Read CLAUDE.md completely before changing anything in this repository. It is the full directive: architecture, invariants, preset and RDF rules, and the reasoning behind each. This file exists because your tool does not load it automatically, and it is a pointer, not a replacement.

Where this file and CLAUDE.md disagree, CLAUDE.md wins. It is maintained in the same commit as the change it describes; this file is deliberately short so it has little room to drift.

SSTIM is the open formalized knowledge standard: specification, RDF vocabulary, semantic infrastructure, documentation, interoperability work, and shared identifiers. The broader SSTIM ecosystem adds its reference tooling and community, plus sensory-stimulation applications and initiatives related by adoption, contribution, support, or documented domain relevance. SSTIM Workbench is SSTIM's non-normative executable reference environment, with a precision multi-engine audiovisual application plus Graph Navigator, annotation, and a SPARQL interface over the SSTIM ontology.

Existing BSC/BSC Lab semantic identities and compatibility formats are preserved provenance, not the public project brand. BioSynCare is a separate closed-source commercial application that adopts and contributes to SSTIM; SSTIM Workbench does not feed it today. Do not conflate them or add BioSynCare-specific logic to SSTIM.

These are the rules where a mistake is expensive or irreversible. They are summarised, not restated in full — CLAUDE.md §3 is authoritative and includes the reasoning, which matters more than the rule.

1. The engine timing context is the only clock. engine.getAudioContext().currentTime is the sole authority for audio-visual sync. Never Date.now(), performance.now(), setTimeout or setInterval for timing.
2. Never bundle static/worklets/. AudioWorklet processors load by URL at runtime. Importing one puts it in Vite's module graph and breaks it.
3. Never allocate inside AudioWorkletProcessor.process(). ~2.67 ms budget on the audio thread; garbage collection there is audible.
4. Never modify ontology files without an explicit instruction naming the file. Everything under static/ontology/ — manifest.json is the authoritative inventory — plus the three defensive publications in docs/technical/, which are timestamped prior-art records.
5. No health, medical or treatment claims in user-facing copy. Wellness framing only. Regulatory, not stylistic. See docs/concept/SCOPE.md.
6. Name your instrument before claiming something is missing. Before writing that anything is absent, untested or undone, say what you checked with and confirm it could see where the thing lives. Use python3 scripts/locate-iri.py <iri> for SSTIM identifiers, and docs/ontology/TERM_INDEX.md before saying a term does not exist. An unreachable instrument is INCOMPLETE, never absence.
7. Svelte 5 runes only. $props(), $state(), $derived(), onclick, {@render children()}. Not export let, $:, on:click, <slot />.

make validate for anything under static/ontology/ — it is the same gate CI runs. make test and make check for application changes. Both take a while; make -j4 -Otarget validate is faster and equivalent.

This repository uses Svelte 5 runes and a modular OWL/SKOS/SHACL ontology. If you are unsure whether SSTIM already defines a term, grep docs/ontology/TERM_INDEX.md — it is generated from the 18 modules and CI-checked. Eighteen modules is more than anyone searches reliably by hand.