DSH Plugins Marketplace

DSH Plugins

Plugins

/

dsh-diwen

P

dsh-diwen

Discovered

dsh-diwen — Diwen (迪文) Concept Coordinate System

What Diwen is: a communication language for machines

Diwen is designed for machine-to-machine communication — a machine-verifiable concept coordinate system for agent-to-agent exchange, the deterministic content in natural-language↔machine interaction, and machine-to-machine interaction.

Natural language is ambiguous and drifts; Diwen anchors meaning to discrete concept coordinates (S3a semantic field × S3b construction), so machines can align precisely, and each concept renders back to a unique natural-language definition. Diwen is not a translator and does not generate natural language; it is a verifiable semantic intermediate representation above natural language.

中文概要见 说明.md。

The two mapping directions (asymmetric by design)

  • Natural language → Diwen is NOT unique, and may be unreachable. A word can carry several senses (polysemy), so mapping needs model word-sense disambiguation (WSD) — it is not mechanical (Axiom ②). When no concept fits, Diwen does not guess or force a nearby concept: it preserves the original text verbatim (Axiom ④, "no evidence, no concept"). Unreachable is a first-class, honest outcome.
  • Diwen → natural language IS unique and deterministic. Each concept renders to exactly one canonical Chinese definition (Axiom ①), machine-verifiable, zero drift. Future upgrade: a concept will render either to a unique deterministic definition or be marked unreachable — both outcomes are possible, but never an uncertain / non-deterministic rendering. Determinism is the invariant.

This asymmetry is the point: machines align on discrete concepts going up, and get guaranteed-stable meaning coming down.

Use cases — and the measured benefit in each

Benefits are measured, not advertised: token counts use real cl100k_base BPE (npm run benefits → scripts/qa/benefits-measure.mjs); accuracy comes from blind WSD evaluation; semanticApproval=false throughout. Token savings are relative to natural-language definitions / disambiguation, not bare Chinese words (Chinese is already token-dense).

ScenarioWhat Diwen addsMeasured benefit
Multi-agent collaborationAgents exchange concept coordinates instead of re-describing meanings in ambiguous NLA concept reference costs 3 tokens (coordinate) vs 13.75 (NL definition) = −78.2%; by id 1.62 tokens = −88.2%. Sense alignment: model WSD 0.9923 (qwen4) / 0.9933 (deepseek) concept accuracy
Sub-agent communicationLead ↔ sub-agents exchange tasks/results as deterministic conceptsSame compact reference; rendering 100% deterministic (3317/3317 round-trip, 0 failures) → no drift across turns or agents
Deterministic knowledge baseKnowledge anchored to machine-verifiable concepts, not fuzzy textCanonical concept index: id vs lemma −12.8% tokens; coordinate vs definition −78.2%; every entry renders identically
Context summaries / promptsCompressed meaning carried as concept coordinates, stable across modelsCompact concept references (−78–88% vs definitions); stable canonical rendering across turns/models
Error correction (any NL→Diwen)Model WSD picks the right sense; honest abstention otherwisePolysemous mapped words: mechanical 0.192 → ~0.99 with WSD (≈+80pp error correction); verbatim recall 0.9931 / precision 0.9983 → no fabrication
Real-time / localLocal decision models, sub-second; deterministic render instantd3:4b 0.09s, clef-flash 0.16s, clef 0.45s per example; cloud cascade routes easy cases to a small model (cost saving)

Honest caveats: token savings are vs NL definitions, not raw sentences (a coordinate list is not always shorter than the original Chinese); coverage is <100% by design (Axiom ④); model accuracies are measured on frozen holdout / pressure subsets, not arbitrary real-world text.

Status: Chinese is deliverable

Current scale: 3317 concepts / 194 semantic fields; 162 surfaces / 1024 senses (mapped 852 / verbatim 172 / approximate 0); 2889 gold sentences / 1024 senses; frozen holdout 116; pinyin 3316/3317; all 3317 concepts reachable via the Chinese surface pack. Private Diwen extension supported (user concepts, id ≥ 699051).

Concept accuracy — blind WSD evaluation (semanticApproval=false):

ModelInterfaceConcept accuracy95% CISpeed
deepseek (cloud)chat0.9933[0.9827, 0.9990]—
qwen4 (cloud)chat0.9923[0.9856, 0.9981]—
Qwen3.6-35B-A3B (local MoE)chat0.9667±~5pp~2s
clef (27B local decision)--ask --choice0.96[0.9155, 0.9815]0.45s
clef-flash (9.1B decision)--ask --choice0.9467[0.8983, 0.9727]0.16s
d3:4b (decision)--ask --choice0.9267[0.8735, 0.9586]0.09s
Deterministic mechanical (lower bound)longest match0.288±2.81pp clusterinstant

Model rows are measured on the round136 gold polysemy pressure subset (1040 examples, blind, zero leakage); the deterministic row is the current full-gold (2889) mechanical lower bound — not the production path.

Production path (enhanced + calibrated; frozen holdout / real domain):

PathConcept accuracyBasis
d3:4b enhanced + calibrated curated t400.9828 (114/116)curated holdout (verbatim recall 0.9839)
cloud deepseek enhanced1.0holdout110
qwen3:8b chat enhanced0.9727holdout110
cross-family ensemble (C3 weighted vote)0.9636holdout110
d3:4b enhanced real-domain t500.8056real domain 36 (small sample, wide CI)

±3pp certification (product metric): enhanced d3:4b on frozen holdout — concept accuracy 0.9727, cluster bootstrap 95% CI ±3.07pp (90 senses, ρ≈0). Deterministic baseline cluster CI ±2.81pp (1024 senses) — baseline ±3pp also achieved (converged ±5.89→±5.05→±4.43→±3.86→±2.97→±2.81pp); concept coverage C_gold=0.78 after filling 101 genuine concept gaps. The baseline is the disclosed mechanical lower bound, not the production path.

Regression: run-all 718/0 + standalone 55/55 + MCP selftest 45/45; Diwen→Chinese round-trip 3055/3055, identityFailures=0 / determinismFailures=0.

Axiom ① (Diwen → unique Chinese definition) is 100% deterministic. semanticApproval=false always holds: passing mechanical validation ≠ semantically correct ≠ human-approved. When no concept exists, the original text is preserved verbatim (Axiom ④) — nothing is fabricated.

Roadmap

  1. Continuously improve Chinese usability (coverage expansion, real-domain generalization).
  2. Complete development of 27 languages, then support more.
  3. Post-hoc, sense-level concept sharing only where coordinates are provably identical (default: do NOT merge; most pairs stay independent and are linked, not merged). Eliminate natural-language partitions as a machine-interoperation overlay — not by deleting language diversity.
  4. Machine-expand concepts and semantics for more natural languages.
  5. Continuously refine humanity's existing concepts and semantics.
  6. Machine-expand concepts that do not yet exist in natural language.
  7. Machine-refine concepts describing the physical world.
  8. Machine-refine abstract concepts describing the world.
  9. Machine-abstract higher-density concepts, evolving structures better suited to each concept.
  10. Continuously improve Diwen so it can clearly describe all things in the world.

Quick start

This repo is Diwen's open-core public distribution: open loader shell + closed core (minified bundle) + encrypted concept data (AES-256-GCM). Clone, then npm install to use offline.

npm package

Published on npm (current diwen@6.10.1) — install directly:

npm install diwen
import { mapChinese, describeConcept } from 'diwen';

// Natural language → Diwen fill contract (Axiom ②: model sense selection; Axiom ④: no concept → verbatim)
const r = await mapChinese('我吃饭', { model: 'deterministic' }); // 'auto'|'decision'|'off'
console.log(r.contract.status, r.contract.words, r.semanticApproval); // semanticApproval always false

// Diwen → unique Chinese definition (Axiom ①)
console.log(describeConcept(1223).lemma_zh); // 折

Or use from this repo: cd packages/diwen && npm install, then import ... from './index.mjs'.

  • model:'decision' uses a local ollama decision model (d3:4b etc.) + calibrated abstention; enhance:true enables the enhancement layer.
  • The runtime segmentation dependency @node-rs/jieba is installed automatically.

Private Diwen (add your own concepts)

You can add your own private concepts without touching the public core; they merge into the lexicon and work across mapping / WSD / describe / render. Private concept IDs must be ≥ 524288 (the custom address partition); glyph coordinates (s3a,s3b) must avoid base concepts (collision-checked at load).

diwen private template -o my-private.json    # scaffold (example concept + a free coordinate)
diwen private free-coords 20                 # list free (s3a,s3b) coordinates
diwen private validate my-private.json       # validate id partition / collisions / schema
export DIWEN_PRIVATE_LEXICON=my-private.json  # load; then map/concept include your concepts

Set DIWEN_PRIVATE_CANDIDATES for private WSD senses. Your private file is your own data (plain JSON, not uploaded/redistributed; no core decryption needed). Programmatic API: loadLexicon(path,{privatePath,noCache}), validatePrivateRoots, freeCoordinates, privateTemplate, mergeLexicon. See docs/PACKAGE-USAGE.md §5.

MCP server (zero third-party protocol dependency; core via npm diwen)

npm install diwen-mcp
npx diwen-mcp        # or repo-local: node packages/diwen-mcp/server.mjs

Configure your MCP client with command=node, args=[<absolute path>/diwen-mcp/server.mjs]; 3 tools: diwen_map{text,model?,enhance?}, diwen_describe{id}, diwen_select{cases,mode?,weights?} (cloud cascade/ensemble sense selection; requires DIWEN_KEYS_FILE). Self-test: node packages/diwen-mcp/selftest.mjs (45/45).

DeepSeek Harness (dsh) plugin

Open Cordis shell + depends on npm diwen core + deterministic force layer (diwen-dsh@0.3.2): registers diwen_map/diwen_describe + skill diwen. dsh plugin --profile <p> add diwen-dsh (or a local path). Force-layer seams (systemPrompt required instruction / pre-step input projection / post-execute result projection) each have a settings toggle (forceSystemPrompt/forcePreStep/forcePostExecute, default on, with descriptions). Self-test: node packages/diwen-dsh/selftest.mjs. See docs/DSH-FORCE-USAGE.md.

Manuals

Decision-model choice capacity (advice + error feedback)

A decision model's --choice option limit varies per model and is not fixed. Measured hard limit is 255 for all (ollama interface); effective discrimination limits: tev1:0.8b=32 / d3:4b=128 / clef-flash=255 / clef=128. You decide which models to use; Diwen does not forbid any: checkModelCapacity gives advisory hints; on a runtime error decisionSelect uses classifyModelError to report "this is a model-capacity issue (e.g. too many options) — switch model or reduce candidates".

Repository layout

packages/diwen/       npm package: index.mjs/env.cjs open shell + dist/core.mjs closed core + dist/data.enc encrypted data
packages/diwen-mcp/   MCP server (open shell, reuses closed core)
packages/diwen-dsh/   dsh plugin (open Cordis shell + npm diwen core + force layer with settings toggles)
docs/USAGE-MANUAL.md  non-technical usage manual
docs/PACKAGE-USAGE.md integration guide
docs/DSH-FORCE-USAGE.md  dsh force layer + settings toggles
说明.md               Chinese overview
SPONSORS.md           sponsor page
sponsors/             donation QR images
LICENSE               license

License

Commercial use permitted (free, no separate authorization), no redistribution, attribution required: credit Diwen / 迪文 + author Pondsi + repository link. Full terms in LICENSE.

Support

If Diwen helps you, consider sponsoring — see SPONSORS.md. Donations buy no extra features; everything stays free.

Limitations

  • Currently Chinese is deliverable — 2 languages: Simplified (zh) + Traditional (zh-TW), sharing the same concept coordinates. The other 25 languages are future roadmap (Axiom ⑤: language-independent mapping, no pivot).
  • Coverage is not 100% (by Axiom ④ design); words without a concept are preserved verbatim, never fabricated.

Author: Pondsi · https://github.com/Pondsi/dsh-diwen

Comments

Loading…