Scope · limits · disclosure

What this is, what it is not, and what we withhold

The project has had almost no publicity so far, and this release is deliberately partial. We would rather say so plainly than let you discover the gaps.

Definitions

Five things it is regularly mistaken for.

Not a search engine. The output is typed entities with declared inputs, outputs, prerequisites and implementations, plus the edges between them — intended for a program, not a reader.
Not a RAG system. RAG derives structure at query time and discards it. A compiler derives it once and keeps it.
Not a vector database. That is a component of the runtime. The content is the typed edge families — hierarchy, ordering, implementation. Vectors give you neighbours; they do not give you precedes.
Not an execution ontology. We model engineering decisions, not action semantics. Complementary to KnowRob and IEEE 1872, not a replacement.
Not classical knowledge compilation. We use the term in a source-level, engineering sense and claim none of the tractability results of the OBDD/d-DNNF line of work. The paper says so explicitly.

Disclosure

Staging the opening, rather than staging a retreat.

We would rather not put the most load-bearing material into the open while the project is still this quiet. Here is exactly where the line currently sits.

Published in full

The three-level taxonomy. Bilingual names and levels for every capability. The asset index with upstream sources. The dependency network's structure — nodes, domains, degrees, topic tags, 61,476 edges. The 126-query retrieval golden set. The adjudicated duplicate ledger. Every number we report, plus the script that regenerates them. Nine topics reproduced completely, with their diagrams and research ledgers.

Withheld for now

Full capability descriptions in bulk — the fields carrying inputs, outputs, constraints and performance envelopes. These are the substance of the artifact, and what would be worth copying wholesale. You can read them in full for the nine case-study topics, which is enough to judge whether the structure is real.

The intention

Open it completely. The value we are building is the compilation pipeline and the standard, not a moat around a name list — and a knowledge layer nobody can inspect is not infrastructure. We would rather stage the opening than stage a retreat.

Ask for what is missing

If you want access to something specific that is not here — for research, for a benchmark, for an integration — write to wangyuqi11and11@163.com. We are far more likely to say yes than to say no.

Honest status

What we have not established.

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.
Build cost is unmeasured. The pipeline records no per-topic time or token cost, so the economics of front-loading work cannot yet be reported.
56.95% of task chains are purely linear even though the schema supports branching. The compiler emits the simplest structure that passes.
The duplicate-detection result is conditional. Pairs in the adjudication ledger are conditioned on having been flagged, so it does not estimate the gate's overall discriminative power, and gate recall is unmeasurable from that data. The lexical row rests on eight positives and is not a finding.

Design intent

What changes, if it works.

Stated as intent — not as a measured outcome.

Route survey: Read dozens of papers over several days → candidate routes with alternatives, in minutes.
Module selection: Search GitHub and find out the hard way → every known implementation, with its attribution.
Orchestration: Hand-wire modules and debug the interfaces → ordered from the dependency graph, adaptation flagged.
Reuse: Each project reinvents → the artifact accumulates; every task strengthens it.
Onboarding: Oral tradition and a long ramp → the graph is the map.

Read the paper → Back to the front page