Criterion Loom MCP server

Criterion Loom

日本語 · Documentation

Make the gap between an engineering claim and its evidence inspectable.

AI-assisted work can leave the difference between fact, hypothesis, missing evidence, and human acceptance trapped in conversation. Criterion Loom exists to externalize those distinctions as versioned, inspectable audit artifacts—so the next correction or decision does not have to trust a fluent summary.

Criterion Loom is the public project name. Its current distribution and CLI are named semantic-guard at development version 1.2.0.dev0; the canonical import package is semantic_guard, and the MCP server entry point is semantic-guard-mcp.

The current v1 public workflow audits one structured functional requirement at a time and, separately, bounded Japanese direction-binding expressions. It retains its own result contract. Plan, diff, completion, and other workflow audits are available through the explicit workflow interface. The separate candidate interface retains the unadopted governed vNext contract.

This development version combines three implementations in one installation. Choose the contract and server explicitly using the integration guide. It is not a release or a deployed-default change.

Why this is different

A single score or pass/fail flag would collapse questions that need different answers. semantic-guard keeps them apart:

• what the selected rule concludes;
• whether that conclusion is provisional or terminal;
• whether a challenge or conflict remains open;
• how much of the declared subject was actually covered;
• whether the configured workflow should continue, warn, or stop;
• whether a human has accepted anything at all.

The result is decision material, not the decision. Analyzer output cannot grant itself supporting authority, and pass never means correctness, release approval, security certification, or human acceptance.

Quickstart: one bounded proof

Prerequisites: Python 3.11 or later and uv. The current project is run from a source checkout; no package-index artifact is claimed here. The projection below uses POSIX shell syntax. Other shells can run the audit command without the pipe and read the same JSON payload directly.

git clone https://github.com/mait2355n/criterion-loom.git
cd criterion-loom

uv run --locked --extra nlp-ja semantic-guard audit-direction-binding \
  --text '横一列を左から右へ辿るとき、Aの次の項目はどれですか？' \
  --morphology sudachi \
  | python3 -c 'import json,sys; r=json.load(sys.stdin); print(json.dumps({"state": r["primary_rule_evaluation"]["state"], "direction": r["direction_binding_summary"]["frames"][0]["direction_binding"]["direction"], "workflow": r["workflow_disposition"]["status"], "human_acceptance": r["acceptance_owner"]["acceptance_status"]}, indent=2))'

The selected fields are:

{
  "state": "satisfied",
  "direction": "left_to_right",
  "workflow": "pass",
  "human_acceptance": "pending"
}

This example exposes two separate boundaries: a rule result does not become human acceptance, and a missing required direction is not filled in from convention. In the latter case, the audit fails closed at the rule boundary by reporting a gap and warning the workflow. The direction-binding reference (Japanese) keeps the paired inputs and exact field states.

Current public surface

Capability	CLI	MCP	Current boundary
Requirement-relation audit	audit-requirement	audit_requirement_relations_tool	One seven-field structured functional requirement; morphology is signal_only, dependency and caller-supplied LLM analysis are candidate_only
Direction-binding audit	audit-direction-binding	audit_direction_binding_tool	Registered scalar and non-scalar Japanese expressions with direct attachment; not unrestricted language understanding
Closed schema access	schema	semantic_guard_schema_tool	25 known schema names; schema availability does not imply an integrated public workflow
Explicit legacy comparison	shadow-compare	shadow_compare_legacy_tool	Operator-owned external 0.1.0 root; trusted comparison requires baseline digest agreement. The MCP route is disabled by default and requires operator configuration; legacy is not a truth oracle

The current source also contains candidate lifecycle and assurance contracts. They are useful design and test material, but they are not silently promoted to public end-to-end features.

The additional CLI routes workflow COMMAND and candidate COMMAND select separate contracts. Their servers are semantic-guard-workflow-mcp (27 tools) and semantic-guard-vnext-mcp (three tools); the canonical server retains its four tools.

Choose an interface

Surface	Use it when	Start here
CLI	A person, script, or CI job owns invocation and JSON persistence	uv run --locked semantic-guard --help
MCP	An agent client needs the same bounded tools over standard input/output	uv run --locked semantic-guard-mcp
Companion Skill	An agent (Codex or Claude Code) should preserve the audit boundary while planning and implementing work	skills/semantic-implementation/ · agent setup (Japanese)

For an MCP client that accepts command, args, and cwd, the equivalent source-checkout configuration is:

{
  "mcpServers": {
    "semantic-guard": {
      "command": "uv",
      "args": ["run", "--locked", "semantic-guard-mcp"],
      "cwd": "/absolute/path/to/criterion-loom"
    }
  }
}

Client field names vary; the server contract is standard input/output transport, not that particular configuration envelope. The companion Skill is repository material, is not included in the wheel or sdist, and is not installed automatically.

Evidence before adjectives

The dated 2026-08-23 1.1.0 evidence records report:

• 608 source tests passing on the recorded Python 3.11 and 3.13 environments;
• 20 packaged-contract checks on one selected local wheel and sdist;
• 24 packaged schemas, four CLI commands, and four MCP tools observed;
• 222 registered direction gap/bound cases replayed on a recorded fresh-wheel Sudachi environment;
• the direction-binding slice merged to GitHub main with the defined four CI jobs passing.

These observations are bound to the subjects recorded there; they are not an automatic verification of the current HEAD. They are contract and registered-case observations, not benchmarks, adoption claims, field accuracy, unrestricted Japanese coverage, production qualification, or human acceptance. Read the exact subjects and limits in Implementation status (Japanese) and the dated direction-binding integration evidence (Japanese).

Documentation

Goal	Document
Understand the current package and non-claims	Current public surface
Run and automate the CLI or MCP server	Operations guide (Japanese)
Compare bound and missing direction-binding results	Direction-binding audit (Japanese)
Inspect implementation and evidence status	Implementation status (Japanese)
Browse current reference, evidence, history, and prototypes	Documentation map
Move from the archived 0.1.0 line	Migration guide

Machine schemas and the verification source outrank explanatory prose for field constraints. Dated reports describe their recorded subject and time; they do not automatically describe the current tree.

Version and legacy boundary

The publication-repaired semantic-guard 0.1.0 predecessor is preserved under legacy/semantic-guard-v0.1.0/. Its original bytes remain anchored by the tag and commit named in the archive manifest. The old request, plan, diff, finish, convention, reviewer, and acceptance-bundle commands are not transparent aliases of v1.

Current source identity, field validity, policy adoption, operational default, and historical preservation are separate states. See the historical 1.0.0 canonical promotion decision and the change log.

Contributing and support

• Contributing
• Support
• Security policy
• Code of Conduct
• MIT License

Final acceptance, risk acceptance, policy adoption, default cutover, and legacy retirement remain human decisions.