Scientific knowledge infrastructure

Knowledge Compiler for Embodied Intelligence

Compile knowledge Structure capabilities Enable robot intelligence

Robotics knowledge already exists — as prose, not as structure. Entropy Box compiles it once into a typed graph that programs can consume.

This page is the static mirror. Complex-query retrieval, one-click plan assembly and the live dashboard run at xiangshang.ngrok.app — no registration, no install.

The compilation path Unstructured sources are compiled once into a typed graph, from which capability chains and executable robot task plans are derived. PAPERS REPOSITORIES DATASETS KNOWLEDGE COMPILER GATED · OFFLINE · AUDITABLE TYPED GRAPH CAPABILITY CHAIN ROBOT TASK PLAN
Research dataset · snapshot 2026-08-30
37,691
Robot capabilities
bilingual, typed
66,609
Knowledge relations
mechanically derived
7,907
Executable task chains
97,379 steps
99.6%
Traceable evidence
provenance resolvable
11,442
Implementation assets
10 asset types
2,524
Compiled topics
of 2,524 leaves
1.57x
Amortization factor
21,380 derivations avoided
2,362
Adjudicated duplicate pairs
embedding AUC 0.51

Regenerated from the artifact on disk by measure.py — no GPU, no network, no service. Snapshot 2026-08-30. The paper's public snapshot (28 August 2026) reports 2,511 topics, 37,673 capabilities and 11,397 assets; the artifact has since grown slightly.

Overview

A language model has no chief engineer's view.

It writes the detector you ask for. It cannot see that one task hides a dozen steps and the dependencies between them.

A craftsman who has mastered every part but never held the drawing. We build the drawing.

Competence without composition

Writes a detector. Cannot deliver “pick up the red cup” end to end.

Blind to dependencies

Perception → localization → planning → control. Which feeds which is what gets lost.

The route is a black box

A new task means reading dozens of papers to compare routes by trial and error.

Orchestration eats the time

Most hours go to wiring modules together, not to the algorithm work.

Why compile

Retrieval compiles on every query. A compiler compiles once.

Query-time derivation

RAG · GraphRAG · agentic search

  • Question
  • Retrieve passages
  • Model derives the structure
  • Answer, in prose
  • Structure discarded — the next query redoes it

Knowledge compilation

Entropy Box

  • Corpus
  • Compile — gated, offline, auditable
  • Typed persistent graph
  • Retrieve over structure
  • Assembly, with the gaps named

Capability reuse is rising, not saturating.

Each compiled unit is consumed 1.57× on average — 21,380 re-derivations avoided — and the curve has not flattened.

22.8% reused across topics max reuse 59×
1.0 1.2 1.4 1.6 50 topics compiled → 2,524

Amortization factor α against topics compiled.

Architecture

Raw knowledge in, robot intelligence out.

Four stages under a persistent control plane. Every transition is mediated by a durable artifact and a mechanical gate.

Stage A1
Chain derivation
A boundary brief fixes the topic, its siblings and exclusions; three technical routes are derived and admitted through the chain-list gate.
2,524TOPICS
Stage A2
Isolated research
One agent per chain — source-grounded search and asset verification — writes an evidence document with decisions, URLs, negatives and a ledger, then passes a document and stage gate.
7,907CHAINS
Stage B
Typed assembly
A topic assembly agent reads one chain at a time, normalizes entities and adjudicates overlap into an incremental topic JSON.
37,691CAPABILITIES
Stage C
Finalization
Schema and graph gates run; one bounded repair is allowed; the record is accepted or written to the failure ledger.
2,362ADJUDICATED
Asynchronous admission plane
Registry sync → duplicate/conflict ledgers → graph materialization → semantic views and incremental index
Agent output never writes the global registry or vector store directly. Acceptance is decided by deterministic programs, not by model self-assessment.
66,609EDGES
Research

Results from the system's own records — and from the benchmark suite.

Embedding similarity cannot decide duplication negative

On 2,362 adjudicated pairs, once the gate has flagged a pair the similarity score carries too little signal to decide the merge (embedding AUC 0.509, CI [0.453, 0.564]). Raising the threshold does not help: precision stays flat near 5–6% while recall of true duplicates collapses from 100% to 4%. Merging automatically above a threshold would be wrong more than nine times in ten — which is why the merge decision is deferred to an LLM adjudicator that reads both records and writes the rationale into the ledger. The similarity score flags; the model judges.

Pathpairsdup. prec.AUC
Embedding1,867103 0.0550.509
Lexical4958 0.0160.689

Conditioned on having been flagged, so gate recall is unmeasurable here; the lexical row rests on eight positives. Full note →

0% 33% 66% 100% 0.862 similarity threshold → 1.0 recall — collapses precision — flat

Tightening the gate throws duplicates away without buying precision.

Provenance resolvability, proposed as a metric 99.6%

Every topic cites the documents it was assembled from, and because those are files the citation is mechanically checkable: retrieval documents resolve at 99.56%, chain documents at 82.0%. Cheap to compute, hard to game — and a system that cannot report it does not have one.

Unverified fields are left unverified by design

2,391 fields across 20.5% of topics are marked unverified rather than filled in. Zero markers would not mean a better artifact — it would mean a pipeline that fills gaps instead of declaring them.

Compilation beats plain RAG — on grounded generation.

This is the result we care about. On the EntropyBench Track-P plan-synthesis benchmark (24 tasks across 8 robotics domains), every ungrounded baseline — LLM direct, BM25 RAG, hybrid RAG — emits plans whose asset and interface claims are 100% unsupported. Entropy Box cuts that to 5.2% (−94.8pp vs hybrid RAG, CI [−97.4, −92.1]). The same grounding transfers to downstream simulation code generation: pass-1 executable plans rise from 0.58 (Vanilla RAG) to 0.92.

Plan synthesis — unsupported claims Track-P
Systemunsup.↓viol.↓cov.↑
LLM direct100%100%0%
BM25 RAG100%100%0%
Hybrid RAG100%100%0%
Entropy Box5.2%66.7%35.4%

Unsupported claims · constraint violations · constraint coverage. Paired diff vs hybrid RAG, 95% CI in the paper §9. Reproduce from evaluation/.

Sim codegen — first-pass execution 12 tasks
Systemimportguardspass-1
Zero-shot0.550.160.33
Web search0.690.420.50
Vanilla RAG0.750.500.58
Entropy Box0.980.880.92

Import validity · constraint guards · pass-1 executable plans. Grounding in the compiled artifact, not raw retrieval.

Demo

Not a paragraph about stair descent — a plan with named implementations.

51 seconds, unedited. A goal is decomposed, matched across topics, and assembled into a plan whose nodes carry attributed open-source assets. The paper also reports two independent development cases that query the same public interface to build simulation applications.

Cropped only, to remove OS chrome. Not sped up.

Explore

Everything here is inspectable right now.

Open data

Published so the measurements reproduce and the resources get reused.

Data CC BY 4.0 · code MIT.

Open API — no key, nothing to install.

Four read-only endpoints, bilingual (中文 / English). Any client that speaks OpenAPI, MCP or REST can integrate — Claude, ChatGPT, Cursor, Trae, WorkBuddy.

EndpointPurpose
/api/consultSolution consult
/api/lookupEntity lookup (CAP_ / AST_)
/api/evidence/searchEvidence RAG
/api/searchHybrid search
curl -X POST "https://xiangshang.ngrok.app/api/evidence/search" \
  -H "Content-Type: application/json" \
  -d '{"query": "robot obstacle avoidance algorithms",
       "top_k": 5, "mode": "hybrid", "rerank": true}'

# OpenAPI schema:  /openapi.json
# MCP server:      /mcp/ontology_mcp_server.py
Honest status

What we have not established.

The benchmarks exist now; the honest gaps are the ones below. We would rather say so than let you find them.

  • No real-robot transfer. The two development cases are simulation applications; no real-robot performance is claimed.
  • Weak retrieval on hard intents. Identity queries are strong (MRR 0.89), but function (0.33) and multi-hop causal (0.30) queries remain weak.
  • Lower graph validity in plan synthesis. The /integrate planner drops graph validity to 62.5% while eliminating unsupported claims.
  • Semantic correctness unmeasured. Compiled descriptions pass schema and graph gates, not human semantic review.
  • 56.95% of chains are purely linear though the schema supports branching.