongest authority-taxonomy lesson; alpha SQLite, not base. |
| doc.haus | 68 · MIT | Private matter indexes, cited answers, Word tracked changes, templates, routed assistants and multi-agent full review. | Fit 4/5 · matter-private UX and template scrubbing probe. |
| LawLink | 86 · MIT | Self-hosted PostgreSQL/Prisma flow from intake and conflict check through formal case, follow-up, archive and export. | Fit 4/5 · case-state baseline; early project with a blocking new-case bug. |
| Tabular Review | 77 · MIT | Docling conversion, dynamic extraction columns, exact source-quote highlighting and analyst chat. | Fit 4/5 · master-pack evidence-grid worker/UI probe. |
| LexReviewer | 18 · metadata absent | Qdrant + BM25 retrieval, LangGraph tool choice, streamed citations, bounding boxes and linked-document awareness. | Fit 4/5 · citation and schedule-link probe; external dependencies. |
| regulated-workflow-demo | 0 · MIT | Local text/CSV/JSON/PDF extraction and old/new diff into canonical JSON, review queues, Excel and offline audit; every decision starts pending. | Fit 5/5 · closest evidence-register fixture; not production/OCR. |
| Temporal approval system | 0 · MIT | FastAPI → Temporal → reviewer signal, reminder, escalation and idempotent audit activity. | Fit 4/5 · durability probe only when direct workers fail. |
| Document Intake Agent | 0 · no metadata | MCP classify → extract → anomaly flag → schema validation → pending approval, with n8n import and 21 adversarial fixtures. | Fit 5/5 · tiny synthetic contract-test harness. |
| AI-Agent-Casebook | 15 · Apache-2.0 | Customer onboarding assistant plus LangGraph/Agents SDK, RAG, evaluator and multi-provider examples. | Fit 4/5 · onboarding/eval shape; open failures are valuable tests. |
| Hermy HQ | 364 · metadata absent | Postgres message bus between a self-hosted agent and dashboard; approval inbox, run history, memory wiki and onboarding. | Fit 3/5 · internal agent front door, never case authority. |
| Open Specter | 31 · AGPL | Projects, versioned documents, assistant/tabular workflows and RLS-backed content tables. | Fit 4/5 · compare data shapes; Supabase conflicts with VPS direction. |
| Aulite | 128 · BUSL-1.1 | Self-hosted AI proxy with deterministic risk rules, hash-chain SQLite audit and PDF reports. | Fit 3/5 · model-egress sidecar after core audit. |
| ContractsPulse | 3 · MIT | Svelte/FastAPI contract cockpit with risk inbox, version verification, redlines and calendar. | Fit 3/5 · UI pattern only; insecure demo defaults. |
| n8n AI starter kit | 15,219 · Apache-2.0 | Compose sandbox for n8n, Ollama, Qdrant and PostgreSQL local AI/RAG experiments. | Fit 2/5 · disposable VPS spike, explicitly not production. |
| ITFlow | 993 · GPL-3.0 | Self-hosted PSA with client/contact/vendor/document records, tickets, billing and portal. | Fit 3/5 · portal/navigation pattern; no evidence semantics. |
| Counsel MVP | 1 · metadata absent | Single matter workspace with persistent assistant plus research, timeline, letter and contract modules. | Fit 4/5 · module-routing sketch, not production. |
The complete machine-readable record, including “how it works,” use boundaries and verdicts, is in github-gold-mine.json. Stars and licence labels are a dated API snapshot, not an endorsement.
Second wave: provenance, truth admission and version control.
These are less about another chat UI and more about making the system honest about where a claim came from, when it became durable and what version a reviewer actually saw.
| Repo | Stars / licence | Gold contribution | Use / verdict |
|---|---|---|---|
| Latence | 11 · Apache-2.0 | Typed CPU-first pipeline with quarantine, PII masking, exact spans/pages, graph edges and Parquet manifests. | Fit 5/5 · ingestion/provenance worker; pre-1.0. |
| Prethinker | 11 · metadata absent | LLM proposes facts; closed predicate packs and deterministic gates decide what becomes durable truth, with contamination lessons documented. | Fit 5/5 · authority-admission research; not general QA. |
| KL4A | 7 · Apache-2.0 | SOP/PDF/DOCX obligation claims carry exact byte spans and approve/reject/edit events before OKF/MCP handoff. | Fit 5/5 · master-pack knowledge probe. |
| DocGraph | 7 · MIT | Document graph with stale-policy, conflicting-claim, supersession and undocumented-export audits. | Fit 4/5 · source/edition drift probe. |
| LedgerKB | 6 · Apache-2.0 | Append-only evidence assertions projected into RAG, graph, OKF wiki and change reports. | Fit 5/5 · claim-history probe; operator bug recorded. |
| Distillery | 27 · MIT | Documents become grounded, deduplicated, provenance-carrying instruction/evaluation data. | Fit 4/5 · de-identified eval fixtures only. |
| Forensic analysis platform | 8 · AGPL | OCR → entity/relation graph → analyst verification → evidence-linked narrative. | Fit 4/5 · chain-of-custody interaction probe. |
| Writing Intelligence | 13 · MIT | Branches and simulates edits, then binds approval to the exact state read before merge. | Fit 4/5 · artifact-version approval probe. |
| GreatCTO | 86 · MIT | Two-gate agent flow (`gate:plan` → work/review/QA/security → `gate:ship`) with decisions and lessons. | Fit 3/5 · operator control pattern. |
| Alea Intake | 2 · MIT | Privacy-first structured legal intake surface with a small applicant-question boundary. | Fit 4/5 · minimum intake UX probe. |
| EP Legal MCPs | 3 · MIT | Local MCP connectors for CRM, client intake and matters; privileged data stays on the operator machine. | Fit 4/5 · explicit local integration pattern. |
| easydocs | 0 · AGPL | Every Word save is immutable; stale edits branch, merge produces redlines, approvals and client-copy push-back are audited. | Fit 4/5 · master-pack versioning probe. |
| apply-for-me | 17 · MIT | Never-guesses application agent: reads full source, parks unknown/legal-declaration fields and produces a read-back dashboard. | Fit 5/5 · missing-fact/declaration boundary pattern. |
| Forensics audit skill | 4 · MIT | Evidence integrity and chain-of-custody checks against NIST/ISO 27037. | Fit 4/5 · audit regression fixture. |
The second-wave records and their source links are also in the machine dataset. They are probes and patterns, not a permission slip to ingest Harrison’s files into third-party systems.
The bug ledger is the valuable part.
Stars show discoverability. Issues show where the abstractions fail under pressure. These receipts turn the search into acceptance tests rather than a shopping list.
| Receipt | Severity | Observed failure | Build response |
|---|---|---|---|
| Legalise #245 | High | pgvector is not load-bearing in the evaluation release. | Prove retrieval need before adding a vector service. |
| Legalise #9 | High | WORM role split, export, deletion retry, durable jobs and runbook remain live-matter gates. | Make each gate a Provider Compliance acceptance test. |
| Legalise #241 | Medium | Ed25519 signatures/external audit anchoring are deferred. | Define checkpoint/key custody before “immutable” claims. |
| Mike #391 | High | Meaning-inverting ellipsis can pass citation verification. | Validate exact spans and preserve the source excerpt. |
| Mike #8 | High | Large uploads hit 524 timeouts and look like CORS failures. | Resumable bounded jobs with status and receipts. |
| LaunchStack #313 | High | DB deletion can orphan storage objects. | Two-phase deletion, reconciliation and restore tests. |
| LaunchStack #365 | High | Embedding can send full document text to a third-party API. | Per-job egress posture and local option. |
| LaunchStack #335 | Medium | Optional Zod fields fail native JSON-schema mode. | Strict contracts plus deterministic fallback. |
| NocoBase #10416 | High | Rich-text API bypass reaches unsanitised HTML. | Adversarial sanitisation; no direct client-content exposure. |
| Directus #28161 | High | Bulk comments can bypass authenticated-user checks. | Narrow domain API and explicit actor checks. |
| Dify #41534 | High | Swagger-tool agent can make unauthorised calls. | Gateway allowlists, scoped credentials and approval. |
| n8n #37360 | High | Personal Agent can hang indefinitely. | Lease, heartbeat, deadline and terminal state. |
| n8n #37377 | Medium | MCP execute-workflow schema differs from runtime. | Pin connector contracts and test deployed endpoints. |
| Cognee #4839 | Critical | Dropped WHERE clause can silently ingest a whole table. | Tenant filter at DB boundary plus negative fixtures. |
| Cognee #4673 | High | Cloud add/search endpoints can hang. | Deadlines, circuit breaker and local fallback. |
| Agent Inbox #38 | High | Resume fails when a checkpoint ID is missing. | Durable case/job/version IDs. |
| Agent Inbox #84 | High | No authentication surface. | Keep identity in the application. |
| AI WorkDeck #200 | Medium | Load-sensitive E2E login truncates credentials. | Assert input values and retry at the test boundary. |
| LawLink #11 | High | Stage selection invalidates cause and blocks submission. | Server-side state-transition validation. |
| Casebook #141 | High | Large PDF indexing fails under embedding rate limits. | Queue/chunk/rate-limit and partial-index receipts. |
| Casebook #134 | High | Onboarding tool called with empty required fields. | Fail-closed tool contracts and missing-fact prompts. |
| Casebook #138 | High | Rejected plan crashes the supervisor. | Rejected-plan state with an explicit next action. |
| LedgerKB #13 | Medium | Embeddings docs disagree with config, and diagnostics claim a setting is locked before first index; the named repair command does not exist. | Operator diagnostics must reflect actual state and every repair command must be executable. |
What to combine—and what to keep apart.
| Layer | Reference pattern | Our boundary |
|---|---|---|
| Case + truth | Legalise, Judicex, Docket | One custom Provider Compliance case and evidence schema in Postgres; never a generic chat history. |
| Workspace UX | Mike, AI WorkDeck, draftLegal, doc.haus | Borrow screens and review interactions; rebind every action to actor, case, version and approval records. |
| Intake/extraction | LaunchStack, Tabular Review, Document Intake Agent, regulated-workflow-demo | Workers emit manifests, claims, source anchors, confidence and pending review; they cannot approve or submit. |
| Long waits | Temporal approval system, Docket reminders | Start with explicit DB jobs; add Temporal only when measured multi-day resume/failure tests fail. |
| Agent front door | Hermy HQ, n8n, Dify | Useful operator/integration surfaces; credentials, tools and side effects remain behind our gateway. |
| AI safety/evals | Aulite, Langfuse, Casebook evaluators | Record model posture and traces; keep business audit and source authority custom. |
The most useful compositionLegalise’s matter loop + Docket’s portal/API/restore mechanics + regulated-workflow-demo’s evidence-register fixtures + LaunchStack’s worker boundary + Judicex’s citable-vs-operational memory split. This is a build map, not a five-repo deployment.
Gold-pass probe gate.
Before forking or adding a repository as a runtime dependency, run it with synthetic/de-identified fixtures and record the exact commit. The probe is accepted only if it passes:
- 1Source and licence review
Read the code, release activity and licence; do not infer maturity from stars or README badges.
- 2Tenant/case negative test
Ask one case for another case’s document, memory, citation and export. It must fail closed.
- 3Version and approval test
Prove source supersession, pending → approved/rejected, resume and exact artifact hash behaviour.
- 4Failure and restore test
Interrupt upload, extraction, deletion and notification; verify retries, receipts, backup and clean-host restore.
- 5Egress and connector test
Pin model/embedding destinations and schema contracts; prove no document leaves the approved posture.
Direction for the next build stepUse the gold pass to sharpen the first vertical slice—not to expand the dependency graph: de-identified successful case + incomplete case → adaptive intake → evidence register → cited draft → gap route → named approval → generated pack → verifiable export.