Add singapore-court-legal — MikeROS connector for Singapore civil court procedure - #95
Add singapore-court-legal — MikeROS connector for Singapore civil court procedure#95Koulsami wants to merge 2 commits into
singapore-court-legal — MikeROS connector for Singapore civil court procedure#95Conversation
… of Court 2021) External plugin wrapping the Reasonex MCP connector: grounded answers over the Singapore Rules of Court 2021, citing governing rules with operator-verified case authorities where they attach, plus relevance-gated reasoning chains. Read-only; three tools (ask, conversation_turn, analyze_situation); API-key auth over streamable HTTP per CONNECTORS.md; honest abstention out of scope. Adds external_plugins/singapore-court-legal/ and one entry in .claude-plugin/marketplace.json. Connector endpoint: https://mcp.reshuffleai.com/mcp Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
All contributors have signed the CLA ✍️ ✅ |
|
I have read the CLA Document and I hereby sign the CLA |
|
Hi — flagging this PR for review when a maintainer has bandwidth. Quick context that may help triage: Plugin + connector are live. The singapore-court-legal plugin pairs with our Reasonex MCP server (mcp.reshuffleai.com/mcp), which is deployed and serving — a neuro-symbolic reasoner over Singapore's Rules of Court 2021 that returns deterministic, citation-grounded procedural determinations (no LLM in the reasoning step). |
…ure, corrected coverage) Refresh the connector config + docs to the live MCP surface post-ADR-0143: - resolve_procedure is the flagship, with routing-first descriptions (determine which ROC rule controls / the sequence / whether a step is compliant; proof-carrying; abstains rather than guess; not for interpretation/drafting/strategy/non-SG). - ask re-aimed (single-rule pinpoint + false-premise); conversation_turn (direct client); analyze_situation noted as a deprecated alias (stays live server-side; consumers unaffected). - Add the abstain-to-source boundary (ROC-rule-deterministic; PD/Forms route-to-source). - Correct coverage: 70 in-scope Orders served + 153 reasoning corridors (was 67/738); state staged/uneven depth honestly. - Bump plugin + skill to 0.3.0; .mcp.json / marketplace.json descriptions updated. Honest-claims: no blanket "hallucination-free"; the stated property is that on a decidable ROC question it returns a proof-carrying determination or abstains, and the zero-LLM resolve path never fabricates a rule, sequence, or citation.
8cdba15 to
5c697e5
Compare
1 participant
Adds an external plugin wrapping the MikeROS / Reasonex MCP server: a neuro-symbolic legal reasoning engine over the Singapore Rules of Court 2021. LLM-based perception feeds a deterministic symbolic core (closed-vocabulary predicate evaluation over an SMT/Z3 solver), which feeds LLM articulation. Runtime reasoning is deterministic — the same facts yield the same determination, with no LLM in the inference loop. Every determination carries a re-checkable proof, or the engine abstains.
What it gives Claude (the differentiated property): on a decidable Singapore civil-procedure question, the engine returns a proof-carrying determination — which rule controls, the required sequence, and per-step compliance — or it abstains. It does not assert a procedural conclusion it cannot ground. This is the one thing an LLM alone can't guarantee on this surface: it cannot be confidently wrong, because the resolve path is zero-LLM and every cited rule is corridor data (citation firewall by construction). It complements the existing research connectors — they retrieve and cite; this one determines and proves on a decidable procedural question.
The perception layer is inspectable and correctable. The deterministic core is only as good as the facts fed into it, and those facts are LLM-extracted — so every response echoes back the engine's perceived facts. Claude (or the reviewing attorney) can see exactly what the engine understood, correct a misread, and re-call. The obvious failure mode for any fact-to-rule system — wrong facts in, confident wrong answer out — is a designed-in correction loop, not a black box.
Why this fits the suite. CONNECTORS.md names regulatory primary sources for Singapore (Singapore Statutes Online) and a curated regulatory knowledge base as wanted connectors. MikeROS is exactly that for the procedural layer — a curated, verified, neuro-symbolic reasoning engine over the Singapore Rules of Court, with citation-ready provenance and a proof-or-abstain guarantee.
Primary tool — resolve_procedure: facts (natural language) → controlling rule chain + procedural sequence + per-step compliance verdicts + re-checkable proof for entailed steps + the Z3-disposed candidate set (each candidate APPLIES / UNDERDETERMINED) + honest abstentions + the engine's perceived facts. Use it to determine which ROC rule controls, what sequence applies, or whether a step is procedurally correct — not for interpretation, drafting, strategy, or non-Singapore-procedure (Claude handles those). Also: ask (single-rule pinpoint + false-premise detection) and conversation_turn (direct end-user client).
See it for yourself (the inspectable artifact). The endpoint is live — run a query and inspect the determination, the candidate set, the proof object, and the abstention behavior on a question you choose:
"I want to enforce a registered foreign judgment" → returns the disposed candidate set (the applicable enforcement routes, each APPLIES / UNDERDETERMINED) so Claude can narrow with the user.
A costs/sequence question → returns the controlling rule chain with a re-checkable proof for each entailed step.
A filing-format / e-service / Form question → abstains to source (returns a pointer to the controlling Practice Direction / Form), rather than fabricating an answer.
Endpoint: mcp.reshuffleai.com/mcp/ (per-key auth over streamable HTTP, per CONNECTORS.md). For this audience, the inspectable proof closes what the description can only open — please run it.
Honest scope. The engine compiles the Rules of Court 2021 rule text — 70 of 70 in-scope Orders, 153 verified reasoning corridors (78 tree + 75 singleton), deployed live. Coverage is staged and uneven in depth; the response carries that honestly (it labels coverage and abstains rather than guess on thin areas). The Practice Directions / Forms layer is not yet compiled — where control passes there, the engine routes to source rather than answering. The connector is honest about that boundary.
Why this helps pro bono lawyers and self-represented litigants. The mistakes that lose a case for an under-resourced litigant are usually procedural, not substantive — a missed deadline, the wrong filing format, a misunderstood e-service requirement, or simply not knowing which rule controls. A pro bono lawyer working outside their usual practice area, and a self-represented litigant with no practice area at all, are the most exposed to exactly these traps. This engine helps them identify the controlling rule and required sequence with a re-checkable proof — and, critically, abstains rather than guess when it can't determine the answer. For an unrepresented litigant especially, the value isn't just getting an answer; it's not getting a confidently-wrong procedural answer they'd act on, because they can't always tell a sound answer from a plausible one. (Every output remains a draft for attorney review, not legal advice.)
We're offering the connector free to pro bono practitioners in Singapore. Happy to make any adjustments to meet the marketplace bar — CLA signed, all checks green. Understood the queue is long; thanks for considering it.