An honest coupling assistant.
hequ is an experiment in systematic cross-domain equation discovery. This site is its public surface: a read-only render of every equation, every coupling hypothesis, every AI review board vote, and the append-only ledger that anchors all of it.
The framing
The sieve is an AI-assisted coupling assistant, not a coupling oracle. Every historical attempt at fully automated discovery of cross-domain scientific relationships has failed or scoped down. hequ generates ranked candidates with explanations; an AI review board of three core members (Claude Opus 4.6, OpenAI GPT-5, Gemini 2.5 Pro) — with Grok-4 as optional fourth seat on high-stakes governance rounds — accepts or rejects each candidate before it enters the ledger as PROVED. This framing is binding — nothing is labeled a "discovery" without a recorded vote, and every round logs whether it seated 3 or 4 reviewers.
The full review protocol, from canonical problem unit tests through live sensor grounding to the four-phase academic review board, is documented on the methodology page.
Provenance
- Every equation is parsed from a YAML source with pint-typed variables and an I-ADOPT compositional descriptor (object, property, context, constraint).
- Every canonical form is simplified by sympy before matching.
- The Layer 5 coupling sieve applies a two-gate prefilter (pint dimensions + explicit domain-adjacency matrix), a six-way bond-graph role classifier, and a structural bijection rescorer.
- The physical-constraint filter runs Tellegen pairing, Onsager reciprocity, and SHO-style energy conservation.
- The emergent-properties analysis computes symbolic steady states, linear stability, Buckingham Π groups, and conserved quantities.
- The AI review board votes independently per model; the ledger records each vote verbatim with verdict, reasoning, and timestamp.
What this site is not
- A published scientific paper. Every finding is a candidate pending continued review.
- A substitute for domain expertise. The AI review board is a sanity check, not a peer review.
- A benchmark against SINDy / AI Feynman / PySR — those tools work within a single domain, not cross-domain.